Dual Pathways of Category-Guided Attentional Selection: Task- Relevance Modulates Hierarchical Integration of Top-Down Control and Salience

Dual Pathways of Category-Guided Attentional Selection: Task- Relevance Modulates Hierarchical Integration of Top-Down Control and Salience | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Dual Pathways of Category-Guided Attentional Selection: Task- Relevance Modulates Hierarchical Integration of Top-Down Control and Salience Xia Wu, Yifan Liu, Junzhe Wang, Xiaoya Sun, Ying Chen, Yunpeng Jiang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6888298/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Category-guided attentional selection (CAS) enables efficient information filtering in complex environments through cognitive category formation. However, systematic comparisons between semantic and prototype-based CAS mechanisms are critically lacking, particularly in how they dynamically integrate top-down control and salience processing. Combining a hybrid experimental design (4 experiments, N = 80) manipulating cognitive load (high vs. low), stimulus salience (oddball vs. standard) and salience relevance (task-related vs. task-irrelevant), we dissected CAS dynamics across semantic (numbers/letters) and prototype-based ("O"/"X" shapes) categories. High cognitive load consistently impaired performance, demonstrating working memory-dependent top-down modulation. Salience effects emerged exclusively under task-relevant conditions, with oddball stimuli eliciting slower RTs and reduced accuracy, indicating goal-contingent weighting. Critically, cognitive load and task-relevant salience interacted during late decision stages, suggesting dynamic resource competition between control processes. Notably, prototype-based categories outperformed semantic categories in oddball trials, likely driven by perceptual similarity-mediated bottom-up integration. These findings extends a refinement to theories of category-guided attention into a dual-pathway model, where semantic categories rely on conceptual templates in working memory, while prototype-based categories leverage perceptual feature binding. category-guided attentional selection top-down stimulus salience prototype-based category Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1 Introduction Given the complexity of sensory input, human attention system must rapidly select goal-relevant stimuli while suppressing irrelevant distractors. While feature-based attention mechanisms allow rapid orientation toward simple attributes (e.g., color, orientation) (Treisman & Gelade, 1980), real-world environments present stimuli with inherent categorical variability that defies feature-level processing. For instance, searching for "vehicles" in traffic requires integrating diverse exemplars (cars, bicycles, scooters) into a unified category rather than tracking discrete features. By grouping stimuli into cognitive categories, category-guided attentional selection (CAS) reduces processing load and enhances recognition efficiency (Wu et al., 2016 ; 2017; 2020 ; Yang & Zelinsky, 2009 ; Wyble et al., 2013 ; Reeder & Peelen, 2013 ). Notably, categories are not monolithic, semantic-based categories (e.g., alphanumeric characters) rely on abstract conceptual rules stored in long-term memory (Goldstone, 1998 ; Giammarco et al., 2016 ; Lech et al., 2016 ; Wu et al., 2018; Wyble et al., 2013 ; Wu et al., 2016 ), whereas prototype-based categories (e.g., "O"-like shapes) emerge from perceptual similarity to a central exemplar through sensory integration (Lech et al., 2016 ; Wu et al., 2020 ; Yildirim et al., 2007 ). For instance, Wu et al. ( 2016 ) demonstrated that semantic-based CAS (e.g., colored letters) elicits delayed attentional orienting compared to feature-based attention, whereas prototype-based CAS (e.g., warm-colored "O" shapes) engages early perceptual integration (Wu et al., 2020 ). Despite these insights, fundamental questions remain unresolved: Do semantic and prototype-based CAS recruit distinct cognitive control mechanisms? How do they differentially integrate top-down goals and stimulus salience during executive tasks? Addressing these questions is critical for understanding how categorical flexibility adapts to real-world demands. Central to this adaptability is top-down control, a mechanism hypothesized to gate CAS efficiency by resolving competition between category-relevant and irrelevant inputs. While prior research establishes top-down modulation of feature-based attention through working memory (WM) operations (Wu et al., 2015 ; He et al., 2021 ), recent studies extend this to CAS contexts. Mihajlović ( 2024 ) demonstrated that top-down attentional control enhances search efficiency when aligned with task goals, though their paradigm focused solely on feature-level targets, leaving category-level mechanisms unexplored. Similarly, Shang et al. ( 2024 ) revealed that category-based top-down attention significantly contributes to memory search, yet their design did not dissociate semantic and prototype-based categories. Crucially, behavioral evidence suggests category-specific top-down dynamics. Semantic CAS (e.g., alphanumeric targets) activates later than feature-based attention (Wu et al., 2016 ), likely due to WM-dependent rule retrieval from long-term memory (Giammarco et al., 2016 ). In contrast, prototype-based CAS (e.g., "O/X" shapes) facilitates rapid guidance through perceptual similarity (Wu et al., 2020 ), potentially bypassing explicit rule maintenance. This dissociation aligns with neurocognitive models: semantic categories engage prefrontal "rule hubs" (Fan, 2014 ), while prototype categories recruit occipitotemporal regions for sensory integration (Lech et al., 2016 ). However, a gap exacerbated by the predominant use of visual search paradigms (Schmidt & Zelinsky, 2009; Yang & Zelinsky, 2009 ; Wyble et al., 2009; Wyble et al., 2013 ), which conflate attentional orienting and executive control, remains. Whether cognitive load differentially impacts these semantic vs. prototype systems still needs to be examined. Beyond top-down control, stimulus salience is another factor that dynamically interacts with CAS. Salience processing by stimulus prominence (the degree to which an item stands out perceptually) typically captures attention in a bottom-up fashion, but its impact on CAS likely depends on task relevance. Task-irrelevant salience (e.g., a uniquely colored distractor unrelated to the task) can involuntarily capture attention in standard feature-search paradigms. Classic studies using color singletons showed that even irrelevant singletons can disrupt visual search (Theeuwes et al., 1991, 1992). However, in CAS contexts, such interference appears attenuated by category-level filtering: Wu et al. ( 2020 ) found that an irrelevant color oddball did not measurably affect search for letters vs. numbers, presumably because the categorical goal allowed participants to filter out low-level color deviations. Conversely, task-relevant salience (e.g., a rare target category occurrence) is expected to amplify attentional effects through goal-directed weighting. Zhang and Gaspelin ( 2024 ) reported that physically salient objects only capture attention when they are task-relevant, supporting the idea of goal-contingent salience processing. Translating this to CAS, if an oddball stimulus aligns with the target category (e.g., an infrequent digit among letters), it should engage top-down control mechanisms to a greater extent, possibly affecting later decision-stage processing. In contrast, an oddball that is unrelated to the task goal might be suppressed early in processing (e.g., via distractor inhibition mechanisms). Neurophysiological evidence supports this temporal dichotomy: task-relevant oddballs modulate late event-related potentials associated with decision-making (P3), whereas task-irrelevant oddballs evoke only early sensory components like the N2pc, reflecting automatic suppression (Chen et al., 2023 ). Despite these insights, it remains unresolved how salience and category architecture jointly influence behavior. For instance, do semantic vs. prototype categories differ in their vulnerability to salient distractors, and does task relevance alter this? We address these issues by manipulating salience relevance across experiments. Notably, In our design, we operationalized “stimulus salience” via an oddball probability manipulation rather than a classic singleton (i.e., the oddball was defined by its rarity across trials, not by a unique feature within a single display). This approach emphasizes expectancy/surprise aspects of salience, which differs from traditional within-trial singleton paradigms. We return to this point in the Discussion when interpreting null vs. significant salience effects. The interplay between top-down control and salience processing epitomizes a core debate in attention research: are these mechanisms independent (serial model) or interactively integrated (parallel model)? In feature-based attention, Wu et al. ( 2015 ) identified temporoparietal junction (TPJ) as a nexus for integrating top-down goals and salience signals, with high cognitive load amplifying TPJ-mediated competition. Extending this to CAS, prototype-based categories that relative to perceptual feature binding (Lech et al., 2016 ) may prioritize salience integration through occipitotemporal pathways, whereas semantic categories that dependent on prefrontal rule hubs (Wu et al., 2020 ) might suppress salience interference via stronger top-down gating. Recent evidence hints at such dissociations, for example, Mihajlović ( 2024 ) found color-category contingencies modulated by task goals, while Shang et al. ( 2024 ) observed category-based memory enhancements contingent on salience alignment. However, whether top-down control and salience processing operate independently or interactively across category architectures and if such interactions are modulated by task relevance remains unclear. To dissociate the effects of top-down control and salience processing across category architectures, we conducted four experiments combining the Majority Function Task (MFT, Fan et al., 2008 ) and oddball paradigm (Theeuwes et al., 1991) orthogonally manipulating two variables in each experiment, including cognitive load (high vs. low) and stimulus salience (oddball vs. standard), and two variable across four experiments, including salience relevance (task-irrelevant in Experiments 1 and 2; task-relevant in Experiments 3 and 4) and category type (semantic-based in Experiment 1 and 3; prototype-based in Experiment 2 and 4). Participants in all experiments performed a majority categorization task, reporting which of two categories was in the majority among three simultaneously presented symbols. Building on prior evidence that semantic CAS relies on WM for rule maintenance (Wu et al., 2016 ; Wang, 2019) and prototype CAS relies on perceptual fluency (Lech et al., 2016 ; Wu et al., 2020 ), we formulated the following hypotheses: (1) High cognitive load will impair performance (slower RTs, lower accuracy) in both category types, but this effect will be especially pronounced for semantic CAS due to its dependence on limited WM resources for maintaining abstract rules. (2) A task-irrelevant oddball (a color singleton unrelated to the categories) will have little to no impact on performance for either category type, as participants’ category-based focus will filter out this bottom-up distraction. In contrast, a task-relevant oddball (a rare target-category trial) will capture attention and disrupt performance, particularly in the prototype-based task which lacks strong rule-based shielding and thus is more susceptible to bottom-up interference. (3) We anticipated an interaction between cognitive load and salience primarily in the semantic condition: under high load, the cost of an oddball (if task-relevant) would be amplified for semantic CAS, reflecting late-stage decision conflicts when WM is taxed. Prototype CAS, by comparison, might show more perceptual-level competition from oddballs regardless of load, given its reliance on sensory processing. In summary, our experimental design allowed us to test whether top-down load effects dominate CAS performance and whether any salience-driven attentional capture depends on both task relevance and category type. This work extends guided search theory (Wolfe, 2012 ) and intentional weighting theory (Memelink & Hommel, 2013) to categorical attention, offering new insights into how conceptual and perceptual hierarchies jointly shape real-world attention control. 2 Experiment 1 2.1 Methods 2.1.1 Participants Forty right-handed university students (31 female; M age = 21.83 years, SD = 2.16 years, range = 18–25 years) with normal or corrected-to-normal visual acuity were recruited through campus advertisements. The sample size was determined through a G*Power analysis, aiming for a statistical power of 0.80, with a significance level (α) set at 0.05 and a medium effect size (f = 0.25) in 2 × 2 within-subject ANOVAs (Cohen, 1992 ). All participants provided informed consent and received compensation. The protocol was approved by the institutional Ethics Committee. 2.1.2 Stimuli Visual stimuli were generated using E-Prime 2.0 on a Windows 7 workstation and displayed on a 17-inch LCD monitor (1024 × 768 resolution, 60 Hz refresh rate) with gamma-corrected gray background (CIE x/y coordinates: 0.313/0.329; luminance = 50 cd/m 2 ). Participants maintained 57 cm head position via chinrest, subtending 0.31°× 0.31°visual angle for all stimuli. Twelve equidistant positions (0.86°eccentricity) formed an invisible circle around central fixation. In each trial, three alphanumeric symbols were shown: either three targets (all same category) or two targets plus one distractor (mixed category). Target symbols were letters and distractor symbols were numbers, or vice versa (counterbalanced). Cognitive load was manipulated via this ratio: low-load trials had 3 targets and 0 distractors (3:0), whereas high-load trials had 2 targets and 1 distractor (2:1). Low-load displays are homogeneous and minimize WM demands, whereas high-load displays are heterogeneous and require active suppression of the incongruent item. Task-irrelevant salience was operationalized by a color oddball: all symbols were either red (CIE x/y: 0.640/0.330) or green (0.300/0.600), with one color occurring on 80% of trials (standard) and the alternate color on 20% of trials (oddball). Importantly, color was entirely unrelated to the category task (e.g., an odd-colored item could be either a target or distractor). The assignment of colors to “standard” vs. “oddball” was counterbalanced across blocks and participants. This probabilistic salience manipulation ensured the oddball item was unexpected but not more physically salient (in terms of luminance or contrast) than standard items. All other stimulus parameters (size, timing) were identical across conditions. 2.1.3 Experiment Design and Procedure Experiments employed a 2 (cognitive load: high vs. low) × 2 (stimulus salience: standard vs. oddball) within-subjects design. Each trial (Fig. 1 ) began with a central fixation cross for 100–600 ms (randomized), followed by a 400 ms presentation of the three-symbol array. Participants pressed one of two keys to indicate which category (letters or numbers) was in the majority. They had up to 1600 ms to respond, after which a blank inter-trial interval of 1500–2000 ms occurred. The experiment comprised 16 practice trials with feedback, then 400 experimental trials divided into 4 blocks of 100 trials (each block containing 80 standard-color trials and 20 oddball-color trials). Trials were self-paced between blocks. Accuracy (ACC) and reaction time (RT) were recorded. 2.2 Results We conducted a 2 (cognitive load: low vs. high) × 2 (stimulus salience: standard vs. oddball) repeated-measures ANOVA on accuracy and RT (Fig. 1 ). Accuracy We found a significant main effect of cognitive load, F (1, 39) = 74.59, p < 0.001, η² = 0.66, with accuracy lower under high load (90.5 ± 1.0%) than low load (96.7 ± 0.8%). No main effect of salience emerged, F (1,39) = 1.14, p = .29, nor load × salience interaction, F (1,39) = 0.008, p = .93. RT A significant main effect of cognitive load, F (1, 39) = 534.36, p < 0.001, η² = 0.93 showed that high load slowed responses (883.16 ± 24 ms high load vs. 703.48 ± 20.36 ms low load). There was no significant RT difference between standard vs. oddball trials ( F (1, 39) = 0.20, p = 0.66) nor interaction ( F (1, 39) = 0.13, p = 0.72). Inverse efficiency score (IES) a combined speed-accuracy measure (RT/ACC, Townsend & Ashby, 1983 ), and higher IES values indicate poorer performance. The analysis of IES also found a significant main effect of cognitive load, F (1, 39) = 571.65, p < 0.001, η² = 0.93, IES was higher (worse) under high load (983.43 ± 32.21) than low load (732.79 ± 25.77). There was no main effect of salience ( F (1, 39) = 0.21, p = 0.65), nor interaction ( F (1, 39) = 0.00, p = 0.99). Thus, in Experiment 1 the odd-colored distractor had no reliable effect on performance, whereas high cognitive load caused a large decrement in accuracy and speed. 2.3 Discussion of Experiment 1 Experiment 1 demonstrates that semantic-based CAS is dominated by top-down control demands, with little influence from task-irrelevant salience. The cognitive load manipulation (comparing 3:0 vs. 2:1 displays) robustly impaired performance: high-load trials led to substantially slower RTs (+ 179 ms on average) and lower accuracy (-6.2%) than low-load trials ( p < .001). These findings are consistent with models positing that semantic CAS relies on WM-mediated maintenance of category templates. In our task, increasing load (mixing categories) likely depleted the resources needed to uphold the task rule (“which category is majority”), in line with evidence that higher working memory demands slow rule-based decisions. Notably, the absence of any salience effect indicates that a task-irrelevant color oddball did not capture attention in this semantic search context. This null result contrasts with classic feature-search studies where an irrelevant color singleton can automatically capture gaze or attention (e.g., Theeuwes, 1992 ). In our study, participants’ categorical focus may have filtered out the odd-color item, or the oddball’s “surprise” value was simply insufficient to overcome the top-down set. In other words, there was no evidence of attentional capture by the irrelevant oddball, indicating the occurrence of category-level gating, although we caution that the null effect could also stem from the subtlety of our salience manipulation (see General Discussion). The dissociation between strong load effects and null salience effects in this experiment supports a competitive resource allocation view. That is, when attention is engaged in the relatively complex process of semantic categorization, little capacity remains for processing an irrelevant feature singletons (Wu et al., 2020 ). Consistent with this, neuroimaging work (Wu et al. 2015 ) shows that under high load, brain regions like the temporoparietal junction (TPJ) are occupied by task demands and then mediate attentional capture, reducing responsiveness to distractions. Consistent with this, neuroimaging study indicate that under high load, brain regions like the temporoparietal junction (TPJ), which are responsible for attentional engagement, are occupied by task demands, reducing responsiveness to distractions. Our results extend this idea by showing that even a potentially attention-grabbing color oddball (20% probability) failed to produce any interference when participants focused on the categorical rule under load. In summary, semantic CAS in our task operated as if protected by a top-down “filter” that prevented irrelevant singletons from affecting choice, at least when working memory was taxed. 3 Experiment 2 3.1 Methods The same 40 participants from Experiment 1 performed Experiment 2 after a ~ 1-hour break (to reduce carryover effects). The stimuli and design of Experiment 2 was identical to Experiment 1 except that the category task switched to prototype-based categories. Participants now judged the majority between “O” vs. “X” shapes, presented in 12 distinct font styles to introduce within-category visual variability (Fig. 2 ). The spatial layout (three items among 12 positions), timing, and color-probability salience manipulation (80% standard-color, 20% oddball-color trials) were the same as Experiment 1. Low-load trials had three shapes of one kind, and high-load trials had two of one kind and one of the other (2:1). This design isolates the effect of category type by holding the load and salience parameters constant while substituting semantic stimuli with perceptual shapes. Participants were instructed exactly as before (identify the majority category per trial). 3.2 Results A 2 (cognitive load: high vs. low) × 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig. 3 ). Accuracy A significant main effect of cognitive load emerged, F (1, 39) = 241.10, p < 0.001, η 2 = 0.86, with accuracy dropping from 97.03 ± 0.4% (low load) to 86.50 ± 1.0% (high load). There was no main effect of salience ( F (1, 39) = 0.25, p = 0.62) nor interaction ( F (1, 39) = 18, p = 0.68). RT A significant main effect of cognitive load ( F (1, 39) = 260.20, p < 0.001, η 2 = 0.87) showed that high load significantly slowed RTs (796.79 ± 21.51 ms high load vs. 670.42 ± 17.47 ms low load). No effect of salience was observed ( F (1, 39) = 0.001, p = 0.89), nor the interaction ( F (1, 39) = 0.14, p = 0.71). IES A significant main effect of cognitive load ( F (1, 39) = 457.91, p < 0.001, η 2 = 0.92) showed that high load increased IES (691 ± 18 low vs. 922 ± 23 high). No effect of stimulus salience ( F (1, 39) = 0.18, p = 0.68) nor the interaction ( F (1, 39) = 0.13, p = 0.72) was found. Thus, as in Experiment 1, cognitive load had a large impact on performance, whereas the irrelevant color oddball again produced no reliable slowdown or accuracy change. 3.3 Discussion of Experiment 2 Experiment 2 examined CAS under the same load and salience conditions, but for prototype-based categories. The results closely mirrored those of Experiment 1. High cognitive load led to substantially worse performance: under low load, participants responded faster (+ 126 ms advantage) and more accurately (+ 10.5% ACC) than under high load ( p < .001). This confirms that top-down control is a critical limiting factor for prototype-based CAS as well, extending the role of WM observed in semantic tasks to perceptual categorization (Wu et al., 2016 ; Wang, 2019). Importantly, we again found no effect of the task-irrelevant oddball, nor any load × salience interaction ( ps > .5). Even in the prototype search, where all stimuli were simple shapes, an odd-colored item did not measurably distract participants. This suggests that, under our conditions, task-irrelevant features failed to override category-level prioritization. In other words, participants allocated attentional resources preferentially to determining the predominant shape category, at the expense of processing an off-colored item that carried no task information. The absence of any salience cost in both Experiments 1 and 2 indicates a domain-general filtering mechanism: whether searching for letters or shapes, observers can ignore an irrelevant oddball when focused on a category judgment (Wu et al., 2020 ; Fan, 2014 ). In summary, Experiments 1–2, which both involved task-irrelevant salience, yielded strong and consistent cognitive load effects but null salience effects for both semantic and prototype CAS. This pattern raised the question of whether salience might influence CAS if it were made task-relevant, that is, if the oddball signaled a change in the target category rather than being a mere color anomaly. We next turned to that question in Experiments 3–4. 4 Experiment 3 The lack of salience effects in Experiments 1–2 (where oddball color was irrelevant to the task) suggested that stimulus salience modulates CAS only when aligned with task goals. Experiment 3 was designed to test this hypothesis by making the oddball task-relevant. Instead of color, we manipulated the probability of target category as the salience cue: one category occurred on 80% of trials and the alternative category on only 20% (the “oddball” category). Thus, the oddball in Experiment 3 was a rare target category trial, directly relevant to task performance. We used the semantic category set (letters vs. numbers) as in Experiment 1, under equivalent high vs. low load conditions. 4.1 Methods 4.1.1 Participants Another forty right-handed university students (30 female; M age = 20.73 years, SD = 2.15 years, range = 18–26 years) with normal or corrected-to-normal visual acuity were recruited through campus advertisements. Sample size and inclusion criteria matched those of Experiments 1–2. All provided consent and were compensated. Procedures were approved by the Ethics Committee. 4.1.2 Stimuli and Procedure Experiment 3 retained the 2 × 2 design of Experiment 1 (cognitive load × salience), but reconceptualized “salience” as a task-relevant category probability. The stimuli were letters and numbers as before. We assigned one category to be frequent (80% of trials, “standard”) and the other category to be rare (20% of trials, “oddball”) in the majority judgment task. For example, in one block, 80% of trials had a majority of letters (with occasional number-majority trials as oddballs); in another block, the probabilities were reversed (80% numbers as majority). Participants were informed about these base rates so that the oddball category trials would be unexpected but clearly tied to the task (they still had to identify the majority category on every trial). Importantly, color was now kept constant, yielding all stimuli were presented in a uniform color (e.g., all white) so that any salience effect would come purely from the rarity of the target category, not from a physical singleton. Load was also manipulated as before (3:0 vs. 2:1 ratio of letters:numbers). Timing, trial structure, and total number of trials were identical to Experiment 1. This design allowed a direct comparison with Experiment 1 to see how making the oddball task-relevant changes its impact. 4.2 Results In this semantic task with task-relevant oddballs, a 2 (cognitive load: high vs. low) × 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig. 4 ). Accuracy A significant main effect of cognitive load was observed, F (1, 38) = 180.74, p < 0.001, η² = 0.83, with lower accuracy under high load (81.5 ± 1.7%) than low load (90.9 ± 1.4%). Importantly, there was now a significant main effect of salience, F (1, 38) = 76.44, p < 0.001, η² = 0.67, with the accuracy on standard trials (94.3 ± 0.8%) was higher than on oddball trials (79.0 ± 2.3%). Furthermore, a significant load × salience interaction emerged, F (1, 38) = 84.48, p < .001, η² = 0.69. Simple effects analysis indicated that the oddball category impaired accuracy under both low-load ( F (1, 38) = 27.94, p < 0.001, η² = 0.42) and high-load ( F (1, 38) = 121.09, p < 0.001, η² = 0.76). Similarly, the load effect (higher accuracy under low-load relative to high-load) was observed for both standard ( F (1, 38) = 92.52, p < 0.001, η² = 0.71) and oddball stimuli ( F (1, 38) = 149.02, p < 0.001, η² = 0.80). RT : A significant main effect of cognitive load ( F (1, 38) = 156.72, p < 0.001, η² = 0.81) showed slower responses on high-load trials (631.22 ± 19.79 ms) than low-load (534.74 ± 15.40 ms). They were also slower on oddball trials (622 ± 17 ms) than standard trials (544 ± 18 ms), F (1, 38) = 135.09, p < 0.001, η² = 0.78. In contrast to accuracy, the RT load × salience interaction was not significant ( F (1, 38) = 1.90, p = 0.18), indicating that the oddball slowed responses by a similar amount in both load conditions. IES Both main effects were significant (main effect of cognitive load, F (1, 38) = 212.47, p < 0.001, η² = 0.85, better performance for low load (594.33 ± 16.84) than high load (808.64 ± 21.40); main effect of salience, F (1, 38) = 66.18, p < 0.001, η² = 0.64, better performance for standard trials (577.93 ± 18.16) than oddball trials (826.04 ± 27.71)). Importantly, the load × salience interaction was significant for IES, F (1, 38) = 51.30, p < 0.001, η² = 0.57. Simple effects analysis mirrored the accuracy interaction, with better performance under low-load relative to high-load in both standard ( F (1, 38) = 343.53, p < 0.001, η² = 0.90) and oddball trials ( F (1, 38) = 131.35, p < 0.001, η² = 0.78), and better performance for standard trials than oddball trials under both low-load ( F (1, 38) = 33.98, p < 0.001, η² = 0.47) and high load conditions ( F (1, 38) = 78.93, p < 0.001, η² = 0.68). Thus, combining speed and accuracy, the cost of oddball trials under high load was disproportionately large. In summary, Experiment 3 revealed that when the oddball was task-relevant (indicating a rare target category), it had a strong detrimental effect on performance, which is a stark contrast to the null results in Experiments 1–2. 4.3 Discussion of Experiment 3 Experiment 3 demonstrates that making the oddball task-relevant fundamentally changes its impact on CAS. In the semantic categorization task, introducing a rare-category oddball led to pronounced interference, especially under high load. Replicating prior experiments, we again saw robust cognitive load effects (participants were ~ 182 ms faster and 12.3% more accurate in low-load vs. high-load trials; p < .001). However, unlike Experiments 1–2, stimulus salience here significantly modulated performance. When an oddball trial occurred (e.g., a number-majority trial in a block where letters were usually majority), participants were slower and less accurate than on standard frequent-category trials. This oddball cost was substantial (overall ~ + 64 ms RT, -14.8% accuracy). Moreover, the cost was amplified under high load (as reflected by the significant interaction in accuracy and IES), indicating that resolving a rare-category occurrence was particularly difficult when working memory was already strained. This pattern contrasts sharply with the null salience effects in Experiments 1–2, highlighting that attentional capture by salience in CAS is highly task-contingent. In short, when the “oddball” is relevant to the search goal, it does capture attention and disrupt performance, but notably, mainly in situations of high cognitive demand. These findings refine our understanding of CAS by showing that goal relevance reconfigures the interplay between perceptual salience and conceptual processing. Under task-relevant conditions, the attentional system cannot simply ignore the oddball; instead, the oddball competes for processing resources, leading to slower and less accurate decisions. Conceptually, this suggests a dynamic allocation: when an oddball aligns with task goals (even as a rare event), the cognitive system prioritizes it, perhaps overly so, at the expense of routine processing, especially if executive resources are limited (high load). We will examine in the General Discussion whether this effect corresponds to a “late-stage” conflict (e.g., deciding that the oddball category is actually the majority) as opposed to an early attentional capture. For now, Experiment 3 confirms that salience can impact CAS when it matters to the task, whereas it was completely filtered out when it did not. 5 Experiment 4 5.1 Methods The same participants from Experiment 3 completed Experiment 4 after a 1-hour washout period to eliminate carryover effects. Experiment 4 extended the task-relevant salience design to the prototype-based categories (“O” vs. “X”). The majority-category probability manipulation (80% vs. 20%) was applied to “O” vs. “X” trials analogously to how letters vs. numbers were treated in Experiment 3. In other words, one shape category was made frequent and the other rare (counterbalanced). All aspects of the apparatus, timing, and load manipulation (3:0 vs. 2:1 shapes) mirrored those of Experiment 3. 5.2 Results A 2 (cognitive load: high vs. low) × 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig. 5 ). Accuracy A significant main effect of cognitive load was observed, F (1, 38) = 172.81, p < 0.001, η² = 0.82 (low load 92.7 ± 1.0% vs. high load 79.3 ± 1.4%), as well as salience, F (1, 38) = 52.86, p < 0.001, η² = 0.58 (standard trials 91.8 ± 0.7% vs. oddball trials 80.2 ± 1.8%). There was also a significant interaction, F (1, 38) = 15.26, p < 0.001, η² = 0.29. The post-hoc tests showed oddball accuracy was lower than standard accuracy under both low load ( F (1, 38) = 42.05, p < 0.001, η² = 0.53) and high-load ( F (1, 38) = 44.68, p < 0.001, η² = 0.54), and load effect for both standard ( F (1, 38) = 129.97, p < 0.001, η² = 0.77) and oddball trials ( F (1, 38) = 99.02, p < 0.001, η² = 0.72). RT High load slowed RTs (529 ± 15 ms low vs. 606 ± 20 ms high), F (1, 38) = 154.49, p < 0.001, η² = 0.80. Oddball trials were significantly slower (603.43 ± 17.44 ms) than standard trials (531.86 ± 18.21 ms), F (1, 38) = 158.53, p < 0.001, η² = 0.80. The interaction for RT was not significant, F (1, 38) = 0.49, p = 0.49. IES : Main effects of load ( F (1, 38) = 295.86, p < 0.001, η² = 0.89) and salience ( F (1, 38) = 85.37, p < 0.001, η² = 0.69) were significant (worse performance for high load (785.67 ± 21.55 vs. 574.75 ± 14.03 low load) and for oddballs (774.85 ± 18.69 vs. 585.57 ± 21.14 standard). Importantly, there was also a significant load × salience interaction, F (1, 38) = 24.25, p < 0.001, η² = 0.39. Simple effects analysis mirrored the accuracy analysis: oddball IES was higher than standard in both low-load ( F (1, 38) = 108.65, p < 0.001, η² = 0.74) and high-load conditions ( F (1, 38) = 62.39, p < 0.001, η² = 0.62), and high load IES was higher than low load in both standard ( F (1, 38) = 181.80, p < 0.001, η² = 0.83) and oddball trials ( F (1, 38) = 147.68, p < 0.001, η² = 0.80). In summary, Experiment 4 showed that for prototype CAS with task-relevant oddballs, we again see clear salience-driven interference (slower, less accurate responses on oddball trials) along with strong load effects. Unlike Exp 3, the salience effect in Exp 4 did not depend as strongly on load because of that it was present under both low and high load to a similar degree. 5.3 Discussion of Experiment 4 Experiment 4 revealed critical differences in how prototype-based CAS handles task-relevant salience, compared to semantic CAS in Experiment 3. Cognitive load effects persisted (participants were ~ 121 ms faster and 11.2% more accurate in low vs. high load, p < .001), indicating that even prototype search suffers under increased task demands. Notably, the salience interference was clearly present for prototype-based CAS: oddball trials induced substantial RT slowing (+ 89 ms on average) and accuracy declines (-18.3%) compared to standard trials. These costs were numerically larger than those observed in the semantic task of Experiment 3 (+ 64 ms RT, -14.7% ACC). This pattern suggests that prototype-based CAS is even more sensitive to salient deviations, perhaps because it relies on occipitotemporal perceptual binding processes that are directly impacted by oddball stimuli. In other words, when a rare target shape appeared, it may have strongly captured perceptual attention in the visual processing stream, competing with the template-matching process for the frequent target shape. Consistent with this idea, the interference by oddballs in prototype search can be seen as a form of “perceptual amplification” of salience effects. The results align with predictive coding accounts whereby the visual system of prototype-based CAS responds to an unexpected stimulus with greater disruption (here, an “oddball” shape violates the expected template, requiring additional processing). By contrast, semantic CAS might handle a rare oddball through higher-level rule evaluation (which we saw was especially taxing under high load, as evidenced by the load × salience interaction in Exp 3). Importantly, although prototype CAS showed greater absolute costs from oddballs, its overall performance remained faster and quite accurate even on oddball trials (e.g., ~ 603 ms RT and 80% ACC on oddball trials, versus semantic CAS’s ~ 622 ms and 79% in Exp 3). This reflects a theme that will be expanded upon: prototype processing trades some vulnerability to distraction for generally higher perceptual efficiency. In sum, Experiments 3–4 confirm that when salience is tied to the task (the oddball signals a rare target event), both semantic and prototype CAS are significantly disrupted by oddballs. The effect is present in both category types, but the magnitude and perhaps the nature of the interference differ: semantic CAS shows a larger relative cost under high cognitive load (pointing to load exacerbating decision-level conflict), whereas prototype CAS shows a robust salience effect regardless of load (consistent with more automatic, perceptual-level capture). We examine these differences directly in the combined analyses below. 6 Comprehensive analysis To directly compare effects across experiments, we performed a comprehensive ANOVA on the inverse efficiency scores (IES) with Experiment factors combined (Fig. 6 ). This mixed-design ANOVA included cognitive load (low vs. high), stimulus salience (standard vs. oddball), salience relevance (task-irrelevant: Exp 1–2 vs. task-relevant: Exp 3–4), and category type (semantic vs. prototype). Cognitive load, stimulus salience, and category were treated as within-subject (note: each pair of experiments shared participants, but salience relevance varied between the group that did Exp 1–2 and the group that did Exp 3–4, so we conservatively treat it as a between-subject factor). This analysis allowed us to assess: (1) differences between semantic and prototype tasks, (2) differences between task-irrelevant vs. task-relevant salience conditions, and (3) interactions among these factors (e.g., three-way interactions indicating differential integration of load and salience). Main effects We found a significant main effect of cognitive load , F (1, 77) = 1070.12, p < 0.001, η² = 0.93, with IES under low load (648.27 ± 12.19) lower (better performance) than that under high load (875.05 ± 15.57). A significant main effect of stimulus salience was also observed, F (1, 77) = 83.51, p < 0.001, η² = 0.52, with IES on standard trials (706.78 ± 14.48) lower (better performance) than that on oddball trials (816.33 ± 15.13). Category type showed a significant main effect as well, F (1, 77) = 6.77, p = 0.011, η² = 0.81, indicating that overall performance was better for prototype-based CAS (743.52 ± 13.26) than for semantic-based CAS (779.80 ± 16.98). This advantage for prototype categories echoes the raw results where prototype tasks often had faster responses for equivalent accuracy. Lastly, salience relevance (task-irrelevant vs. task-relevant conditions) had a significant main effect, F (1, 77) = 27.14, p < 0.001, η² = 0.26, with overall IES was lower in the task-relevant salience experiments (691.10 ± 19.28) than in the task-irrelevant experiments (832.22 ± 19.08). On face value, this might seem surprising (since one might expect performance to be worse when dealing with oddballs that matter), but it reflects that Exps 3–4 participants responded faster overall than Exps 1–2 participants. This difference likely arises because different participant groups were used and perhaps improved with practice by the time they did Exps 3–4; we interpret interaction effects below with caution for this between-group factor. Interactions There was a strong cognitive load × stimulus salience interaction, F (1, 77) = 47.74, p < 0.001, η² = 0.38. As already implied by individual experiments, load and salience each impaired performance, and their combination was especially detrimental. The oddball vs. standard difference in IES was significant under both low-load ( F (1, 77) = 60.65, p < 0.001, η² = 0.44) and high-load conditions ( F (1, 77) = 81.88, p < 0.001, η² = 0.52). Likewise, the load effect (high vs. low) was significant for both standard ( F (1, 77) = 1007.95, p < 0.001, η² = 0.93) and oddball trials ( F (1, 77) = 563.91, p < 0.001, η² = 0.88). However, the magnitude of the load effect was even larger for standard trials than oddball trials (since performance was already quite degraded on oddball-high-load trials, leaving slightly less additional room for load to worsen it). This pattern aligns with the notion that having both high load and an oddball creates a ceiling of difficulty, where performance is pushed to a floor. We also found a significant category type × stimulus salience interaction, F (1, 77) = 5.24, p = 0.025, η² = 0.06. Interestingly, although both semantic ( F (1, 77) = 63.23, p < 0.001, η² = 0.45) and prototype tasks ( F (1, 77) = 80.07, p < 0.001, η² = 0.51) showed worse performance on oddball trials than standard trials, prototype-based CAS was overall more efficient than semantic CAS specifically on oddball trials ( F (1, 77) = 9.43, p = 0.003, η² = 0.11). This indicates that prototypes had a performance edge even when dealing with oddballs (likely due to generally faster processing), despite the fact that in absolute terms they suffered a larger slowing from oddballs. In essence, prototype CAS was relatively resilient in final outcome, even though oddballs did disrupt it. There was also a category type × salience relevance × salience (3-way) interaction, F (1, 77) = 7.90, p = 0.006, η² = 0.09. This reflected subtle differences in how category type mattered across the irrelevant vs. relevant experiments. In the standard-stimulus condition, making salience task-relevant reliably improved efficiency for both category systems: in the semantic task, task-relevant IES was significantly lower than its task-irrelevant counterpart, F (1, 77) = 67.26, p < 0.001, η² = 0.47; the same pattern held for the prototype task, F (1, 77) = 54.17, p < 0.001, η² = 0.41. Nevertheless, the prototype pathway still maintained a speed–accuracy edge in most situations: IES for prototype-based categories was significantly lower than for semantic categories in task-relevant/standard trials ( F (1, 77) = 7.49, p = 0.008, η² = 0.09), task-relevant/oddball trials ( F (1, 77) = 4.45, p = 0.038, η² = 0.06), and even task-irrelevant/oddball trials ( F (1, 77) = 4.98, p = 0.028, η² = 0.06). Within each category system, oddballs imposed a clear cost whenever they were task-relevant, F (1, 77) = 127.95, p < 0.001, η² = 0.62. Taken together, task relevance globally boosts processing efficiency in both CAS pathways, yet the prototype route retains a better speed-accuracy trade-off across all salience contexts, whereas the semantic route suffers the greatest performance drop when a task-relevant oddball (especially under high load) is added, underscoring fundamental differences in how the two pathways allocate resources and integrate salience. Lastly, the 3-way interaction of load × salience relevance × salience was significant, F (1, 77) = 44.70, p < 0.001, η² = 0.37. Under every combination of salience relevance and stimulus salience, low load produced markedly lower IES (better performance) than high load: task-irrelevant/standard, F (1, 77) = 829.39, p < 0.001, η² = 0.92, task-irrelevant/oddball, F (1, 77) = 235.92, p < 0.001, η² = 0.75, task-relevant/standard, F (1, 77) = 261.81, p < 0.001, η² = 0.77, and task-relevant/ oddball, F (1, 77) = 331.47, p < 0.001, η² = 0.81. Turning to salience relevance, the overall performance in task-relevant experiments (Exp. 3–4) was better than task-irrelevant experiments (Exp. 1–2) in low load/standard, F (1, 77) = 60.52, p < 0.001, η² = 0.44, and high load/standard, F (1, 77) = 83.52, p < 0.001, η² = 0.52, and low load/oddball trials, F (1, 77) = 4.50, p = 0.037, η² = 0.06. Finally, the disadvantages of oddball trials were found in task-relevant experiments, no matter in high load ( F (1, 77) = 121.93, p < 0.001, η² = 0.61) or low-load conditions ( F (1, 77) = 160.31, p < 0.001, η² = 0.68). In summary, the comprehensive analysis statistically supports conclusions drawn from individual experiments: (a) cognitive load had a strong deleterious effect regardless of category type or salience relevance; (b) stimulus oddballs significantly affected performance only in the task-relevant context; (c) prototype-based CAS was overall more efficient than semantic CAS (lower IES), including during oddball trials; (d) the worst performance occurred when both high load and a task-relevant oddball were present (especially for semantic CAS). 7 General Discussion This study systematically dissected how top-down cognitive control and stimulus salience jointly shape category-guided attentional selection (CAS) across semantic and prototype-based systems through four behavioral experiments. By orthogonally manipulating cognitive load (high vs. low), stimulus salience (oddball vs. standard), salience relevance (task-irrelevant vs. task-relevant), and category architecture (semantic: alphanumeric characters; prototype: "O/X" shapes), we identified three principal mechanisms. First, cognitive load exerted a domain-general effect, with high-load conditions impairing accuracy (6.2–10.5%) and prolonging reaction times (126–179 ms) across both semantic and prototype tasks. This is consistent with the notion that maintaining multiple items or rules in working memory under high load consumes attentional resources needed for CAS (Wu et al., 2015 ; Wang, 2019). Second, stimulus salience affected performance exclusively when it was task-relevant: in our task-irrelevant conditions (Exp 1–2), oddball stimuli produced no significant interference, whereas in task-relevant conditions (Exp 3–4), rare-category oddballs incurred clear performance costs (RT delays of ~ 64–89 ms and accuracy drops of ~ 15–18%). This finding demonstrates goal-contingent prioritization and aligns with recent evidence that salience matters only when it matches current goals, while challenging classical views of automatic capture by salient distractors regardless of goals. Third, prototype-based CAS showed overall superior efficiency on oddball trials compared to semantic CAS (approximately a 12.3% IES advantage). We attribute this to perceptual similarity-driven guidance in the prototype task, which may allow faster grouping and processing of stimuli, effectively bypassing the need for slow rule retrieval that semantic tasks require (Lech et al., 2016 ; Wu et al., 2020 ). Importantly, the interactions between load and salience provide additional insight: under high load, semantic CAS faced amplified conflict resolution demands (evidenced by disproportionately slowed decisions and reduced accuracy when a task-relevant oddball occurred), suggesting that the bottleneck was at a late decision stage (e.g., making the category judgment with conflicting evidence). In contrast, prototype CAS under high load exhibited signs of early-stage perceptual competition when oddballs appeared, that is, the interference was manifest in RT costs that did not strongly depend on load, implying that even with ample resources, a perceptually salient oddball engages early visual attention in the prototype task. Together, these findings support a dual-pathway account: semantic CAS primarily operates via conceptual, rule-based control susceptible to WM load, whereas prototype CAS operates via perceptual grouping and is thus faster and somewhat more “bottom-up” albeit still subject to strategic control. 7.1 Dominance of Top-Down Control in CAS Across all experiments, cognitive load was the dominant factor limiting CAS performance. High-load trials (requiring integration of a distractor category) consistently produced worse outcomes in both accuracy and RT. This underscores that WM capacity is a domain-general bottleneck for CAS. Our findings align with models suggesting attentional selection efficiency hinges on maintaining attentional templates in WM (Fan, 2014 ; Wu et al., 2015 ). When load increased (2:1 vs. 3:0), participants had to compare and suppress multiple category representations, leaving fewer resources available for timely target selection. where increased cognitive load depletes resources critical for sustaining category-specific representations. Notably, this bottleneck manifests differently across category architectures: semantic-based CAS, reliant on rule retrieval from long-term memory (Giammarco et al., 2016 ), exhibited delayed response thresholds under high load, reflecting protracted conflict resolution in prefrontal-parietal networks (Wang, 2019), whereas prototype-based CAS, driven by perceptual similarity gradients (Lech et al., 2016 ), showed attenuated accuracy declines, suggesting perceptual fluency compensates through feature binding (Wu et al., 2020 ). The three-way interaction (category × load × salience) further confirms this dissociation, with semantic systems prioritize abstract rule arbitration under resource competition, while prototype systems exploit perceptual fluency to mitigate WM demands. A dual-pathway CAS model interprets this divergence: semantic processing engages prefrontal-parietal networks for hierarchical control, while prototype processing utilizes occipitotemporal circuits for sensory optimization (Wu et al., 2016 ; Lech et al., 2016 ; Wu et al., 2015 ; Wu et al., 2020 ). Overall, these findings extend Wu et al.’s ( 2015 ) framework of temporoparietal junction-mediated control integration to categorical contexts, demonstrating that cognitive load amplifies competition between category templates and salient distractors across architectures. This hierarchical constraint aligns with capacity-limited models of cognitive control (Fan, 2014 ; Wu et al., 2018, 2019), where CAS efficiency reflects dynamic trade-offs between rule precision and perceptual integration, rather than mere feature-level competition. 7.2 Task-Contingent Salience Effects on CAS In contrast to the ever-present load effects, stimulus salience influenced CAS only under specific conditions, namely, when the salient feature was tied to the task. This finding supports the idea of goal-contingent attentional capture (Zhang & Gaspelin, 2024 ). In Experiments 1–2 (task-irrelevant oddballs), we saw no hint of the typical distraction caused by an odd-colored item; participants effectively filtered out the oddball color, focusing on category identity. This aligns with studies showing that when observers are engaged in a demanding task, irrelevant singletons often fail to capture attention (especially if they fall outside the “attentional set”). Our results go further by demonstrating such filtering in a categorical context: a color singleton that had no bearing on the category decision was behaviorally suppressed. Neuroimaging evidence concurs that task-irrelevant salient stimuli mainly activate visual and parietal regions (Clark et al., 2000 ; Downar et al., 2001 ; Indovina & Macaluso, 2006 ) but do not engage the frontal executive network (Fockert et al., 2004 ). We infer that in our task-irrelevant conditions, the oddball color’s signal likely remained confined to low-level visual areas and was prevented from influencing the decision process, essentially, category-based focusing provided “protective gating” against the irrelevant oddball. Conversely, when the oddball carried categorical significance (Exps 3–4), it reliably disrupted performance: participants could not ignore a salient event that was directly tied to target identity. Under task-relevant conditions, rare-category oddballs drew attention and processing priority, consistent with the “attentional control settings” being tuned to that category. The goal-contingent capture framework (Zhang & Gaspelin, 2024 ) predicts that salient stimuli capture attention only if they match the observer’s active goal. Our data strongly support this principle in CAS. Indeed, the absence of an oddball effect in Exp. 1–2 versus its presence in Exp. 3–4 created a significant salience relevance × salience interaction in the omnibus analysis. In practical terms, CAS is highly adaptive: it ignores irrelevant singletons but prioritizes (to a fault) unexpected targets. It is important to note that the lack of capture by irrelevant oddballs in our study does not mean salience is never potent, instead, it likely reflects the nature of our salience manipulation and the strength of top-down focus (Folk et al., 1992 ; Lavie, 2005 ). In classic additional-singleton paradigms (Theeuwes, 1992 ), a salient color distractor in a visual search can capture attention involuntarily. Why did our color oddball not capture attention? We suspect two reasons: first, our task was a coarse categorical judgment that might not require shifting attention to individual items, so the oddball distractor might have been processed superficially without drawing attention away from the overall category count. Second, our oddball was defined by probability across trials rather than a unique singleton in a display, so that, on any given trial, the odd-colored item was not actually more salient in a bottom-up sense than the others (all items were isoluminant, just different in color). Instead, it was the expectation violation over trials that constituted “salience”, consistent with predictive-coding accounts of attention (Summerfield & Egner, 2009 ). Such probabilistic oddballs are known to have weaker and more delayed effects compared to outright singletons (Biggs et al., 2014 ). They can influence decision criteria or arousal, but they do not automatically capture spatial attention the way an abrupt onset or high-contrast singleton might. In our task-irrelevant case, apparently neither mechanism (arousal nor spatial capture) was sufficient to impact behavior. Thus, we interpret our null oddball findings in Exp. 1–2 as evidence that categorical goals override subtle salience signals, but we also acknowledge that a more salient distractor (e.g., a flashing item or a uniquely shaped singleton) might have produced a different result. Supporting this, Wu et al. ( 2015 ) did find attentional capture by an oddball in a feature-based task (they used a spatial cueing design with color singletons), but there the salient distractor likely shared some task relevance (e.g., a directional-cue context). Our findings diverged from Wu et al. ( 2015 ) precisely in that our distractors had no spatial or response relevance, allowing category-based focus to suppress them. When salience was task-relevant in our study, the effects on CAS were unambiguous: performance suffered significantly on oddball trials. Participants took longer and were more error-prone when the majority category was the rare one. This can be interpreted as a form of “oddball cost” due to violated expectations, where observers were biased toward the frequent category and had to overcome this bias on oddball trials, leading to delays and mistakes (Biggs et al., 2014 ). The effect was akin to a contingent capture: attention might have momentarily been captured by the oddball item (e.g., focusing on it or re-checking it), or decision thresholds may have been adjusted (e.g., requiring more evidence on oddball trials due to lower prior probability, which could increase RT and errors). Neurocognitively, task-relevant oddballs engage frontal regions such as the anterior cingulate cortex that are implicated in conflict monitoring and cognitive control (Botvinick et al., 2004 ). This suggests our participants, upon encountering an oddball trial, had to invoke additional control processing (e.g., “Is this trial really an odd category?”) resulting in slower, less accurate responses. Interestingly, our data showed this oddball cost was generally larger in prototype CAS than semantic CAS (absolute RT/ACC differences). We discuss this further below, but it implies that even though prototype tasks are faster overall, they might experience a proportionally greater disruption from an unexpected event, perhaps because they rely on a more stimulus-driven mode of processing that is directly perturbed by any irregular input. From a theoretical standpoint, our results emphasize that attentional selection in CAS is neither purely stimulus-driven nor purely goal-driven, but a conditional combination. Under high cognitive load, even typically “automatic” capture by salient stimuli can be eliminated, in line with load theory’s prediction that limited central resources curtail distractor processing (Lavie, 2005 ; de Fockert et al., 2001 ). Under normal load but with strong top-down goals, capture is also prevented, consonant with contingent-capture findings that attention prioritises goal-matching features (Folk et al., 1992 ; Bacon & Egeth, 1994 ). Only when the salience coincides with an active goal do we see a full effect. Thus, CAS appears to implement a hierarchical priority map (cf. Wolfe, 2012 ) where task goals gate the influence of salient signals. In our semantic CAS, this meant that only when the oddball was categorically relevant did it make it onto the “priority map” to compete for attention. Furthermore, our data provide insight into the stages of processing affected by salience in CAS. The semantic task under relevant oddball showed an interaction with load primarily in accuracy (not RT), suggesting that oddballs under high load caused participants to miss targets or misjudge the majority more often, consistent with a late-stage decision error or threshold adjustment, often indexed by the P3 component (Polich, 2007 ). Prototype task under relevant oddball showed a robust RT cost without an interaction with load, implying a more uniform perceptual slow-down due to oddballs; such effects are typically associated with extended N2pc or target-selection activity (Luck & Hillyard, 1994 ; Eimer, 1996 ). We cautiously interpret that task-relevant salience triggers late decision-stage conflict in semantic CAS, but earlier perceptual-stage competition in prototype CAS. We stress that our behavioural data cannot directly pinpoint neural timing, yet this inference accords with the differing accuracy patterns observed in the two cases. In any event, salience integration in CAS remains highly context-dependent, a theme elaborated in the next section. 7.3 Integrative Effects of Top-Down Control and Salience on CAS One of our motivating questions was whether top-down control and salience operate independently or interactively in CAS. Our experiments indicate strong interactive integration: cognitive load (top-down demand) and stimulus salience did not simply have additive effects; instead, they jointly determined performance in important ways (Lavie, 2005 ; de Fockert et al., 2001 ). Specifically, under task-relevant salience conditions, we observed that high cognitive load exacerbated the negative impact of oddballs, especially in the semantic task. When the rare-category oddball appeared and participants were already under strain (high load, juggling two categories in mind), the result was a pronounced conflict, which reflected in markedly increased decision time and errors. In our comprehensive analysis, this showed up as a three-way interaction: the difference between oddball and standard trials was greatest in the high-load, task-relevant, semantic condition. This finding can be framed as top-down and bottom-up factors competing for shared resources (Fan, 2014 ). Under easier conditions (low load), participants had the capacity to deal with an oddball fairly well (RT cost but small accuracy impact). Under maximal demand (high load), that oddball tipped performance over the edge, causing participants to struggle (slow and error-prone responses). This aligns with models where central resources (like WM or decision-making circuits) mediate between goal-directed and stimulus-driven inputs (Duncan & Humphreys, 1989 ). Our data resonate with Fan’s ( 2014 ) information-theoretic account and others, wherein cognitive control under uncertainty must arbitrate between internal goals and external salient events. In our case, high load amplified this arbitration conflict, particularly when an oddball signaled a deviation from the expected rule. Conversely, in task-irrelevant conditions, no such integration or competition occurred: load effects were present (because that’s internally driven), but salience did nothing and thus there was no additional conflict. This emphasizes that the cognitive system can effectively segregate an irrelevant source of salience, preventing it from draining resources when focus is required on something else. Our findings here dovetail with Lavie’s load theory (Fockert et al., 2004 ), which posits that under high cognitive load, processing of irrelevant stimuli is reduced. We observed that even under low load, irrelevant oddballs didn’t matter. This likely because participants strategically ignored color from the start (given it was never useful), illustrating a strong top-down set (Folk et al., 1992 ). Under high load, that top-down filtering may have been even more absolute (as attention was fully occupied). Thus, when salience was irrelevant, top-down control effectively nullified any integration. This indicates that the two operated on separate tracks, with category decisions proceeding unaffected by oddball presence. Our results also shed light on how category architecture (semantic vs. prototype) influences this integration. We found that semantic-based CAS and prototype-based CAS, despite similar overall trends, had quantitative and qualitative differences in their responses to load + salience. Semantic CAS under high load exhibited what we interpret as conceptual interference, needing to resolve conflicting category cues (majority vs. oddball minority) perhaps via executive processes (hence more errors when taxed). Prototype CAS under high load showed a slightly different pattern: oddly, the presence of high load reduced the relative cost of oddballs in some measures (e.g., the interaction in accuracy was smaller in Exp 4 than Exp 3). This hints that under high load, prototype searchers might have inadvertently filtered some distractors, even possibly the oddball shape, by focusing on the dominant perceptual features (Lech et al., 2016 ). Another way to view it is through perceptual-load theory: when prototype displays were heterogeneous (high load), visual attention was likely more occupied, leaving less room for the oddball shape to capture attention. This mechanism analogous to high perceptual load reducing distractibility (Lavie, 2005 ). Indeed, we speculated earlier that in Exp 4, high load didn’t increase oddball cost; intriguingly, it slightly decreased the oddball effect on accuracy relative to low load (the prototype oddball effect was big at both loads, but not bigger at high load). This could reflect that participants, when confronted with a difficult prototype array (2 : 1 with a rare shape), sometimes missed the oddball shape entirely or treated it as just another distractor, whereas in easier arrays (low load), the oddball shape stood out more and thus interfered relatively more. In any case, category architecture modulated how load and salience interplay. Semantic tasks brought the conflict to a head at decision-making (leading to late-stage interference under combined strain), whereas prototype tasks handled it in a more distributed, perceptual manner (steady interference that didn’t double-dip with load). This distinction mirrors Ashby et al.’s (2004, 2005) dual-process account where rule-based systems (semantic CAS) rely on WM and executive control, thereby two sources of demand (load and oddball) overload the system. Whereas, similarity-based systems (prototype CAS) rely more on visual processing, thereby they can maintain performance speed but will show interference in accuracy that doesn’t scale massively with cognitive load. Finally, by extending Memelink & Hommel’s (2013) intentional-weighting theory to our results, we can say that the “weighting” of task features in CAS is highly flexible. When the feature “category frequency” was irrelevant, it was weighted at zero (oddball frequency ignored). When it became relevant, that feature’s weight shot up and influenced both perception and decision. CAS thus dynamically gates incoming information based on task demands, balancing conceptual precision and perceptual efficiency as needed. In high-load semantic CAS, the system favored conceptual precision (hence ignoring oddballs until it perhaps was too late, causing conflicts); in prototype CAS, the system leaned on perceptual processes (hence oddballs always intruded a bit, but decisions were fast). This adaptive gating underscores the context-dependent nature of attentional selection in categorical tasks. 7.4 Dual-Pathway CAS Model Involving Conceptual and Perceptual Hierarchies Bringing these findings together, we propose a dual-pathway model for CAS that reflects two hierarchies: a conceptual hierarchy (semantic CAS) and a perceptual hierarchy (prototype CAS). Our results provide direct evidence that prototype-based CAS can systematically outperform semantic-based CAS across varying conditions (Lech et al., 2016 ; Wu et al., 2020 ). This was evident in the main effect of category type on IES (prototypes had overall lower IES) and in their higher resilience in absolute performance even when oddballs and load were present. This supports the idea that prototype categories tap into fast, efficient visual processing strategies (a “perceptual pathway”) that do not require heavy WM involvement (Giammarco et al., 2016 ). On the other hand, semantic categories rely on a “conceptual pathway” that uses learned rules and is tied to WM and long-term memory retrieval (Fan, 2014 ). This makes semantic CAS more accurate but slower, and more vulnerable to WM depletion. For example, in our experiments the semantic tasks maintained higher accuracy under oddball conditions by incurring an RT cost (especially under high load), whereas the prototype tasks tended to respond quickly at the expense of some accuracy. This trade-off aligns with the notion that semantic CAS prioritizes rule-based precision (ensuring the category decision is correct, even if slower), while prototype CAS prioritizes speed/efficiency (leveraging visual similarities to respond quickly, even if occasionally an oddball leads to an error) (Ashby & Maddox, 2004 ). The dual-pathway model is supported by the three-way interaction of category × load × salience in our data. It extends Ashby’s dual-system theory (rule vs. exemplar) to attentional selection: under resource competition, semantic (rule-based) systems slow down to maintain accuracy, whereas prototype (similarity-based) systems sacrifice a bit of accuracy to maintain speed (Ashby & Valentin, 2005 ). We observed exactly that pattern. This divergence likely stems from the underlying neural circuits: semantic CAS engages prefrontal-parietal circuits (associated with executive control and sustained rule maintenance), which have limited capacity (Miller & Cohen, 2001 ). Once that capacity is strained (by load or dual-tasking), performance slows and any additional conflict (like an oddball) further taxes the system. Prototype CAS, conversely, relies on occipitotemporal circuits (associated with visual pattern recognition and grouping), which operate in parallel and with high efficiency (Ungerleider & Haxby, 1994 ). These circuits can rapidly integrate features and are less disrupted by multitasking until a certain point; thus, performance remains quick, though not immune to errors if conflicting inputs (like an oddball shape) are present. Indeed, evidence of occipital alpha-band suppressions unique to prototype tasks (suggesting heightened visual processing) would be a neural marker to confirm this pathway (Worden et al., 2000 ). Crucially, our model posits that semantic CAS and prototype CAS are not just quantitatively different, but qualitatively employ distinct control strategies. Semantic CAS behaves like a “late selection” system and it filters and decides at a later stage, allowing it to ignore irrelevant salience entirely but struggling if multiple demands coincide (late-stage conflict) (Luck & Vogel, 1997). Prototype CAS behaves more like an “early selection” system, and it integrates all perceptual inputs (including oddballs) early on, leading to immediate competition but perhaps resolving it at a sensory level (early competition resolved via perceptual coding) (Eimer, 1996 ). This dichotomy extends Wolfe’s guided search: semantic CAS adds a top-down rule-based “priority map” in prefrontal cortex, whereas prototype CAS leans on feature-based “maps” in visual cortex (Wolfe & Horowitz, 2017 ). Under our dual-pathway model, when faced with distractions or load, the semantic pathway will enforce its priorities through frontal control (slowing things down but keeping on task), and the prototype pathway will continue to leverage fast sensory processing (maintaining speed but letting in some distraction). Neither pathway is inherently superior; each excels under certain conditions. Semantic CAS might dominate in situations requiring high accuracy and when distractors can be categorically excluded (since it can completely ignore irrelevant stimuli given enough focus). Prototype CAS might excel in visually complex scenes or when speed is essential, as it can quickly home in on target-like features. 7.5 Strengths, Limitations and implications This study provides a novel, integrated perspective by directly comparing semantic and prototype-based attention under systematically varied load and salience conditions. We demonstrated a clear goal-dependence of salience effects in CAS and empirically distinguished two modes of category-guided attention, thereby resolving to some extent the debate on whether CAS is more like feature-based attention (perceptual) or like executive control (conceptual), and our findings suggest it can be both, depending on category type and context (Folk et al., 1992 ; Zhang & Gaspelin, 2024 ). The dual-pathway model emerging from our results offers a framework for future research on how different category representations (rule-based vs. similarity-based) recruit distinct neural resources and cope with distractions (Ashby & Valentin, 2005 ). Our use of multiple experiments and a combined analysis is a strength, as it allowed us to see consistent patterns and cross-experiment interactions that a single experiment might miss (Efron & Tibshirani, 1993 ). We also introduce an important methodological point: manipulating salience via probability (oddball frequency) in a CAS task is a subtle but insightful way to study contingent capture in a categorical domain, complementing traditional singleton methods (Biggs, Adamo, & Mitroff, 2014 ). However, several limitations should be acknowledged. First, while our salience manipulation via color probability provided a novel angle, it deviates from classic definitions of “physical salience”. As discussed, a uniquely colored item in a display can capture attention, whereas an oddball color event across trials may not, and our interpretation of “no capture = top-down filtering” rests on the assumption that our oddball was a meaningful bottom-up cue (Theeuwes, 1992 ). It might be argued that the oddball color simply wasn’t salient enough; thus, our conclusion of protective gating in semantic CAS should be tempered by that consideration. Future work could incorporate a true singleton distractor in a category search to confirm that semantic guidance can overcome even strong singletons (Zhang & Gaspelin, 2024 suggests it can under certain conditions). Second, our cognitive-load manipulation was labeled as such because of prior literature and our intention (more items to hold/compare = higher load). Nevertheless, as reviewers pointed out, this “load” may also reflect decision difficulty or perceptual load, not purely working-memory load (Lavie, 2005 ). In a 2 : 1 display, one must scrutinize the stimuli more carefully to identify the majority, thereby engage perceptual processes as much as WM. Thus, while we interpret our load effects in terms of WM depletion, an alternative view is that high-load trials increased task difficulty overall, and our results do not exclusively pinpoint WM (de Fockert et al., 2001 ). Future studies could separate memory load (e.g., requiring an actual memory set to hold) from decision difficulty to see if the effects differ. Our use of the term “cognitive load” should be understood as the combined challenge of a heterogeneous display requiring suppression of an oddball item, which indeed could involve both WM (keeping the category rule in mind) and additional decision effort. We have reframed some of our interpretations accordingly in this revision. It’s also worth noting that our interpretations of “early” vs. “late” stage processing differences between tasks are inferences based on behavioural patterns and existing literature. Without direct neurophysiological measures (EEG/ERP or eye-tracking), we cannot conclusively state at what processing stage the load–salience interactions occur for each category type. Future experiments could record ERPs in a similar paradigm to see if, for example, the N2pc (an index of early attentional allocation) is modulated by oddballs more in prototype tasks, whereas the P3 (decision-related) is more modulated in semantic tasks (Luck, 2014 ; Polich, 2007 ). Our discussion of this point is meant to generate hypotheses rather than final answers. Despite these limitations, our study has important implications. From a theoretical standpoint, it reconciles prior mixed results on attentional capture in categorical search by showing that task context (irrelevant vs. relevant) is key (Theeuwes, 2010 ). It also bridges literature on visual search and working memory by examining their interplay in CAS. Practically, understanding how different kinds of categories handle distraction can inform user-interface design or training protocols. For example, if one needs to design a display to grab a person’s attention during a task, our results suggest that the cue must be relevant to their task or else it may be ignored; conversely, if one wants to minimise distraction, ensuring that salient visual features are unlinked to the user’s goals will help them stay focused (Wickens & McCarley, 2008 ). In applied settings such as airport baggage screening (a semantic category search) or medical image analysis (often more prototype-like, searching for anomalous shapes), different strategies might be required to mitigate distractions or manage cognitive load. In conclusion, category-guided attention operates via dual pathways that differentially integrate top-down and bottom-up influences. Semantic CAS functions like a focused spotlight, governed by WM and executive control, highly effective at excluding irrelevant stimuli but at risk of overload when multitasking. Prototype CAS functions like a wide-angle filter, rapidly responsive to visual features and robust in speed, yet allowing more bottom-up intrusion. Understanding these dual pathways enriches our comprehension of attention in complex, real-world tasks and could guide the development of techniques to improve attention, for example through adaptive interfaces or training that either bolsters rule maintenance or harnesses perceptual fluency as appropriate. Declarations Acknowledgments This work was supported by National Education Scientific Planning Project (DBA230368). Declaration of Interest statement The authors declare no conflict of interest. Author contributions X. W., H. Z., and X. S. conceived the study; X. W., Y. L and X. S. analyzed the data. All authors discussed the results and contributed to the writing of the article. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work the authors used Deepseek (https://www.deepseek.com/) in order to improve the readability and language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. Open practices statement Data, scripts, materials, and analyses are available at https://osf.io/6wvqc/ Correspondence Huan Zhang, Faculty of Psychology, Tianjin Normal University, Tianjin, China. Email: [email protected] References Alvarez, G. A., & Cavanagh, P. (2004). The capacity of visual short-term memory is set both by visual information load and by number of objects. Psychological Science , 15 (2), 106–111. Ashby, F. G., Maddox, W. T., & Bohil, C. J. (2002). Observational versus feedback training in rule-based and information-integration category learning. Memory & Cognition , 30 (5), 666–677. Ashby, F. G., & Maddox, W. T. (2004). Human category learning. Annual Review of Psychology , 56 (1), 149–178. Ashby, F. G., & Valentin, V. V. (2005). Multiple systems of perceptual category learning: theory and cognitive tests. Handbook of Categorization in Cognitive Science , 547–572. Bacon, W. F., & Egeth, H. E. (1994). Overriding stimulus-driven attentional capture. Perception & Psychophysics , 55 , 485–496. Biggs, A. T., Adamo, S. H., & Mitroff, S. R. (2014). Rare, but obviously there: effects of target frequency and salience on visual-search accuracy. Acta Psychologica , 152 , 158–165. Botvinick, M. M., Cohen, J. D., & Carter, C. S. (2004). Conflict monitoring and anterior cingulate cortex: an update. Trends in Cognitive Sciences , 8 , 539–546. Chen, X., Xu, B., Chen, Y., Zeng, X., Zhang, Y., & Fu, S. (2023). Saliency affects attentional capture and suppression of abrupt-onset and color singleton distractors: Evidence from event‐related potential studies. Psychophysiology , 60(8), e14290. Chu, H., & Zhou, X. (2004). Attentional capture and top-down processes. Advances in Psychological Science , 12 (5), 680–687. Clark, V. P., Fannon, S., Lai, S., Benson, R., & Bauer, L. (2000). Responses to rare visual target and distractor stimuli using event-related fmri. Journal of Neurophysiology , 83 (5), 3133–3139. Cohen, J. (1992). A power primer. Psychological Bulletin , 112 (1), 155–159. de Fockert, J. W., Rees, G., Frith, C. D., & Lavie, N. (2001). The role of working memory in visual selective attention. Science , 291 , 1803–1806. Downar, J., Crawley, A. P., Mikulis, D. J., & Davis, K. D. (2001). The effect of task relevance on the cortical response to changes in visual and auditory stimuli: an event-related fMRI study. NeuroImage , 14( 6),1256–1267. Duncan, J., & Humphreys, G. W. (1989). Visual search and stimulus similarity. Psychological Review , 96 (3), 433–458. Efron, B., & Tibshirani, R. J. (1993). An Introduction to the Bootstrap . Chapman & Hall/CRC. Eimer, M. (1996). The N2pc component as an indicator of attentional selective processing. Electroencephalography and Clinical Neurophysiology , 99 , 225–234. Ell, S. W., & Ashby, F. G. (2004). Dynamical trajectories in category learning. Perception & Psychophysics , 66 (8), 1318–1340. Fan, J., Guise, K. G., Xun, L., Wang, H., & Jan, L. (2008). Searching for the majority: algorithms of voluntary control. Plos One , 3(10), e3522. Fan, J. (2014). An information theory account of cognitive control. Frontiers in Human Neuroscience , 8 , 680. Fockert, J. D., Rees, G., Frith, C., & Lavie, N. (2004). Neural correlates of attentional capture in visual search. Journal of Cognitive Neuroscience , 16 (5), 751–759. Folk, C. L., Remington, R. W., & Johnston, J. C. (1992). Involuntary covert orienting is contingent on attentional control settings. Journal of Experimental Psychology: Human Perception and Performance , 18 , 1030–1044. Giammarco, M., Schmidtke, J., & Arnell, K. (2016). Dissociating semantic and perceptual contributions to the suppression of distractors during working memory maintenance. Journal of Experimental Psychology: Human Perception and Performance , 42 , 1550–1562. Goldstone, R. L. (1998). Perceptual learning. Annual Review of Psychology , 49 , 585–612. He, H., Chen, Y., Li, X., Hu, X., Wang, J., Wu, T., & Guan, Q. (2021). Decline in the integration of top-down and bottom-up attentional control in older adults with mild cognitive impairment. Neuropsychologia , 161 , 108014. Indovina, I., & Macaluso, E. (2006). Dissociation of stimulus relevance and saliency factors during shifts of visuospatial attention. Cerebral Cortex , 17 (7), 1701–1711. Lech, R. K., Gunturkun, O., & Suchan, B. (2016). An interplay of fusiform gyrus and hippocampus enables prototype- and exemplar-based category learning. Behavioural Brain Research , 311 , 239–246. Lavie, N. (2005). Distracted and confused? Selective attention under load. Trends in Cognitive Sciences , 9 , 75–82. Luck, S. J. (2014). An Introduction to the Event-Related Potential Technique (2nd ed.). MIT Press. Luck, S. J., & Hillyard, S. A. (1994). Spatial filtering during visual search: evidence from human electrophysiology. Journal of Experimental Psychology: Human Perception and Performance , 20 , 1000–1014. Luck, Steven, J., Vogel, & Edward, K. (1997). The capacity of visual working memory for features and conjunctions. Nature , 390 (6657), 279–281. Giammarco, M., Paoletti, A., Guild, E. B., & Al-Aidroos, N. (2016). Attentional capture by items that match episodic long-term memory representations. Visual Cognition , 1–24. Memelink, J., & Hommel, B. (2012). Intentional weighting: a basic principle in cognitive control. Psychological Research , 77 (3). Miller, E. K., & Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. Annual Review of Neuroscience , 24 , 167–202. Mihajlović, N., & Zdravković, S. (2024). Contingent capture by color is sensitive to categorical color perception. Attention Perception & Psychophysics , 86 (1), 36–48. Peelen, M. V., & Kastner, S. (2014). Attention in the real world: toward understanding its neural basis. Trends in Cognitive Sciences , 18 (5), 242–250. Polich, J. (2007). Updating P300: an integrative theory of P3a and P3b. Clinical Neurophysiology , 118 , 2128–2148. Reeder, R. R., & Peelen, M. V. (2013). The contents of the search template for category-level search in natural scenes. Journal of vision , 13 , 1–13. Shang, L., Yeh, L. C., Zhao, Y., Wiegand, I., & Peelen, M. V. (2024). Category-based attention facilitates memory search. Eneuro , 11 (2). Summerfield, C., & Egner, T. (2009). Expectation (and attention) in visual cognition. Trends in Cognitive Sciences , 13 , 403–409. Theeuwes, J. (1992). Perceptual selectivity for color and form. Perception & Psychophysics , 51 , 599–606. Theeuwes, J. (2010). Top-down and bottom-up control of visual selection. Acta Psychologica , 135 , 77–99. Townsend, J. T., & Ashby, F. G. (1983). Stochastic modelling of elementary psychological processes . Cambridge University Press. Ungerleider, L. G., & Haxby, J. V. (1994). What’ and ‘where’ in the human brain. Current Opinion in Neurobiology , 4 , 157–165. Wickens, C. D., & McCarley, J. S. (2008). Applied Attention Theory . CRC. Wolfe, J. M. (2012). Guided search 4.0: current progress with a model of visual search. Integrated Models of Cognitive Systems . Wolfe, J. M., & Horowitz, T. S. (2017). Five factors that guide attention in visual search. In M. W. Current Topics in Behavioral Neurosciences (Vol. 41, pp. 117–141). Springer. Woodman, G. F., Carlisle, N. B., & Reinhart, R., M. G (2013). Where do we store the memory representations that guide attention? Journal of Vision , 13 (3), 1, 1–17. Worden, M. S., Foxe, J. J., Wang, N., & Simpson, G. V. (2000). Anticipatory biasing of visuospatial attention indexed by retinotopically specific alpha-band electroencephalography increases over occipital cortex. Journal of Neuroscience , 20 , RC63. Wu, Q., Chang, C., Xi, S., Huang, I., Liu, Z., Juan, C., et al. (2015). A critical role of temporoparietal junction in the integration of top-down and bottom‐up attentional control. Human Brain Mapping , 36 (11), 4317–4333. Wu, T., Chen, C., Spagna, A. (2020). The functional anatomy of cognitive control: A domain-general brain network for uncertainty processing. Journal of Comparative Neurology , 528 (8). Wu, X., Liu, X., & Fu, S. (2016). Feature- and category-specific attentional control settings are differently affected by attentional engagement in contingent attentional capture. Biological Psychology , 118 , 8–16. Wu, X., Wang, X., Saab, R., & Jiang, Y. (2020). Category-based attentional capture can be influenced by colorand shape-dimensions independently in the conjunction search task. Psychophysiology , 57 (4). Wyble, B., Folk, C., & Potter, M. C. (2013). Contingent attentional capture by conceptually relevant images. Journal of experimental psychology Human perception and performance , 39 (3), 861–871. Yang, H., & Zelinsky, G. J. (2009). Visual search is guided to categorically defined targets. Vision Research , 49 , 2095–2103. Yildirim, K., Akalin-Baskaya, A., & Hidayetoglu, M. L. (2007). Effects of indoor color on mood and cognitive performance. Building & Environment , 42 (9), 3233–3240. Zhang, Y., & Gaspelin, N. (2024). Salience effects on attentional selection are enabled by task relevance. Journal of Experimental Psychology: Human Perception and Performance , 50 (11), 1131–1142. Additional Declarations No competing interests reported. Supplementary Files Openpracticesstatement.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6888298","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":472946677,"identity":"20e59e4d-ba71-4b99-aabe-4127ddefbaa1","order_by":0,"name":"Xia Wu","email":"","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":false,"prefix":"","firstName":"Xia","middleName":"","lastName":"Wu","suffix":""},{"id":472946678,"identity":"cbdd7574-191a-4e52-a77e-49e5ef784e3b","order_by":1,"name":"Yifan Liu","email":"","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":false,"prefix":"","firstName":"Yifan","middleName":"","lastName":"Liu","suffix":""},{"id":472946681,"identity":"d9e2ad91-08c7-49e8-bd47-13f7813aefa8","order_by":2,"name":"Junzhe Wang","email":"","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":false,"prefix":"","firstName":"Junzhe","middleName":"","lastName":"Wang","suffix":""},{"id":472946682,"identity":"d56726d6-bbf3-4394-9020-8a51aa2f08e7","order_by":3,"name":"Xiaoya Sun","email":"","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":false,"prefix":"","firstName":"Xiaoya","middleName":"","lastName":"Sun","suffix":""},{"id":472946683,"identity":"f60dbde9-fa07-4346-a041-ca8a236926c2","order_by":4,"name":"Ying Chen","email":"","orcid":"","institution":"Tianjin University of Technology and Education","correspondingAuthor":false,"prefix":"","firstName":"Ying","middleName":"","lastName":"Chen","suffix":""},{"id":472946684,"identity":"57b06b4e-17bb-4d15-9222-a6589eee6c90","order_by":5,"name":"Yunpeng Jiang","email":"","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":false,"prefix":"","firstName":"Yunpeng","middleName":"","lastName":"Jiang","suffix":""},{"id":472946686,"identity":"2de74548-7bc6-42cf-8b38-0c2bb9180221","order_by":6,"name":"Huan Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA6klEQVRIiWNgGAWjYDACCSBmbAAS7A0MBmCRA0Rr4TlMshaJZKgIIS3ys5ufPfy647C8ueT7A8WFbQxyfDcSGD8X4NHCOOeYubHsmcOGO2cnMxjPbGMwlryRwCw9A48WZokEM2nJtsOMG24DtfC2MSRuuJHAxsyDRwubRPo3kBb7DTcPg7XUE9TCI5FjJvmx7TDQcGawlgQDQlokJHLKpBnb0pM3nEk2MOY5J2E488zDZml8WuRnpG+T/Nlmbbvh+MFnxjxlNvJ8x5MPfsanBQRgzmAzQIom/IDxB1TrA4JKR8EoGAWjYEQCAIBYSTMZu5KdAAAAAElFTkSuQmCC","orcid":"","institution":"Tianjin Normal University","correspondingAuthor":true,"prefix":"","firstName":"Huan","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2025-06-13 12:38:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6888298/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6888298/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":85031652,"identity":"efc59945-1c21-4b36-b325-afafd7138830","added_by":"auto","created_at":"2025-06-20 07:31:41","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":47100,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental design and behavioral outcomes for Experiment 1 (semantic-based categories, task-irrelevant salience).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(Left) \u003c/strong\u003eSchematic of the Majority Function Task: Participants reported the majority category (letters/numbers) in three-symbol arrays (400 ms presentation). Cognitive load was manipulated via symbol ratios (low load: 3:0; high load: 2:1), with task-irrelevant salience operationalized by color oddballs (20% trials). \u003cstrong\u003e(Right)\u003c/strong\u003e Reaction times (RT) and accuracy (ACC) across conditions. Error bars represent standard deviation.\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/06c555407aba1716c67025a7.png"},{"id":85031627,"identity":"43de347a-f58d-42f5-8513-164245ae3bcd","added_by":"auto","created_at":"2025-06-20 07:31:41","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":51096,"visible":true,"origin":"","legend":"\u003cp\u003eStimulus examples for prototype-based categories. The 'O' and 'X' shapes were rendered in 12 distinct fonts to create perceptual variability while maintaining category membership.\u003c/p\u003e","description":"","filename":"image2.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/b93ed2d43efd65446f99b1a8.png"},{"id":85031633,"identity":"dc775d8b-2fa8-453f-a3ce-17b3944dea30","added_by":"auto","created_at":"2025-06-20 07:31:41","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":50890,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental design and behavioral outcomes for Experiment 2 (prototype-based categories, task-irrelevant salience).\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/ce8f904706807c04038fcde4.png"},{"id":85032658,"identity":"c319739d-f288-4f21-ac83-0e8c4f648744","added_by":"auto","created_at":"2025-06-20 07:39:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":46245,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental design and behavioral outcomes for Experiment 3 (semantic-based categories, task-relevant salience).\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/17b2cc6311c6d6c6d48f2aa2.png"},{"id":85032660,"identity":"68a2a4b2-abab-43e4-8887-469eb656e268","added_by":"auto","created_at":"2025-06-20 07:39:41","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":49388,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental design and behavioral outcomes for Experiment 4 (prototype-based categories, task-relevant salience).\u003c/p\u003e","description":"","filename":"image5.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/43d395cafe5d5390e7ea89e8.png"},{"id":85031637,"identity":"eaf967b7-a1d0-44a4-a1c4-f2fedac2b6bd","added_by":"auto","created_at":"2025-06-20 07:31:41","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":36775,"visible":true,"origin":"","legend":"\u003cp\u003eInverse Efficiency Scores (IES) across experiments.\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/0aeec2f4691283be0c5d5197.png"},{"id":91440597,"identity":"f6dc8fa4-87b2-4268-8b96-c77e1f44565e","added_by":"auto","created_at":"2025-09-16 14:02:10","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1420928,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/dbedee30-2bf3-4211-927f-736445fb2d01.pdf"},{"id":85031628,"identity":"1f2f0b7d-7247-467d-baa5-43ec40ac9004","added_by":"auto","created_at":"2025-06-20 07:31:41","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":10488,"visible":true,"origin":"","legend":"","description":"","filename":"Openpracticesstatement.docx","url":"https://assets-eu.researchsquare.com/files/rs-6888298/v1/0f2e91b0145ad4bb80cc25aa.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Dual Pathways of Category-Guided Attentional Selection: Task- Relevance Modulates Hierarchical Integration of Top-Down Control and Salience","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eGiven the complexity of sensory input, human attention system must rapidly select goal-relevant stimuli while suppressing irrelevant distractors. While feature-based attention mechanisms allow rapid orientation toward simple attributes (e.g., color, orientation) (Treisman \u0026amp; Gelade, 1980), real-world environments present stimuli with inherent categorical variability that defies feature-level processing. For instance, searching for \"vehicles\" in traffic requires integrating diverse exemplars (cars, bicycles, scooters) into a unified category rather than tracking discrete features. By grouping stimuli into cognitive categories, category-guided attentional selection (CAS) reduces processing load and enhances recognition efficiency (Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; 2017; \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Yang \u0026amp; Zelinsky, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Wyble et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Reeder \u0026amp; Peelen, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Notably, categories are not monolithic, semantic-based categories (e.g., alphanumeric characters) rely on abstract conceptual rules stored in long-term memory (Goldstone, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e1998\u003c/span\u003e; Giammarco et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., 2018; Wyble et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), whereas prototype-based categories (e.g., \"O\"-like shapes) emerge from perceptual similarity to a central exemplar through sensory integration (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Yildirim et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). For instance, Wu et al. (\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) demonstrated that semantic-based CAS (e.g., colored letters) elicits delayed attentional orienting compared to feature-based attention, whereas prototype-based CAS (e.g., warm-colored \"O\" shapes) engages early perceptual integration (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Despite these insights, fundamental questions remain unresolved: Do semantic and prototype-based CAS recruit distinct cognitive control mechanisms? How do they differentially integrate top-down goals and stimulus salience during executive tasks? Addressing these questions is critical for understanding how categorical flexibility adapts to real-world demands.\u003c/p\u003e \u003cp\u003eCentral to this adaptability is top-down control, a mechanism hypothesized to gate CAS efficiency by resolving competition between category-relevant and irrelevant inputs. While prior research establishes top-down modulation of feature-based attention through working memory (WM) operations (Wu et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; He et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), recent studies extend this to CAS contexts. Mihajlović (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) demonstrated that top-down attentional control enhances search efficiency when aligned with task goals, though their paradigm focused solely on feature-level targets, leaving category-level mechanisms unexplored. Similarly, Shang et al. (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) revealed that category-based top-down attention significantly contributes to memory search, yet their design did not dissociate semantic and prototype-based categories. Crucially, behavioral evidence suggests category-specific top-down dynamics. Semantic CAS (e.g., alphanumeric targets) activates later than feature-based attention (Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), likely due to WM-dependent rule retrieval from long-term memory (Giammarco et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). In contrast, prototype-based CAS (e.g., \"O/X\" shapes) facilitates rapid guidance through perceptual similarity (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), potentially bypassing explicit rule maintenance. This dissociation aligns with neurocognitive models: semantic categories engage prefrontal \"rule hubs\" (Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e), while prototype categories recruit occipitotemporal regions for sensory integration (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). However, a gap exacerbated by the predominant use of visual search paradigms (Schmidt \u0026amp; Zelinsky, 2009; Yang \u0026amp; Zelinsky, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Wyble et al., 2009; Wyble et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2013\u003c/span\u003e), which conflate attentional orienting and executive control, remains. Whether cognitive load differentially impacts these semantic vs. prototype systems still needs to be examined.\u003c/p\u003e \u003cp\u003eBeyond top-down control, stimulus salience is another factor that dynamically interacts with CAS. Salience processing by stimulus prominence (the degree to which an item stands out perceptually) typically captures attention in a bottom-up fashion, but its impact on CAS likely depends on task relevance. Task-irrelevant salience (e.g., a uniquely colored distractor unrelated to the task) can involuntarily capture attention in standard feature-search paradigms. Classic studies using color singletons showed that even irrelevant singletons can disrupt visual search (Theeuwes et al., 1991, 1992). However, in CAS contexts, such interference appears attenuated by category-level filtering: Wu et al. (\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) found that an irrelevant color oddball did not measurably affect search for letters vs. numbers, presumably because the categorical goal allowed participants to filter out low-level color deviations. Conversely, task-relevant salience (e.g., a rare target category occurrence) is expected to amplify attentional effects through goal-directed weighting. Zhang and Gaspelin (\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) reported that physically salient objects only capture attention when they are task-relevant, supporting the idea of goal-contingent salience processing. Translating this to CAS, if an oddball stimulus aligns with the target category (e.g., an infrequent digit among letters), it should engage top-down control mechanisms to a greater extent, possibly affecting later decision-stage processing. In contrast, an oddball that is unrelated to the task goal might be suppressed early in processing (e.g., via distractor inhibition mechanisms). Neurophysiological evidence supports this temporal dichotomy: task-relevant oddballs modulate late event-related potentials associated with decision-making (P3), whereas task-irrelevant oddballs evoke only early sensory components like the N2pc, reflecting automatic suppression (Chen et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Despite these insights, it remains unresolved how salience and category architecture jointly influence behavior. For instance, do semantic vs. prototype categories differ in their vulnerability to salient distractors, and does task relevance alter this? We address these issues by manipulating salience relevance across experiments. Notably, In our design, we operationalized \u0026ldquo;stimulus salience\u0026rdquo; via an oddball probability manipulation rather than a classic singleton (i.e., the oddball was defined by its rarity across trials, not by a unique feature within a single display). This approach emphasizes expectancy/surprise aspects of salience, which differs from traditional within-trial singleton paradigms. We return to this point in the Discussion when interpreting null vs. significant salience effects.\u003c/p\u003e \u003cp\u003eThe interplay between top-down control and salience processing epitomizes a core debate in attention research: are these mechanisms independent (serial model) or interactively integrated (parallel model)? In feature-based attention, Wu et al. (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) identified temporoparietal junction (TPJ) as a nexus for integrating top-down goals and salience signals, with high cognitive load amplifying TPJ-mediated competition. Extending this to CAS, prototype-based categories that relative to perceptual feature binding (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) may prioritize salience integration through occipitotemporal pathways, whereas semantic categories that dependent on prefrontal rule hubs (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) might suppress salience interference via stronger top-down gating. Recent evidence hints at such dissociations, for example, Mihajlović (\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) found color-category contingencies modulated by task goals, while Shang et al. (\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) observed category-based memory enhancements contingent on salience alignment. However, whether top-down control and salience processing operate independently or interactively across category architectures and if such interactions are modulated by task relevance remains unclear.\u003c/p\u003e \u003cp\u003eTo dissociate the effects of top-down control and salience processing across category architectures, we conducted four experiments combining the Majority Function Task (MFT, Fan et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2008\u003c/span\u003e) and oddball paradigm (Theeuwes et al., 1991) orthogonally manipulating two variables in each experiment, including cognitive load (high vs. low) and stimulus salience (oddball vs. standard), and two variable across four experiments, including salience relevance (task-irrelevant in Experiments 1 and 2; task-relevant in Experiments 3 and 4) and category type (semantic-based in Experiment 1 and 3; prototype-based in Experiment 2 and 4). Participants in all experiments performed a majority categorization task, reporting which of two categories was in the majority among three simultaneously presented symbols. Building on prior evidence that semantic CAS relies on WM for rule maintenance (Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wang, 2019) and prototype CAS relies on perceptual fluency (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), we formulated the following hypotheses: (1) High cognitive load will impair performance (slower RTs, lower accuracy) in both category types, but this effect will be especially pronounced for semantic CAS due to its dependence on limited WM resources for maintaining abstract rules. (2) A task-irrelevant oddball (a color singleton unrelated to the categories) will have little to no impact on performance for either category type, as participants\u0026rsquo; category-based focus will filter out this bottom-up distraction. In contrast, a task-relevant oddball (a rare target-category trial) will capture attention and disrupt performance, particularly in the prototype-based task which lacks strong rule-based shielding and thus is more susceptible to bottom-up interference. (3) We anticipated an interaction between cognitive load and salience primarily in the semantic condition: under high load, the cost of an oddball (if task-relevant) would be amplified for semantic CAS, reflecting late-stage decision conflicts when WM is taxed. Prototype CAS, by comparison, might show more perceptual-level competition from oddballs regardless of load, given its reliance on sensory processing. In summary, our experimental design allowed us to test whether top-down load effects dominate CAS performance and whether any salience-driven attentional capture depends on both task relevance and category type. This work extends guided search theory (Wolfe, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) and intentional weighting theory (Memelink \u0026amp; Hommel, 2013) to categorical attention, offering new insights into how conceptual and perceptual hierarchies jointly shape real-world attention control.\u003c/p\u003e"},{"header":"2 Experiment 1","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Methods\u003c/h2\u003e \u003cdiv id=\"Sec4\" class=\"Section3\"\u003e \u003ch2\u003e2.1.1 Participants\u003c/h2\u003e \u003cp\u003eForty right-handed university students (31 female; M\u003csub\u003eage\u003c/sub\u003e = 21.83 years, SD\u0026thinsp;=\u0026thinsp;2.16 years, range\u0026thinsp;=\u0026thinsp;18\u0026ndash;25 years) with normal or corrected-to-normal visual acuity were recruited through campus advertisements. The sample size was determined through a G*Power analysis, aiming for a statistical power of 0.80, with a significance level (α) set at 0.05 and a medium effect size (f\u0026thinsp;=\u0026thinsp;0.25) in 2 \u0026times; 2 within-subject ANOVAs (Cohen, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1992\u003c/span\u003e). All participants provided informed consent and received compensation. The protocol was approved by the institutional Ethics Committee.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.1.2 Stimuli\u003c/h2\u003e \u003cp\u003eVisual stimuli were generated using E-Prime 2.0 on a Windows 7 workstation and displayed on a 17-inch LCD monitor (1024 \u0026times; 768 resolution, 60 Hz refresh rate) with gamma-corrected gray background (CIE x/y coordinates: 0.313/0.329; luminance\u0026thinsp;=\u0026thinsp;50 cd/m\u003csup\u003e2\u003c/sup\u003e). Participants maintained 57 cm head position via chinrest, subtending 0.31\u0026deg;\u0026times; 0.31\u0026deg;visual angle for all stimuli. Twelve equidistant positions (0.86\u0026deg;eccentricity) formed an invisible circle around central fixation. In each trial, three alphanumeric symbols were shown: either three targets (all same category) or two targets plus one distractor (mixed category). Target symbols were letters and distractor symbols were numbers, or vice versa (counterbalanced). \u003cb\u003eCognitive load\u003c/b\u003e was manipulated via this ratio: low-load trials had 3 targets and 0 distractors (3:0), whereas high-load trials had 2 targets and 1 distractor (2:1). Low-load displays are homogeneous and minimize WM demands, whereas high-load displays are heterogeneous and require active suppression of the incongruent item. \u003cb\u003eTask-irrelevant salience\u003c/b\u003e was operationalized by a color oddball: all symbols were either red (CIE x/y: 0.640/0.330) or green (0.300/0.600), with one color occurring on 80% of trials (standard) and the alternate color on 20% of trials (oddball). Importantly, color was entirely unrelated to the category task (e.g., an odd-colored item could be either a target or distractor). The assignment of colors to \u0026ldquo;standard\u0026rdquo; vs. \u0026ldquo;oddball\u0026rdquo; was counterbalanced across blocks and participants. This probabilistic salience manipulation ensured the oddball item was unexpected but not more physically salient (in terms of luminance or contrast) than standard items. All other stimulus parameters (size, timing) were identical across conditions.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e\u003cb\u003e2.1.3 Experiment Design and Procedure\u003c/b\u003e\u003c/h2\u003e \u003cp\u003eExperiments employed a 2 (cognitive load: high vs. low) \u0026times; 2 (stimulus salience: standard vs. oddball) within-subjects design. Each trial (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e) began with a central fixation cross for 100\u0026ndash;600 ms (randomized), followed by a 400 ms presentation of the three-symbol array. Participants pressed one of two keys to indicate which category (letters or numbers) was in the majority. They had up to 1600 ms to respond, after which a blank inter-trial interval of 1500\u0026ndash;2000 ms occurred. The experiment comprised 16 practice trials with feedback, then 400 experimental trials divided into 4 blocks of 100 trials (each block containing 80 standard-color trials and 20 oddball-color trials). Trials were self-paced between blocks. Accuracy (ACC) and reaction time (RT) were recorded.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Results\u003c/h2\u003e \u003cp\u003eWe conducted a 2 (cognitive load: low vs. high) \u0026times; 2 (stimulus salience: standard vs. oddball) repeated-measures ANOVA on accuracy and RT (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAccuracy\u003c/strong\u003e \u003cp\u003eWe found a significant main effect of cognitive load, \u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;74.59, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.66, with accuracy lower under high load (90.5\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0%) than low load (96.7\u0026thinsp;\u0026plusmn;\u0026thinsp;0.8%). No main effect of salience emerged, \u003cem\u003eF\u003c/em\u003e(1,39)\u0026thinsp;=\u0026thinsp;1.14, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.29, nor load \u0026times; salience interaction, \u003cem\u003eF\u003c/em\u003e(1,39)\u0026thinsp;=\u0026thinsp;0.008, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;.93.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRT\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load, \u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;534.36, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.93 showed that high load slowed responses (883.16\u0026thinsp;\u0026plusmn;\u0026thinsp;24 ms high load vs. 703.48\u0026thinsp;\u0026plusmn;\u0026thinsp;20.36 ms low load). There was no significant RT difference between standard vs. oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.20, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.66) nor interaction (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.13, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.72).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInverse efficiency score (IES)\u003c/strong\u003e \u003cp\u003ea combined speed-accuracy measure (RT/ACC, Townsend \u0026amp; Ashby, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e1983\u003c/span\u003e), and higher IES values indicate poorer performance. The analysis of IES also found a significant main effect of cognitive load, \u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;571.65, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.93, IES was higher (worse) under high load (983.43\u0026thinsp;\u0026plusmn;\u0026thinsp;32.21) than low load (732.79\u0026thinsp;\u0026plusmn;\u0026thinsp;25.77). There was no main effect of salience (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.21, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.65), nor interaction (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.00, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.99). Thus, in Experiment 1 the odd-colored distractor had no reliable effect on performance, whereas high cognitive load caused a large decrement in accuracy and speed.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e\u003cb\u003e2.3 Discussion of Experiment 1\u003c/b\u003e\u003c/h2\u003e \u003cp\u003eExperiment 1 demonstrates that semantic-based CAS is dominated by top-down control demands, with little influence from task-irrelevant salience. The cognitive load manipulation (comparing 3:0 vs. 2:1 displays) robustly impaired performance: high-load trials led to substantially slower RTs (+\u0026thinsp;179 ms on average) and lower accuracy (-6.2%) than low-load trials (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001). These findings are consistent with models positing that semantic CAS relies on WM-mediated maintenance of category templates. In our task, increasing load (mixing categories) likely depleted the resources needed to uphold the task rule (\u0026ldquo;which category is majority\u0026rdquo;), in line with evidence that higher working memory demands slow rule-based decisions. Notably, the absence of any salience effect indicates that a task-irrelevant color oddball did not capture attention in this semantic search context. This null result contrasts with classic feature-search studies where an irrelevant color singleton can automatically capture gaze or attention (e.g., Theeuwes, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1992\u003c/span\u003e). In our study, participants\u0026rsquo; categorical focus may have filtered out the odd-color item, or the oddball\u0026rsquo;s \u0026ldquo;surprise\u0026rdquo; value was simply insufficient to overcome the top-down set. In other words, there was no evidence of attentional capture by the irrelevant oddball, indicating the occurrence of category-level gating, although we caution that the null effect could also stem from the subtlety of our salience manipulation (see General Discussion).\u003c/p\u003e \u003cp\u003eThe dissociation between strong load effects and null salience effects in this experiment supports a competitive resource allocation view. That is, when attention is engaged in the relatively complex process of semantic categorization, little capacity remains for processing an irrelevant feature singletons (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Consistent with this, neuroimaging work (Wu et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) shows that under high load, brain regions like the temporoparietal junction (TPJ) are occupied by task demands and then mediate attentional capture, reducing responsiveness to distractions.\u003c/p\u003e \u003cp\u003eConsistent with this, neuroimaging study indicate that under high load, brain regions like the temporoparietal junction (TPJ), which are responsible for attentional engagement, are occupied by task demands, reducing responsiveness to distractions. Our results extend this idea by showing that even a potentially attention-grabbing color oddball (20% probability) failed to produce any interference when participants focused on the categorical rule under load. In summary, semantic CAS in our task operated as if protected by a top-down \u0026ldquo;filter\u0026rdquo; that prevented irrelevant singletons from affecting choice, at least when working memory was taxed.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Experiment 2","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Methods\u003c/h2\u003e \u003cp\u003eThe same 40 participants from Experiment 1 performed Experiment 2 after a\u0026thinsp;~\u0026thinsp;1-hour break (to reduce carryover effects). The stimuli and design of Experiment 2 was identical to Experiment 1 except that the category task switched to prototype-based categories. Participants now judged the majority between \u0026ldquo;O\u0026rdquo; vs. \u0026ldquo;X\u0026rdquo; shapes, presented in 12 distinct font styles to introduce within-category visual variability (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The spatial layout (three items among 12 positions), timing, and color-probability salience manipulation (80% standard-color, 20% oddball-color trials) were the same as Experiment 1. Low-load trials had three shapes of one kind, and high-load trials had two of one kind and one of the other (2:1). This design isolates the effect of category type by holding the load and salience parameters constant while substituting semantic stimuli with perceptual shapes. Participants were instructed exactly as before (identify the majority category per trial).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Results\u003c/h2\u003e \u003cp\u003eA 2 (cognitive load: high vs. low) \u0026times; 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAccuracy\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load emerged, \u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;241.10, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.86, with accuracy dropping from 97.03\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4% (low load) to 86.50\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0% (high load). There was no main effect of salience (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.25, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.62) nor interaction (\u003cem\u003eF\u003c/em\u003e (1, 39)\u0026thinsp;=\u0026thinsp;18, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.68).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRT\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;260.20, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.87) showed that high load significantly slowed RTs (796.79\u0026thinsp;\u0026plusmn;\u0026thinsp;21.51 ms high load vs. 670.42\u0026thinsp;\u0026plusmn;\u0026thinsp;17.47 ms low load). No effect of salience was observed (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.001, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.89), nor the interaction (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.14, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.71).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eIES\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;457.91, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.92) showed that high load increased IES (691\u0026thinsp;\u0026plusmn;\u0026thinsp;18 low vs. 922\u0026thinsp;\u0026plusmn;\u0026thinsp;23 high). No effect of stimulus salience (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.18, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.68) nor the interaction (\u003cem\u003eF\u003c/em\u003e(1, 39)\u0026thinsp;=\u0026thinsp;0.13, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.72) was found. Thus, as in Experiment 1, cognitive load had a large impact on performance, whereas the irrelevant color oddball again produced no reliable slowdown or accuracy change.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Discussion of Experiment 2\u003c/h2\u003e \u003cp\u003eExperiment 2 examined CAS under the same load and salience conditions, but for prototype-based categories. The results closely mirrored those of Experiment 1. High cognitive load led to substantially worse performance: under low load, participants responded faster (+\u0026thinsp;126 ms advantage) and more accurately (+\u0026thinsp;10.5% ACC) than under high load (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001). This confirms that top-down control is a critical limiting factor for prototype-based CAS as well, extending the role of WM observed in semantic tasks to perceptual categorization (Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wang, 2019). Importantly, we again found no effect of the task-irrelevant oddball, nor any load \u0026times; salience interaction (\u003cem\u003eps\u003c/em\u003e\u0026thinsp;\u0026gt;\u0026thinsp;.5). Even in the prototype search, where all stimuli were simple shapes, an odd-colored item did not measurably distract participants. This suggests that, under our conditions, task-irrelevant features failed to override category-level prioritization. In other words, participants allocated attentional resources preferentially to determining the predominant shape category, at the expense of processing an off-colored item that carried no task information. The absence of any salience cost in both Experiments 1 and 2 indicates a domain-general filtering mechanism: whether searching for letters or shapes, observers can ignore an irrelevant oddball when focused on a category judgment (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). In summary, Experiments 1\u0026ndash;2, which both involved task-irrelevant salience, yielded strong and consistent cognitive load effects but null salience effects for both semantic and prototype CAS. This pattern raised the question of whether salience might influence CAS if it were made task-relevant, that is, if the oddball signaled a change in the target category rather than being a mere color anomaly. We next turned to that question in Experiments 3\u0026ndash;4.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Experiment 3","content":"\u003cp\u003eThe lack of salience effects in Experiments 1\u0026ndash;2 (where oddball color was irrelevant to the task) suggested that stimulus salience modulates CAS only when aligned with task goals. Experiment 3 was designed to test this hypothesis by making the oddball task-relevant. Instead of color, we manipulated the probability of target category as the salience cue: one category occurred on 80% of trials and the alternative category on only 20% (the \u0026ldquo;oddball\u0026rdquo; category). Thus, the oddball in Experiment 3 was a rare target category trial, directly relevant to task performance. We used the semantic category set (letters vs. numbers) as in Experiment 1, under equivalent high vs. low load conditions.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Methods\u003c/h2\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e4.1.1 Participants\u003c/h2\u003e \u003cp\u003eAnother forty right-handed university students (30 female; M\u003csub\u003eage\u003c/sub\u003e = 20.73 years, SD\u0026thinsp;=\u0026thinsp;2.15 years, range\u0026thinsp;=\u0026thinsp;18\u0026ndash;26 years) with normal or corrected-to-normal visual acuity were recruited through campus advertisements. Sample size and inclusion criteria matched those of Experiments 1\u0026ndash;2. All provided consent and were compensated. Procedures were approved by the Ethics Committee.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e4.1.2 Stimuli and Procedure\u003c/h2\u003e \u003cp\u003eExperiment 3 retained the 2 \u0026times; 2 design of Experiment 1 (cognitive load \u0026times; salience), but reconceptualized \u0026ldquo;salience\u0026rdquo; as a task-relevant category probability. The stimuli were letters and numbers as before. We assigned one category to be frequent (80% of trials, \u0026ldquo;standard\u0026rdquo;) and the other category to be rare (20% of trials, \u0026ldquo;oddball\u0026rdquo;) in the majority judgment task. For example, in one block, 80% of trials had a majority of letters (with occasional number-majority trials as oddballs); in another block, the probabilities were reversed (80% numbers as majority). Participants were informed about these base rates so that the oddball category trials would be unexpected but clearly tied to the task (they still had to identify the majority category on every trial). Importantly, color was now kept constant, yielding all stimuli were presented in a uniform color (e.g., all white) so that any salience effect would come purely from the rarity of the target category, not from a physical singleton. Load was also manipulated as before (3:0 vs. 2:1 ratio of letters:numbers). Timing, trial structure, and total number of trials were identical to Experiment 1. This design allowed a direct comparison with Experiment 1 to see how making the oddball task-relevant changes its impact.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Results\u003c/h2\u003e \u003cp\u003eIn this semantic task with task-relevant oddballs, a 2 (cognitive load: high vs. low) \u0026times; 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAccuracy\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load was observed, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;180.74, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.83, with lower accuracy under high load (81.5\u0026thinsp;\u0026plusmn;\u0026thinsp;1.7%) than low load (90.9\u0026thinsp;\u0026plusmn;\u0026thinsp;1.4%). Importantly, there was now a significant main effect of salience, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;76.44, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.67, with the accuracy on standard trials (94.3\u0026thinsp;\u0026plusmn;\u0026thinsp;0.8%) was higher than on oddball trials (79.0\u0026thinsp;\u0026plusmn;\u0026thinsp;2.3%). Furthermore, a significant load \u0026times; salience interaction emerged, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;84.48, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001, η\u0026sup2; = 0.69. Simple effects analysis indicated that the oddball category impaired accuracy under both low-load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;27.94, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.42) and high-load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;121.09, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.76). Similarly, the load effect (higher accuracy under low-load relative to high-load) was observed for both standard (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;92.52, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.71) and oddball stimuli (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;149.02, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.80).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eRT\u003c/b\u003e: A significant main effect of cognitive load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;156.72, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.81) showed slower responses on high-load trials (631.22\u0026thinsp;\u0026plusmn;\u0026thinsp;19.79 ms) than low-load (534.74\u0026thinsp;\u0026plusmn;\u0026thinsp;15.40 ms). They were also slower on oddball trials (622\u0026thinsp;\u0026plusmn;\u0026thinsp;17 ms) than standard trials (544\u0026thinsp;\u0026plusmn;\u0026thinsp;18 ms), \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;135.09, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.78. In contrast to accuracy, the RT load \u0026times; salience interaction was not significant (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;1.90, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.18), indicating that the oddball slowed responses by a similar amount in both load conditions.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eIES\u003c/strong\u003e \u003cp\u003eBoth main effects were significant (main effect of cognitive load, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;212.47, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.85, better performance for low load (594.33\u0026thinsp;\u0026plusmn;\u0026thinsp;16.84) than high load (808.64\u0026thinsp;\u0026plusmn;\u0026thinsp;21.40); main effect of salience, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;66.18, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.64, better performance for standard trials (577.93\u0026thinsp;\u0026plusmn;\u0026thinsp;18.16) than oddball trials (826.04\u0026thinsp;\u0026plusmn;\u0026thinsp;27.71)). Importantly, the load \u0026times; salience interaction was significant for IES, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;51.30, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.57. Simple effects analysis mirrored the accuracy interaction, with better performance under low-load relative to high-load in both standard (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;343.53, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.90) and oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;131.35, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.78), and better performance for standard trials than oddball trials under both low-load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;33.98, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.47) and high load conditions (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;78.93, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.68). Thus, combining speed and accuracy, the cost of oddball trials under high load was disproportionately large. In summary, Experiment 3 revealed that when the oddball was task-relevant (indicating a rare target category), it had a strong detrimental effect on performance, which is a stark contrast to the null results in Experiments 1\u0026ndash;2.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Discussion of Experiment 3\u003c/h2\u003e \u003cp\u003eExperiment 3 demonstrates that making the oddball task-relevant fundamentally changes its impact on CAS. In the semantic categorization task, introducing a rare-category oddball led to pronounced interference, especially under high load. Replicating prior experiments, we again saw robust cognitive load effects (participants were ~\u0026thinsp;182 ms faster and 12.3% more accurate in low-load vs. high-load trials; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001). However, unlike Experiments 1\u0026ndash;2, stimulus salience here significantly modulated performance. When an oddball trial occurred (e.g., a number-majority trial in a block where letters were usually majority), participants were slower and less accurate than on standard frequent-category trials. This oddball cost was substantial (overall\u0026thinsp;~\u0026thinsp;+\u0026thinsp;64 ms RT, -14.8% accuracy). Moreover, the cost was amplified under high load (as reflected by the significant interaction in accuracy and IES), indicating that resolving a rare-category occurrence was particularly difficult when working memory was already strained. This pattern contrasts sharply with the null salience effects in Experiments 1\u0026ndash;2, highlighting that attentional capture by salience in CAS is highly task-contingent. In short, when the \u0026ldquo;oddball\u0026rdquo; is relevant to the search goal, it does capture attention and disrupt performance, but notably, mainly in situations of high cognitive demand.\u003c/p\u003e \u003cp\u003eThese findings refine our understanding of CAS by showing that goal relevance reconfigures the interplay between perceptual salience and conceptual processing. Under task-relevant conditions, the attentional system cannot simply ignore the oddball; instead, the oddball competes for processing resources, leading to slower and less accurate decisions. Conceptually, this suggests a dynamic allocation: when an oddball aligns with task goals (even as a rare event), the cognitive system prioritizes it, perhaps overly so, at the expense of routine processing, especially if executive resources are limited (high load). We will examine in the General Discussion whether this effect corresponds to a \u0026ldquo;late-stage\u0026rdquo; conflict (e.g., deciding that the oddball category is actually the majority) as opposed to an early attentional capture. For now, Experiment 3 confirms that salience can impact CAS when it matters to the task, whereas it was completely filtered out when it did not.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Experiment 4","content":"\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e5.1 Methods\u003c/h2\u003e \u003cp\u003eThe same participants from Experiment 3 completed Experiment 4 after a 1-hour washout period to eliminate carryover effects. Experiment 4 extended the task-relevant salience design to the prototype-based categories (\u0026ldquo;O\u0026rdquo; vs. \u0026ldquo;X\u0026rdquo;). The majority-category probability manipulation (80% vs. 20%) was applied to \u0026ldquo;O\u0026rdquo; vs. \u0026ldquo;X\u0026rdquo; trials analogously to how letters vs. numbers were treated in Experiment 3. In other words, one shape category was made frequent and the other rare (counterbalanced). All aspects of the apparatus, timing, and load manipulation (3:0 vs. 2:1 shapes) mirrored those of Experiment 3.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003e5.2 Results\u003c/h2\u003e \u003cp\u003eA 2 (cognitive load: high vs. low) \u0026times; 2 (stimulus salience: standard vs. oddball) ANOVA was conducted on accuracy and RT (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eAccuracy\u003c/strong\u003e \u003cp\u003eA significant main effect of cognitive load was observed, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;172.81, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.82 (low load 92.7\u0026thinsp;\u0026plusmn;\u0026thinsp;1.0% vs. high load 79.3\u0026thinsp;\u0026plusmn;\u0026thinsp;1.4%), as well as salience, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;52.86, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.58 (standard trials 91.8\u0026thinsp;\u0026plusmn;\u0026thinsp;0.7% vs. oddball trials 80.2\u0026thinsp;\u0026plusmn;\u0026thinsp;1.8%). There was also a significant interaction, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;15.26, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.29. The post-hoc tests showed oddball accuracy was lower than standard accuracy under both low load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;42.05, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.53) and high-load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;44.68, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.54), and load effect for both standard (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;129.97, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.77) and oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;99.02, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.72).\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eRT\u003c/strong\u003e \u003cp\u003eHigh load slowed RTs (529\u0026thinsp;\u0026plusmn;\u0026thinsp;15 ms low vs. 606\u0026thinsp;\u0026plusmn;\u0026thinsp;20 ms high), \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;154.49, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.80. Oddball trials were significantly slower (603.43\u0026thinsp;\u0026plusmn;\u0026thinsp;17.44 ms) than standard trials (531.86\u0026thinsp;\u0026plusmn;\u0026thinsp;18.21 ms), \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;158.53, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.80. The interaction for RT was not significant, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;0.49, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.49.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eIES\u003c/b\u003e: Main effects of load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;295.86, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.89) and salience (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;85.37, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.69) were significant (worse performance for high load (785.67\u0026thinsp;\u0026plusmn;\u0026thinsp;21.55 vs. 574.75\u0026thinsp;\u0026plusmn;\u0026thinsp;14.03 low load) and for oddballs (774.85\u0026thinsp;\u0026plusmn;\u0026thinsp;18.69 vs. 585.57\u0026thinsp;\u0026plusmn;\u0026thinsp;21.14 standard). Importantly, there was also a significant load \u0026times; salience interaction, \u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;24.25, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.39. Simple effects analysis mirrored the accuracy analysis: oddball IES was higher than standard in both low-load (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;108.65, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.74) and high-load conditions (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;62.39, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.62), and high load IES was higher than low load in both standard (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;181.80, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.83) and oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 38)\u0026thinsp;=\u0026thinsp;147.68, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.80). In summary, Experiment 4 showed that for prototype CAS with task-relevant oddballs, we again see clear salience-driven interference (slower, less accurate responses on oddball trials) along with strong load effects. Unlike Exp 3, the salience effect in Exp 4 did not depend as strongly on load because of that it was present under both low and high load to a similar degree.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003e5.3 Discussion of Experiment 4\u003c/h2\u003e \u003cp\u003eExperiment 4 revealed critical differences in how prototype-based CAS handles task-relevant salience, compared to semantic CAS in Experiment 3. Cognitive load effects persisted (participants were ~\u0026thinsp;121 ms faster and 11.2% more accurate in low vs. high load, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;.001), indicating that even prototype search suffers under increased task demands. Notably, the salience interference was clearly present for prototype-based CAS: oddball trials induced substantial RT slowing (+\u0026thinsp;89 ms on average) and accuracy declines (-18.3%) compared to standard trials. These costs were numerically larger than those observed in the semantic task of Experiment 3 (+\u0026thinsp;64 ms RT, -14.7% ACC). This pattern suggests that prototype-based CAS is even more sensitive to salient deviations, perhaps because it relies on occipitotemporal perceptual binding processes that are directly impacted by oddball stimuli. In other words, when a rare target shape appeared, it may have strongly captured perceptual attention in the visual processing stream, competing with the template-matching process for the frequent target shape. Consistent with this idea, the interference by oddballs in prototype search can be seen as a form of \u0026ldquo;perceptual amplification\u0026rdquo; of salience effects. The results align with predictive coding accounts whereby the visual system of prototype-based CAS responds to an unexpected stimulus with greater disruption (here, an \u0026ldquo;oddball\u0026rdquo; shape violates the expected template, requiring additional processing). By contrast, semantic CAS might handle a rare oddball through higher-level rule evaluation (which we saw was especially taxing under high load, as evidenced by the load \u0026times; salience interaction in Exp 3).\u003c/p\u003e \u003cp\u003eImportantly, although prototype CAS showed greater absolute costs from oddballs, its overall performance remained faster and quite accurate even on oddball trials (e.g., ~\u0026thinsp;603 ms RT and 80% ACC on oddball trials, versus semantic CAS\u0026rsquo;s\u0026thinsp;~\u0026thinsp;622 ms and 79% in Exp 3). This reflects a theme that will be expanded upon: prototype processing trades some vulnerability to distraction for generally higher perceptual efficiency. In sum, Experiments 3\u0026ndash;4 confirm that when salience is tied to the task (the oddball signals a rare target event), both semantic and prototype CAS are significantly disrupted by oddballs. The effect is present in both category types, but the magnitude and perhaps the nature of the interference differ: semantic CAS shows a larger relative cost under high cognitive load (pointing to load exacerbating decision-level conflict), whereas prototype CAS shows a robust salience effect regardless of load (consistent with more automatic, perceptual-level capture). We examine these differences directly in the combined analyses below.\u003c/p\u003e \u003c/div\u003e"},{"header":"6 Comprehensive analysis","content":"\u003cp\u003eTo directly compare effects across experiments, we performed a comprehensive ANOVA on the inverse efficiency scores (IES) with Experiment factors combined (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e). This mixed-design ANOVA included cognitive load (low vs. high), stimulus salience (standard vs. oddball), salience relevance (task-irrelevant: Exp 1\u0026ndash;2 vs. task-relevant: Exp 3\u0026ndash;4), and category type (semantic vs. prototype). Cognitive load, stimulus salience, and category were treated as within-subject (note: each pair of experiments shared participants, but salience relevance varied between the group that did Exp 1\u0026ndash;2 and the group that did Exp 3\u0026ndash;4, so we conservatively treat it as a between-subject factor). This analysis allowed us to assess: (1) differences between semantic and prototype tasks, (2) differences between task-irrelevant vs. task-relevant salience conditions, and (3) interactions among these factors (e.g., three-way interactions indicating differential integration of load and salience).\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eMain effects\u003c/strong\u003e \u003cp\u003eWe found a significant main effect of \u003cb\u003ecognitive load\u003c/b\u003e, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;1070.12, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.93, with IES under low load (648.27\u0026thinsp;\u0026plusmn;\u0026thinsp;12.19) lower (better performance) than that under high load (875.05\u0026thinsp;\u0026plusmn;\u0026thinsp;15.57). A significant main effect of \u003cb\u003estimulus salience\u003c/b\u003e was also observed, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;83.51, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.52, with IES on standard trials (706.78\u0026thinsp;\u0026plusmn;\u0026thinsp;14.48) lower (better performance) than that on oddball trials (816.33\u0026thinsp;\u0026plusmn;\u0026thinsp;15.13). \u003cb\u003eCategory type\u003c/b\u003e showed a significant main effect as well, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;6.77, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.011, η\u0026sup2; = 0.81, indicating that overall performance was better for prototype-based CAS (743.52\u0026thinsp;\u0026plusmn;\u0026thinsp;13.26) than for semantic-based CAS (779.80\u0026thinsp;\u0026plusmn;\u0026thinsp;16.98). This advantage for prototype categories echoes the raw results where prototype tasks often had faster responses for equivalent accuracy. Lastly, \u003cb\u003esalience relevance\u003c/b\u003e (task-irrelevant vs. task-relevant conditions) had a significant main effect, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;27.14, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.26, with overall IES was lower in the task-relevant salience experiments (691.10\u0026thinsp;\u0026plusmn;\u0026thinsp;19.28) than in the task-irrelevant experiments (832.22\u0026thinsp;\u0026plusmn;\u0026thinsp;19.08). On face value, this might seem surprising (since one might expect performance to be worse when dealing with oddballs that matter), but it reflects that Exps 3\u0026ndash;4 participants responded faster overall than Exps 1\u0026ndash;2 participants. This difference likely arises because different participant groups were used and perhaps improved with practice by the time they did Exps 3\u0026ndash;4; we interpret interaction effects below with caution for this between-group factor.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInteractions\u003c/strong\u003e \u003cp\u003eThere was a strong \u003cb\u003ecognitive load \u0026times; stimulus salience\u003c/b\u003e interaction, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;47.74, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.38. As already implied by individual experiments, load and salience each impaired performance, and their combination was especially detrimental. The oddball vs. standard difference in IES was significant under both low-load (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;60.65, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.44) and high-load conditions (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;81.88, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.52). Likewise, the load effect (high vs. low) was significant for both standard (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;1007.95, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.93) and oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;563.91, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.88). However, the magnitude of the load effect was even larger for standard trials than oddball trials (since performance was already quite degraded on oddball-high-load trials, leaving slightly less additional room for load to worsen it). This pattern aligns with the notion that having both high load and an oddball creates a ceiling of difficulty, where performance is pushed to a floor.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eWe also found a significant \u003cb\u003ecategory type \u0026times; stimulus salience\u003c/b\u003e interaction, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;5.24, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.025, η\u0026sup2; = 0.06. Interestingly, although both semantic (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;63.23, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.45) and prototype tasks (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;80.07, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.51) showed worse performance on oddball trials than standard trials, prototype-based CAS was overall more efficient than semantic CAS specifically on oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;9.43, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003, η\u0026sup2; = 0.11). This indicates that prototypes had a performance edge even when dealing with oddballs (likely due to generally faster processing), despite the fact that in absolute terms they suffered a larger slowing from oddballs. In essence, prototype CAS was relatively resilient in final outcome, even though oddballs did disrupt it.\u003c/p\u003e \u003cp\u003eThere was also a \u003cb\u003ecategory type \u0026times; salience relevance \u0026times; salience\u003c/b\u003e (3-way) interaction, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;7.90, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.006, η\u0026sup2; = 0.09. This reflected subtle differences in how category type mattered across the irrelevant vs. relevant experiments. In the standard-stimulus condition, making salience task-relevant reliably improved efficiency for both category systems: in the semantic task, task-relevant IES was significantly lower than its task-irrelevant counterpart, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;67.26, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.47; the same pattern held for the prototype task, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;54.17, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.41. Nevertheless, the prototype pathway still maintained a speed\u0026ndash;accuracy edge in most situations: IES for prototype-based categories was significantly lower than for semantic categories in task-relevant/standard trials (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;7.49, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.008, η\u0026sup2; = 0.09), task-relevant/oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;4.45, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.038, η\u0026sup2; = 0.06), and even task-irrelevant/oddball trials (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;4.98, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.028, η\u0026sup2; = 0.06). Within each category system, oddballs imposed a clear cost whenever they were task-relevant, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;127.95, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.62. Taken together, task relevance globally boosts processing efficiency in both CAS pathways, yet the prototype route retains a better speed-accuracy trade-off across all salience contexts, whereas the semantic route suffers the greatest performance drop when a task-relevant oddball (especially under high load) is added, underscoring fundamental differences in how the two pathways allocate resources and integrate salience.\u003c/p\u003e \u003cp\u003eLastly, the 3-way interaction of \u003cb\u003eload \u0026times; salience relevance \u0026times; salience\u003c/b\u003e was significant, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;44.70, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.37. Under every combination of salience relevance and stimulus salience, low load produced markedly lower IES (better performance) than high load: task-irrelevant/standard, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;829.39, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.92, task-irrelevant/oddball, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;235.92, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.75, task-relevant/standard, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;261.81, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.77, and task-relevant/ oddball, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;331.47, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.81. Turning to salience relevance, the overall performance in task-relevant experiments (Exp. 3\u0026ndash;4) was better than task-irrelevant experiments (Exp. 1\u0026ndash;2) in low load/standard, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;60.52, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.44, and high load/standard, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;83.52, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.52, and low load/oddball trials, \u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;4.50, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.037, η\u0026sup2; = 0.06. Finally, the disadvantages of oddball trials were found in task-relevant experiments, no matter in high load (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;121.93, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.61) or low-load conditions (\u003cem\u003eF\u003c/em\u003e(1, 77)\u0026thinsp;=\u0026thinsp;160.31, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001, η\u0026sup2; = 0.68).\u003c/p\u003e \u003cp\u003eIn summary, the comprehensive analysis statistically supports conclusions drawn from individual experiments: (a) cognitive load had a strong deleterious effect regardless of category type or salience relevance; (b) stimulus oddballs significantly affected performance only in the task-relevant context; (c) prototype-based CAS was overall more efficient than semantic CAS (lower IES), including during oddball trials; (d) the worst performance occurred when both high load and a task-relevant oddball were present (especially for semantic CAS).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"7 General Discussion","content":"\u003cp\u003eThis study systematically dissected how top-down cognitive control and stimulus salience jointly shape category-guided attentional selection (CAS) across semantic and prototype-based systems through four behavioral experiments. By orthogonally manipulating cognitive load (high vs. low), stimulus salience (oddball vs. standard), salience relevance (task-irrelevant vs. task-relevant), and category architecture (semantic: alphanumeric characters; prototype: \"O/X\" shapes), we identified three principal mechanisms. First, cognitive load exerted a domain-general effect, with high-load conditions impairing accuracy (6.2\u0026ndash;10.5%) and prolonging reaction times (126\u0026ndash;179 ms) across both semantic and prototype tasks. This is consistent with the notion that maintaining multiple items or rules in working memory under high load consumes attentional resources needed for CAS (Wu et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Wang, 2019). Second, stimulus salience affected performance exclusively when it was task-relevant: in our task-irrelevant conditions (Exp 1\u0026ndash;2), oddball stimuli produced no significant interference, whereas in task-relevant conditions (Exp 3\u0026ndash;4), rare-category oddballs incurred clear performance costs (RT delays of ~\u0026thinsp;64\u0026ndash;89 ms and accuracy drops of ~\u0026thinsp;15\u0026ndash;18%). This finding demonstrates goal-contingent prioritization and aligns with recent evidence that salience matters only when it matches current goals, while challenging classical views of automatic capture by salient distractors regardless of goals. Third, prototype-based CAS showed overall superior efficiency on oddball trials compared to semantic CAS (approximately a 12.3% IES advantage). We attribute this to perceptual similarity-driven guidance in the prototype task, which may allow faster grouping and processing of stimuli, effectively bypassing the need for slow rule retrieval that semantic tasks require (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Importantly, the interactions between load and salience provide additional insight: under high load, semantic CAS faced amplified conflict resolution demands (evidenced by disproportionately slowed decisions and reduced accuracy when a task-relevant oddball occurred), suggesting that the bottleneck was at a late decision stage (e.g., making the category judgment with conflicting evidence). In contrast, prototype CAS under high load exhibited signs of early-stage perceptual competition when oddballs appeared, that is, the interference was manifest in RT costs that did not strongly depend on load, implying that even with ample resources, a perceptually salient oddball engages early visual attention in the prototype task. Together, these findings support a dual-pathway account: semantic CAS primarily operates via conceptual, rule-based control susceptible to WM load, whereas prototype CAS operates via perceptual grouping and is thus faster and somewhat more \u0026ldquo;bottom-up\u0026rdquo; albeit still subject to strategic control.\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e7.1 Dominance of Top-Down Control in CAS\u003c/h2\u003e \u003cp\u003eAcross all experiments, cognitive load was the dominant factor limiting CAS performance. High-load trials (requiring integration of a distractor category) consistently produced worse outcomes in both accuracy and RT. This underscores that WM capacity is a domain-general bottleneck for CAS. Our findings align with models suggesting attentional selection efficiency hinges on maintaining attentional templates in WM (Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). When load increased (2:1 vs. 3:0), participants had to compare and suppress multiple category representations, leaving fewer resources available for timely target selection. where increased cognitive load depletes resources critical for sustaining category-specific representations. Notably, this bottleneck manifests differently across category architectures: semantic-based CAS, reliant on rule retrieval from long-term memory (Giammarco et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), exhibited delayed response thresholds under high load, reflecting protracted conflict resolution in prefrontal-parietal networks (Wang, 2019), whereas prototype-based CAS, driven by perceptual similarity gradients (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e), showed attenuated accuracy declines, suggesting perceptual fluency compensates through feature binding (Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The three-way interaction (category \u0026times; load \u0026times; salience) further confirms this dissociation, with semantic systems prioritize abstract rule arbitration under resource competition, while prototype systems exploit perceptual fluency to mitigate WM demands. A dual-pathway CAS model interprets this divergence: semantic processing engages prefrontal-parietal networks for hierarchical control, while prototype processing utilizes occipitotemporal circuits for sensory optimization (Wu et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Overall, these findings extend Wu et al.\u0026rsquo;s (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) framework of temporoparietal junction-mediated control integration to categorical contexts, demonstrating that cognitive load amplifies competition between category templates and salient distractors across architectures. This hierarchical constraint aligns with capacity-limited models of cognitive control (Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Wu et al., 2018, 2019), where CAS efficiency reflects dynamic trade-offs between rule precision and perceptual integration, rather than mere feature-level competition.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e7.2 Task-Contingent Salience Effects on CAS\u003c/h2\u003e \u003cp\u003eIn contrast to the ever-present load effects, stimulus salience influenced CAS only under specific conditions, namely, when the salient feature was tied to the task. This finding supports the idea of goal-contingent attentional capture (Zhang \u0026amp; Gaspelin, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). In Experiments 1\u0026ndash;2 (task-irrelevant oddballs), we saw no hint of the typical distraction caused by an odd-colored item; participants effectively filtered out the oddball color, focusing on category identity. This aligns with studies showing that when observers are engaged in a demanding task, irrelevant singletons often fail to capture attention (especially if they fall outside the \u0026ldquo;attentional set\u0026rdquo;). Our results go further by demonstrating such filtering in a categorical context: a color singleton that had no bearing on the category decision was behaviorally suppressed. Neuroimaging evidence concurs that task-irrelevant salient stimuli mainly activate visual and parietal regions (Clark et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2000\u003c/span\u003e; Downar et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2001\u003c/span\u003e; Indovina \u0026amp; Macaluso, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2006\u003c/span\u003e) but do not engage the frontal executive network (Fockert et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). We infer that in our task-irrelevant conditions, the oddball color\u0026rsquo;s signal likely remained confined to low-level visual areas and was prevented from influencing the decision process, essentially, category-based focusing provided \u0026ldquo;protective gating\u0026rdquo; against the irrelevant oddball. Conversely, when the oddball carried categorical significance (Exps 3\u0026ndash;4), it reliably disrupted performance: participants could not ignore a salient event that was directly tied to target identity. Under task-relevant conditions, rare-category oddballs drew attention and processing priority, consistent with the \u0026ldquo;attentional control settings\u0026rdquo; being tuned to that category. The goal-contingent capture framework (Zhang \u0026amp; Gaspelin, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) predicts that salient stimuli capture attention only if they match the observer\u0026rsquo;s active goal. Our data strongly support this principle in CAS. Indeed, the absence of an oddball effect in Exp. 1\u0026ndash;2 versus its presence in Exp. 3\u0026ndash;4 created a significant salience relevance \u0026times; salience interaction in the omnibus analysis. In practical terms, CAS is highly adaptive: it ignores irrelevant singletons but prioritizes (to a fault) unexpected targets.\u003c/p\u003e \u003cp\u003eIt is important to note that the lack of capture by irrelevant oddballs in our study does not mean salience is never potent, instead, it likely reflects the nature of our salience manipulation and the strength of top-down focus (Folk et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1992\u003c/span\u003e; Lavie, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). In classic additional-singleton paradigms (Theeuwes, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1992\u003c/span\u003e), a salient color distractor in a visual search can capture attention involuntarily. Why did our color oddball not capture attention? We suspect two reasons: first, our task was a coarse categorical judgment that might not require shifting attention to individual items, so the oddball distractor might have been processed superficially without drawing attention away from the overall category count. Second, our oddball was defined by probability across trials rather than a unique singleton in a display, so that, on any given trial, the odd-colored item was not actually more salient in a bottom-up sense than the others (all items were isoluminant, just different in color). Instead, it was the expectation violation over trials that constituted \u0026ldquo;salience\u0026rdquo;, consistent with predictive-coding accounts of attention (Summerfield \u0026amp; Egner, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Such probabilistic oddballs are known to have weaker and more delayed effects compared to outright singletons (Biggs et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). They can influence decision criteria or arousal, but they do not automatically capture spatial attention the way an abrupt onset or high-contrast singleton might. In our task-irrelevant case, apparently neither mechanism (arousal nor spatial capture) was sufficient to impact behavior. Thus, we interpret our null oddball findings in Exp. 1\u0026ndash;2 as evidence that categorical goals override subtle salience signals, but we also acknowledge that a more salient distractor (e.g., a flashing item or a uniquely shaped singleton) might have produced a different result. Supporting this, Wu et al. (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) did find attentional capture by an oddball in a feature-based task (they used a spatial cueing design with color singletons), but there the salient distractor likely shared some task relevance (e.g., a directional-cue context). Our findings diverged from Wu et al. (\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) precisely in that our distractors had no spatial or response relevance, allowing category-based focus to suppress them.\u003c/p\u003e \u003cp\u003eWhen salience was task-relevant in our study, the effects on CAS were unambiguous: performance suffered significantly on oddball trials. Participants took longer and were more error-prone when the majority category was the rare one. This can be interpreted as a form of \u0026ldquo;oddball cost\u0026rdquo; due to violated expectations, where observers were biased toward the frequent category and had to overcome this bias on oddball trials, leading to delays and mistakes (Biggs et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). The effect was akin to a contingent capture: attention might have momentarily been captured by the oddball item (e.g., focusing on it or re-checking it), or decision thresholds may have been adjusted (e.g., requiring more evidence on oddball trials due to lower prior probability, which could increase RT and errors). Neurocognitively, task-relevant oddballs engage frontal regions such as the anterior cingulate cortex that are implicated in conflict monitoring and cognitive control (Botvinick et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). This suggests our participants, upon encountering an oddball trial, had to invoke additional control processing (e.g., \u0026ldquo;Is this trial really an odd category?\u0026rdquo;) resulting in slower, less accurate responses. Interestingly, our data showed this oddball cost was generally larger in prototype CAS than semantic CAS (absolute RT/ACC differences). We discuss this further below, but it implies that even though prototype tasks are faster overall, they might experience a proportionally greater disruption from an unexpected event, perhaps because they rely on a more stimulus-driven mode of processing that is directly perturbed by any irregular input.\u003c/p\u003e \u003cp\u003eFrom a theoretical standpoint, our results emphasize that attentional selection in CAS is neither purely stimulus-driven nor purely goal-driven, but a conditional combination. Under high cognitive load, even typically \u0026ldquo;automatic\u0026rdquo; capture by salient stimuli can be eliminated, in line with load theory\u0026rsquo;s prediction that limited central resources curtail distractor processing (Lavie, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; de Fockert et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). Under normal load but with strong top-down goals, capture is also prevented, consonant with contingent-capture findings that attention prioritises goal-matching features (Folk et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1992\u003c/span\u003e; Bacon \u0026amp; Egeth, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e1994\u003c/span\u003e). Only when the salience coincides with an active goal do we see a full effect. Thus, CAS appears to implement a hierarchical priority map (cf. Wolfe, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) where task goals gate the influence of salient signals. In our semantic CAS, this meant that only when the oddball was categorically relevant did it make it onto the \u0026ldquo;priority map\u0026rdquo; to compete for attention.\u003c/p\u003e \u003cp\u003eFurthermore, our data provide insight into the stages of processing affected by salience in CAS. The semantic task under relevant oddball showed an interaction with load primarily in accuracy (not RT), suggesting that oddballs under high load caused participants to miss targets or misjudge the majority more often, consistent with a late-stage decision error or threshold adjustment, often indexed by the P3 component (Polich, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). Prototype task under relevant oddball showed a robust RT cost without an interaction with load, implying a more uniform perceptual slow-down due to oddballs; such effects are typically associated with extended N2pc or target-selection activity (Luck \u0026amp; Hillyard, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e1994\u003c/span\u003e; Eimer, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e1996\u003c/span\u003e). We cautiously interpret that task-relevant salience triggers late decision-stage conflict in semantic CAS, but earlier perceptual-stage competition in prototype CAS. We stress that our behavioural data cannot directly pinpoint neural timing, yet this inference accords with the differing accuracy patterns observed in the two cases. In any event, salience integration in CAS remains highly context-dependent, a theme elaborated in the next section.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e7.3 Integrative Effects of Top-Down Control and Salience on CAS\u003c/h2\u003e \u003cp\u003eOne of our motivating questions was whether top-down control and salience operate independently or interactively in CAS. Our experiments indicate strong interactive integration: cognitive load (top-down demand) and stimulus salience did not simply have additive effects; instead, they jointly determined performance in important ways (Lavie, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; de Fockert et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). Specifically, under task-relevant salience conditions, we observed that high cognitive load exacerbated the negative impact of oddballs, especially in the semantic task. When the rare-category oddball appeared and participants were already under strain (high load, juggling two categories in mind), the result was a pronounced conflict, which reflected in markedly increased decision time and errors. In our comprehensive analysis, this showed up as a three-way interaction: the difference between oddball and standard trials was greatest in the high-load, task-relevant, semantic condition. This finding can be framed as top-down and bottom-up factors competing for shared resources (Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Under easier conditions (low load), participants had the capacity to deal with an oddball fairly well (RT cost but small accuracy impact). Under maximal demand (high load), that oddball tipped performance over the edge, causing participants to struggle (slow and error-prone responses). This aligns with models where central resources (like WM or decision-making circuits) mediate between goal-directed and stimulus-driven inputs (Duncan \u0026amp; Humphreys, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e1989\u003c/span\u003e). Our data resonate with Fan\u0026rsquo;s (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e) information-theoretic account and others, wherein cognitive control under uncertainty must arbitrate between internal goals and external salient events. In our case, high load amplified this arbitration conflict, particularly when an oddball signaled a deviation from the expected rule.\u003c/p\u003e \u003cp\u003eConversely, in task-irrelevant conditions, no such integration or competition occurred: load effects were present (because that\u0026rsquo;s internally driven), but salience did nothing and thus there was no additional conflict. This emphasizes that the cognitive system can effectively segregate an irrelevant source of salience, preventing it from draining resources when focus is required on something else. Our findings here dovetail with Lavie\u0026rsquo;s load theory (Fockert et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2004\u003c/span\u003e), which posits that under high cognitive load, processing of irrelevant stimuli is reduced. We observed that even under low load, irrelevant oddballs didn\u0026rsquo;t matter. This likely because participants strategically ignored color from the start (given it was never useful), illustrating a strong top-down set (Folk et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1992\u003c/span\u003e). Under high load, that top-down filtering may have been even more absolute (as attention was fully occupied). Thus, when salience was irrelevant, top-down control effectively nullified any integration. This indicates that the two operated on separate tracks, with category decisions proceeding unaffected by oddball presence.\u003c/p\u003e \u003cp\u003eOur results also shed light on how category architecture (semantic vs. prototype) influences this integration. We found that semantic-based CAS and prototype-based CAS, despite similar overall trends, had quantitative and qualitative differences in their responses to load\u0026thinsp;+\u0026thinsp;salience. Semantic CAS under high load exhibited what we interpret as conceptual interference, needing to resolve conflicting category cues (majority vs. oddball minority) perhaps via executive processes (hence more errors when taxed). Prototype CAS under high load showed a slightly different pattern: oddly, the presence of high load reduced the relative cost of oddballs in some measures (e.g., the interaction in accuracy was smaller in Exp 4 than Exp 3). This hints that under high load, prototype searchers might have inadvertently filtered some distractors, even possibly the oddball shape, by focusing on the dominant perceptual features (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Another way to view it is through perceptual-load theory: when prototype displays were heterogeneous (high load), visual attention was likely more occupied, leaving less room for the oddball shape to capture attention. This mechanism analogous to high perceptual load reducing distractibility (Lavie, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). Indeed, we speculated earlier that in Exp 4, high load didn\u0026rsquo;t increase oddball cost; intriguingly, it slightly decreased the oddball effect on accuracy relative to low load (the prototype oddball effect was big at both loads, but not bigger at high load). This could reflect that participants, when confronted with a difficult prototype array (2 : 1 with a rare shape), sometimes missed the oddball shape entirely or treated it as just another distractor, whereas in easier arrays (low load), the oddball shape stood out more and thus interfered relatively more. In any case, category architecture modulated how load and salience interplay. Semantic tasks brought the conflict to a head at decision-making (leading to late-stage interference under combined strain), whereas prototype tasks handled it in a more distributed, perceptual manner (steady interference that didn\u0026rsquo;t double-dip with load). This distinction mirrors Ashby et al.\u0026rsquo;s (2004, 2005) dual-process account where rule-based systems (semantic CAS) rely on WM and executive control, thereby two sources of demand (load and oddball) overload the system. Whereas, similarity-based systems (prototype CAS) rely more on visual processing, thereby they can maintain performance speed but will show interference in accuracy that doesn\u0026rsquo;t scale massively with cognitive load.\u003c/p\u003e \u003cp\u003eFinally, by extending Memelink \u0026amp; Hommel\u0026rsquo;s (2013) intentional-weighting theory to our results, we can say that the \u0026ldquo;weighting\u0026rdquo; of task features in CAS is highly flexible. When the feature \u0026ldquo;category frequency\u0026rdquo; was irrelevant, it was weighted at zero (oddball frequency ignored). When it became relevant, that feature\u0026rsquo;s weight shot up and influenced both perception and decision. CAS thus dynamically gates incoming information based on task demands, balancing conceptual precision and perceptual efficiency as needed. In high-load semantic CAS, the system favored conceptual precision (hence ignoring oddballs until it perhaps was too late, causing conflicts); in prototype CAS, the system leaned on perceptual processes (hence oddballs always intruded a bit, but decisions were fast). This adaptive gating underscores the context-dependent nature of attentional selection in categorical tasks.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e7.4 Dual-Pathway CAS Model Involving Conceptual and Perceptual Hierarchies\u003c/h2\u003e \u003cp\u003eBringing these findings together, we propose a dual-pathway model for CAS that reflects two hierarchies: a conceptual hierarchy (semantic CAS) and a perceptual hierarchy (prototype CAS). Our results provide direct evidence that prototype-based CAS can systematically outperform semantic-based CAS across varying conditions (Lech et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Wu et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). This was evident in the main effect of category type on IES (prototypes had overall lower IES) and in their higher resilience in absolute performance even when oddballs and load were present. This supports the idea that prototype categories tap into fast, efficient visual processing strategies (a \u0026ldquo;perceptual pathway\u0026rdquo;) that do not require heavy WM involvement (Giammarco et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). On the other hand, semantic categories rely on a \u0026ldquo;conceptual pathway\u0026rdquo; that uses learned rules and is tied to WM and long-term memory retrieval (Fan, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). This makes semantic CAS more accurate but slower, and more vulnerable to WM depletion. For example, in our experiments the semantic tasks maintained higher accuracy under oddball conditions by incurring an RT cost (especially under high load), whereas the prototype tasks tended to respond quickly at the expense of some accuracy. This trade-off aligns with the notion that semantic CAS prioritizes rule-based precision (ensuring the category decision is correct, even if slower), while prototype CAS prioritizes speed/efficiency (leveraging visual similarities to respond quickly, even if occasionally an oddball leads to an error) (Ashby \u0026amp; Maddox, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2004\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe dual-pathway model is supported by the three-way interaction of category \u0026times; load \u0026times; salience in our data. It extends Ashby\u0026rsquo;s dual-system theory (rule vs. exemplar) to attentional selection: under resource competition, semantic (rule-based) systems slow down to maintain accuracy, whereas prototype (similarity-based) systems sacrifice a bit of accuracy to maintain speed (Ashby \u0026amp; Valentin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). We observed exactly that pattern. This divergence likely stems from the underlying neural circuits: semantic CAS engages prefrontal-parietal circuits (associated with executive control and sustained rule maintenance), which have limited capacity (Miller \u0026amp; Cohen, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). Once that capacity is strained (by load or dual-tasking), performance slows and any additional conflict (like an oddball) further taxes the system. Prototype CAS, conversely, relies on occipitotemporal circuits (associated with visual pattern recognition and grouping), which operate in parallel and with high efficiency (Ungerleider \u0026amp; Haxby, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e1994\u003c/span\u003e). These circuits can rapidly integrate features and are less disrupted by multitasking until a certain point; thus, performance remains quick, though not immune to errors if conflicting inputs (like an oddball shape) are present. Indeed, evidence of occipital alpha-band suppressions unique to prototype tasks (suggesting heightened visual processing) would be a neural marker to confirm this pathway (Worden et al., \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2000\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eCrucially, our model posits that semantic CAS and prototype CAS are not just quantitatively different, but qualitatively employ distinct control strategies. Semantic CAS behaves like a \u0026ldquo;late selection\u0026rdquo; system and it filters and decides at a later stage, allowing it to ignore irrelevant salience entirely but struggling if multiple demands coincide (late-stage conflict) (Luck \u0026amp; Vogel, 1997). Prototype CAS behaves more like an \u0026ldquo;early selection\u0026rdquo; system, and it integrates all perceptual inputs (including oddballs) early on, leading to immediate competition but perhaps resolving it at a sensory level (early competition resolved via perceptual coding) (Eimer, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e1996\u003c/span\u003e). This dichotomy extends Wolfe\u0026rsquo;s guided search: semantic CAS adds a top-down rule-based \u0026ldquo;priority map\u0026rdquo; in prefrontal cortex, whereas prototype CAS leans on feature-based \u0026ldquo;maps\u0026rdquo; in visual cortex (Wolfe \u0026amp; Horowitz, \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Under our dual-pathway model, when faced with distractions or load, the semantic pathway will enforce its priorities through frontal control (slowing things down but keeping on task), and the prototype pathway will continue to leverage fast sensory processing (maintaining speed but letting in some distraction). Neither pathway is inherently superior; each excels under certain conditions. Semantic CAS might dominate in situations requiring high accuracy and when distractors can be categorically excluded (since it can completely ignore irrelevant stimuli given enough focus). Prototype CAS might excel in visually complex scenes or when speed is essential, as it can quickly home in on target-like features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e7.5 Strengths, Limitations and implications\u003c/h2\u003e \u003cp\u003eThis study provides a novel, integrated perspective by directly comparing semantic and prototype-based attention under systematically varied load and salience conditions. We demonstrated a clear goal-dependence of salience effects in CAS and empirically distinguished two modes of category-guided attention, thereby resolving to some extent the debate on whether CAS is more like feature-based attention (perceptual) or like executive control (conceptual), and our findings suggest it can be both, depending on category type and context (Folk et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e1992\u003c/span\u003e; Zhang \u0026amp; Gaspelin, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The dual-pathway model emerging from our results offers a framework for future research on how different category representations (rule-based vs. similarity-based) recruit distinct neural resources and cope with distractions (Ashby \u0026amp; Valentin, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). Our use of multiple experiments and a combined analysis is a strength, as it allowed us to see consistent patterns and cross-experiment interactions that a single experiment might miss (Efron \u0026amp; Tibshirani, \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1993\u003c/span\u003e). We also introduce an important methodological point: manipulating salience via probability (oddball frequency) in a CAS task is a subtle but insightful way to study contingent capture in a categorical domain, complementing traditional singleton methods (Biggs, Adamo, \u0026amp; Mitroff, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eHowever, several limitations should be acknowledged. First, while our salience manipulation via color probability provided a novel angle, it deviates from classic definitions of \u0026ldquo;physical salience\u0026rdquo;. As discussed, a uniquely colored item in a display can capture attention, whereas an oddball color event across trials may not, and our interpretation of \u0026ldquo;no capture\u0026thinsp;=\u0026thinsp;top-down filtering\u0026rdquo; rests on the assumption that our oddball was a meaningful bottom-up cue (Theeuwes, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1992\u003c/span\u003e). It might be argued that the oddball color simply wasn\u0026rsquo;t salient enough; thus, our conclusion of protective gating in semantic CAS should be tempered by that consideration. Future work could incorporate a true singleton distractor in a category search to confirm that semantic guidance can overcome even strong singletons (Zhang \u0026amp; Gaspelin, \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2024\u003c/span\u003e suggests it can under certain conditions). Second, our cognitive-load manipulation was labeled as such because of prior literature and our intention (more items to hold/compare\u0026thinsp;=\u0026thinsp;higher load). Nevertheless, as reviewers pointed out, this \u0026ldquo;load\u0026rdquo; may also reflect decision difficulty or perceptual load, not purely working-memory load (Lavie, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). In a 2 : 1 display, one must scrutinize the stimuli more carefully to identify the majority, thereby engage perceptual processes as much as WM. Thus, while we interpret our load effects in terms of WM depletion, an alternative view is that high-load trials increased task difficulty overall, and our results do not exclusively pinpoint WM (de Fockert et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). Future studies could separate memory load (e.g., requiring an actual memory set to hold) from decision difficulty to see if the effects differ. Our use of the term \u0026ldquo;cognitive load\u0026rdquo; should be understood as the combined challenge of a heterogeneous display requiring suppression of an oddball item, which indeed could involve both WM (keeping the category rule in mind) and additional decision effort. We have reframed some of our interpretations accordingly in this revision.\u003c/p\u003e \u003cp\u003eIt\u0026rsquo;s also worth noting that our interpretations of \u0026ldquo;early\u0026rdquo; vs. \u0026ldquo;late\u0026rdquo; stage processing differences between tasks are inferences based on behavioural patterns and existing literature. Without direct neurophysiological measures (EEG/ERP or eye-tracking), we cannot conclusively state at what processing stage the load\u0026ndash;salience interactions occur for each category type. Future experiments could record ERPs in a similar paradigm to see if, for example, the N2pc (an index of early attentional allocation) is modulated by oddballs more in prototype tasks, whereas the P3 (decision-related) is more modulated in semantic tasks (Luck, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Polich, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). Our discussion of this point is meant to generate hypotheses rather than final answers.\u003c/p\u003e \u003cp\u003eDespite these limitations, our study has important implications. From a theoretical standpoint, it reconciles prior mixed results on attentional capture in categorical search by showing that task context (irrelevant vs. relevant) is key (Theeuwes, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). It also bridges literature on visual search and working memory by examining their interplay in CAS. Practically, understanding how different kinds of categories handle distraction can inform user-interface design or training protocols. For example, if one needs to design a display to grab a person\u0026rsquo;s attention during a task, our results suggest that the cue must be relevant to their task or else it may be ignored; conversely, if one wants to minimise distraction, ensuring that salient visual features are unlinked to the user\u0026rsquo;s goals will help them stay focused (Wickens \u0026amp; McCarley, \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). In applied settings such as airport baggage screening (a semantic category search) or medical image analysis (often more prototype-like, searching for anomalous shapes), different strategies might be required to mitigate distractions or manage cognitive load.\u003c/p\u003e \u003cp\u003eIn conclusion, category-guided attention operates via dual pathways that differentially integrate top-down and bottom-up influences. Semantic CAS functions like a focused spotlight, governed by WM and executive control, highly effective at excluding irrelevant stimuli but at risk of overload when multitasking. Prototype CAS functions like a wide-angle filter, rapidly responsive to visual features and robust in speed, yet allowing more bottom-up intrusion. Understanding these dual pathways enriches our comprehension of attention in complex, real-world tasks and could guide the development of techniques to improve attention, for example through adaptive interfaces or training that either bolsters rule maintenance or harnesses perceptual fluency as appropriate.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by National Education Scientific Planning Project (DBA230368).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Interest statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflict of interest.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eX. W., H. Z., and X. S. conceived the study; X. W., Y. L and X. S. analyzed the data. All authors discussed the results and contributed to the writing of the article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of generative AI and AI-assisted technologies in the writing process\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this work the authors used Deepseek (https://www.deepseek.com/) in order to improve the readability and language of the manuscript. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOpen practices statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData, scripts, materials, and analyses are available at https://osf.io/6wvqc/\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorrespondence\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHuan Zhang, Faculty of Psychology, Tianjin Normal University, Tianjin, China.\u003c/p\u003e\n\u003cp\u003eEmail: [email protected]\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAlvarez, G. A., \u0026amp; Cavanagh, P. (2004). The capacity of visual short-term memory is set both by visual information load and by number of objects. \u003cem\u003ePsychological Science\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(2), 106\u0026ndash;111.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAshby, F. G., Maddox, W. T., \u0026amp; Bohil, C. J. (2002). Observational versus feedback training in rule-based and information-integration category learning. \u003cem\u003eMemory \u0026amp; Cognition\u003c/em\u003e, \u003cem\u003e30\u003c/em\u003e(5), 666\u0026ndash;677.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAshby, F. G., \u0026amp; Maddox, W. T. (2004). Human category learning. \u003cem\u003eAnnual Review of Psychology\u003c/em\u003e, \u003cem\u003e56\u003c/em\u003e(1), 149\u0026ndash;178.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAshby, F. G., \u0026amp; Valentin, V. V. (2005). Multiple systems of perceptual category learning: theory and cognitive tests. \u003cem\u003eHandbook of Categorization in Cognitive Science\u003c/em\u003e, 547\u0026ndash;572.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBacon, W. F., \u0026amp; Egeth, H. E. (1994). Overriding stimulus-driven attentional capture. \u003cem\u003ePerception \u0026amp; Psychophysics\u003c/em\u003e, \u003cem\u003e55\u003c/em\u003e, 485\u0026ndash;496.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiggs, A. T., Adamo, S. H., \u0026amp; Mitroff, S. R. (2014). Rare, but obviously there: effects of target frequency and salience on visual-search accuracy. \u003cem\u003eActa Psychologica\u003c/em\u003e, \u003cem\u003e152\u003c/em\u003e, 158\u0026ndash;165.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBotvinick, M. M., Cohen, J. D., \u0026amp; Carter, C. S. (2004). Conflict monitoring and anterior cingulate cortex: an update. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 539\u0026ndash;546.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, X., Xu, B., Chen, Y., Zeng, X., Zhang, Y., \u0026amp; Fu, S. (2023). Saliency affects attentional capture and suppression of abrupt-onset and color singleton distractors: Evidence from event‐related potential studies. \u003cem\u003ePsychophysiology\u003c/em\u003e, 60(8), e14290.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChu, H., \u0026amp; Zhou, X. (2004). Attentional capture and top-down processes. \u003cem\u003eAdvances in Psychological Science\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(5), 680\u0026ndash;687.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClark, V. P., Fannon, S., Lai, S., Benson, R., \u0026amp; Bauer, L. (2000). Responses to rare visual target and distractor stimuli using event-related fmri. \u003cem\u003eJournal of Neurophysiology\u003c/em\u003e, \u003cem\u003e83\u003c/em\u003e(5), 3133\u0026ndash;3139.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCohen, J. (1992). A power primer. \u003cem\u003ePsychological Bulletin\u003c/em\u003e, \u003cem\u003e112\u003c/em\u003e(1), 155\u0026ndash;159.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Fockert, J. W., Rees, G., Frith, C. D., \u0026amp; Lavie, N. (2001). The role of working memory in visual selective attention. \u003cem\u003eScience\u003c/em\u003e, \u003cem\u003e291\u003c/em\u003e, 1803\u0026ndash;1806.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDownar, J., Crawley, A. P., Mikulis, D. J., \u0026amp; Davis, K. D. (2001). The effect of task relevance on the cortical response to changes in visual and auditory stimuli: an event-related fMRI study. \u003cem\u003eNeuroImage\u003c/em\u003e, \u003cem\u003e14(\u003c/em\u003e6),1256\u0026ndash;1267.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDuncan, J., \u0026amp; Humphreys, G. W. (1989). Visual search and stimulus similarity. \u003cem\u003ePsychological Review\u003c/em\u003e, \u003cem\u003e96\u003c/em\u003e(3), 433\u0026ndash;458.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEfron, B., \u0026amp; Tibshirani, R. J. (1993). \u003cem\u003eAn Introduction to the Bootstrap\u003c/em\u003e. Chapman \u0026amp; Hall/CRC.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEimer, M. (1996). The N2pc component as an indicator of attentional selective processing. \u003cem\u003eElectroencephalography and Clinical Neurophysiology\u003c/em\u003e, \u003cem\u003e99\u003c/em\u003e, 225\u0026ndash;234.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEll, S. W., \u0026amp; Ashby, F. G. (2004). Dynamical trajectories in category learning. \u003cem\u003ePerception \u0026amp; Psychophysics\u003c/em\u003e, \u003cem\u003e66\u003c/em\u003e(8), 1318\u0026ndash;1340.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan, J., Guise, K. G., Xun, L., Wang, H., \u0026amp; Jan, L. (2008). Searching for the majority: algorithms of voluntary control. \u003cem\u003ePlos One\u003c/em\u003e, 3(10), e3522.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFan, J. (2014). An information theory account of cognitive control. \u003cem\u003eFrontiers in Human Neuroscience\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 680.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFockert, J. D., Rees, G., Frith, C., \u0026amp; Lavie, N. (2004). Neural correlates of attentional capture in visual search. \u003cem\u003eJournal of Cognitive Neuroscience\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(5), 751\u0026ndash;759.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFolk, C. L., Remington, R. W., \u0026amp; Johnston, J. C. (1992). Involuntary covert orienting is contingent on attentional control settings. \u003cem\u003eJournal of Experimental Psychology: Human Perception and Performance\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e, 1030\u0026ndash;1044.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiammarco, M., Schmidtke, J., \u0026amp; Arnell, K. (2016). Dissociating semantic and perceptual contributions to the suppression of distractors during working memory maintenance. \u003cem\u003eJournal of Experimental Psychology: Human Perception and Performance\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e, 1550\u0026ndash;1562.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGoldstone, R. L. (1998). Perceptual learning. \u003cem\u003eAnnual Review of Psychology\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e, 585\u0026ndash;612.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, H., Chen, Y., Li, X., Hu, X., Wang, J., Wu, T., \u0026amp; Guan, Q. (2021). Decline in the integration of top-down and bottom-up attentional control in older adults with mild cognitive impairment. \u003cem\u003eNeuropsychologia\u003c/em\u003e, \u003cem\u003e161\u003c/em\u003e, 108014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIndovina, I., \u0026amp; Macaluso, E. (2006). Dissociation of stimulus relevance and saliency factors during shifts of visuospatial attention. \u003cem\u003eCerebral Cortex\u003c/em\u003e, \u003cem\u003e17\u003c/em\u003e(7), 1701\u0026ndash;1711.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLech, R. K., Gunturkun, O., \u0026amp; Suchan, B. (2016). An interplay of fusiform gyrus and hippocampus enables prototype- and exemplar-based category learning. \u003cem\u003eBehavioural Brain Research\u003c/em\u003e, \u003cem\u003e311\u003c/em\u003e, 239\u0026ndash;246.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLavie, N. (2005). Distracted and confused? Selective attention under load. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 75\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuck, S. J. (2014). \u003cem\u003eAn Introduction to the Event-Related Potential Technique\u003c/em\u003e (2nd ed.). MIT Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuck, S. J., \u0026amp; Hillyard, S. A. (1994). Spatial filtering during visual search: evidence from human electrophysiology. \u003cem\u003eJournal of Experimental Psychology: Human Perception and Performance\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e, 1000\u0026ndash;1014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLuck, Steven, J., Vogel, \u0026amp; Edward, K. (1997). The capacity of visual working memory for features and conjunctions. \u003cem\u003eNature\u003c/em\u003e, \u003cem\u003e390\u003c/em\u003e(6657), 279\u0026ndash;281.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGiammarco, M., Paoletti, A., Guild, E. B., \u0026amp; Al-Aidroos, N. (2016). Attentional capture by items that match episodic long-term memory representations. \u003cem\u003eVisual Cognition\u003c/em\u003e, 1\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMemelink, J., \u0026amp; Hommel, B. (2012). Intentional weighting: a basic principle in cognitive control. \u003cem\u003ePsychological Research\u003c/em\u003e, \u003cem\u003e77\u003c/em\u003e(3).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiller, E. K., \u0026amp; Cohen, J. D. (2001). An integrative theory of prefrontal cortex function. \u003cem\u003eAnnual Review of Neuroscience\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e, 167\u0026ndash;202.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMihajlović, N., \u0026amp; Zdravković, S. (2024). Contingent capture by color is sensitive to categorical color perception. \u003cem\u003eAttention Perception \u0026amp; Psychophysics\u003c/em\u003e, \u003cem\u003e86\u003c/em\u003e(1), 36\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePeelen, M. V., \u0026amp; Kastner, S. (2014). Attention in the real world: toward understanding its neural basis. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e18\u003c/em\u003e(5), 242\u0026ndash;250.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePolich, J. (2007). Updating P300: an integrative theory of P3a and P3b. \u003cem\u003eClinical Neurophysiology\u003c/em\u003e, \u003cem\u003e118\u003c/em\u003e, 2128\u0026ndash;2148.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReeder, R. R., \u0026amp; Peelen, M. V. (2013). The contents of the search template for category-level search in natural scenes. \u003cem\u003eJournal of vision\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e, 1\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShang, L., Yeh, L. C., Zhao, Y., Wiegand, I., \u0026amp; Peelen, M. V. (2024). Category-based attention facilitates memory search. \u003cem\u003eEneuro\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(2).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSummerfield, C., \u0026amp; Egner, T. (2009). Expectation (and attention) in visual cognition. \u003cem\u003eTrends in Cognitive Sciences\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e, 403\u0026ndash;409.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTheeuwes, J. (1992). Perceptual selectivity for color and form. \u003cem\u003ePerception \u0026amp; Psychophysics\u003c/em\u003e, \u003cem\u003e51\u003c/em\u003e, 599\u0026ndash;606.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTheeuwes, J. (2010). Top-down and bottom-up control of visual selection. \u003cem\u003eActa Psychologica\u003c/em\u003e, \u003cem\u003e135\u003c/em\u003e, 77\u0026ndash;99.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTownsend, J. T., \u0026amp; Ashby, F. G. (1983). \u003cem\u003eStochastic modelling of elementary psychological processes\u003c/em\u003e. Cambridge University Press.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUngerleider, L. G., \u0026amp; Haxby, J. V. (1994). What\u0026rsquo; and \u0026lsquo;where\u0026rsquo; in the human brain. \u003cem\u003eCurrent Opinion in Neurobiology\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e, 157\u0026ndash;165.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWickens, C. D., \u0026amp; McCarley, J. S. (2008). \u003cem\u003eApplied Attention Theory\u003c/em\u003e. CRC.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolfe, J. M. (2012). Guided search 4.0: current progress with a model of visual search. \u003cem\u003eIntegrated Models of Cognitive Systems\u003c/em\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolfe, J. M., \u0026amp; Horowitz, T. S. (2017). \u003cem\u003eFive factors that guide attention in visual search. In M. W. Current Topics in Behavioral Neurosciences (Vol. 41, pp. 117\u0026ndash;141).\u003c/em\u003e Springer.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWoodman, G. F., Carlisle, N. B., \u0026amp; Reinhart, R., M. G (2013). Where do we store the memory representations that guide attention? \u003cem\u003eJournal of Vision\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(3), 1, 1\u0026ndash;17.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorden, M. S., Foxe, J. J., Wang, N., \u0026amp; Simpson, G. V. (2000). Anticipatory biasing of visuospatial attention indexed by retinotopically specific alpha-band electroencephalography increases over occipital cortex. \u003cem\u003eJournal of Neuroscience\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e, RC63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, Q., Chang, C., Xi, S., Huang, I., Liu, Z., Juan, C., et al. (2015). A critical role of temporoparietal junction in the integration of top-down and bottom‐up attentional control. \u003cem\u003eHuman Brain Mapping\u003c/em\u003e, \u003cem\u003e36\u003c/em\u003e(11), 4317\u0026ndash;4333.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, T., Chen, C., Spagna, A. (2020). The functional anatomy of cognitive control: A domain-general brain network for uncertainty processing. \u003cem\u003eJournal of Comparative Neurology\u003c/em\u003e, \u003cem\u003e528\u003c/em\u003e(8).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, X., Liu, X., \u0026amp; Fu, S. (2016). Feature- and category-specific attentional control settings are differently affected by attentional engagement in contingent attentional capture. \u003cem\u003eBiological Psychology\u003c/em\u003e, \u003cem\u003e118\u003c/em\u003e, 8\u0026ndash;16.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, X., Wang, X., Saab, R., \u0026amp; Jiang, Y. (2020). Category-based attentional capture can be influenced by colorand shape-dimensions independently in the conjunction search task. \u003cem\u003ePsychophysiology\u003c/em\u003e, \u003cem\u003e57\u003c/em\u003e(4).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWyble, B., Folk, C., \u0026amp; Potter, M. C. (2013). Contingent attentional capture by conceptually relevant images. \u003cem\u003eJournal of experimental psychology Human perception and performance\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e(3), 861\u0026ndash;871.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, H., \u0026amp; Zelinsky, G. J. (2009). Visual search is guided to categorically defined targets. \u003cem\u003eVision Research\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e, 2095\u0026ndash;2103.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYildirim, K., Akalin-Baskaya, A., \u0026amp; Hidayetoglu, M. L. (2007). Effects of indoor color on mood and cognitive performance. \u003cem\u003eBuilding \u0026amp; Environment\u003c/em\u003e, \u003cem\u003e42\u003c/em\u003e(9), 3233\u0026ndash;3240.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, Y., \u0026amp; Gaspelin, N. (2024). Salience effects on attentional selection are enabled by task relevance. \u003cem\u003eJournal of Experimental Psychology: Human Perception and Performance\u003c/em\u003e, \u003cem\u003e50\u003c/em\u003e(11), 1131\u0026ndash;1142.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"category-guided attentional selection, top-down, stimulus salience, prototype-based category","lastPublishedDoi":"10.21203/rs.3.rs-6888298/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6888298/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eCategory-guided attentional selection (CAS) enables efficient information filtering in complex environments through cognitive category formation. However, systematic comparisons between semantic and prototype-based CAS mechanisms are critically lacking, particularly in how they dynamically integrate top-down control and salience processing. Combining a hybrid experimental design (4 experiments, N\u0026thinsp;=\u0026thinsp;80) manipulating cognitive load (high vs. low), stimulus salience (oddball vs. standard) and salience relevance (task-related vs. task-irrelevant), we dissected CAS dynamics across semantic (numbers/letters) and prototype-based (\"O\"/\"X\" shapes) categories. High cognitive load consistently impaired performance, demonstrating working memory-dependent top-down modulation. Salience effects emerged exclusively under task-relevant conditions, with oddball stimuli eliciting slower RTs and reduced accuracy, indicating goal-contingent weighting. Critically, cognitive load and task-relevant salience interacted during late decision stages, suggesting dynamic resource competition between control processes. Notably, prototype-based categories outperformed semantic categories in oddball trials, likely driven by perceptual similarity-mediated bottom-up integration. These findings extends a refinement to theories of category-guided attention into a dual-pathway model, where semantic categories rely on conceptual templates in working memory, while prototype-based categories leverage perceptual feature binding.\u003c/p\u003e","manuscriptTitle":"Dual Pathways of Category-Guided Attentional Selection: Task- Relevance Modulates Hierarchical Integration of Top-Down Control and Salience","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-20 07:31:36","doi":"10.21203/rs.3.rs-6888298/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e38140f0-05bc-4b1f-bf54-9b5b957f6662","owner":[],"postedDate":"June 20th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-16T13:54:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-20 07:31:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6888298","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6888298","identity":"rs-6888298","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}