{"paper_id":"232a238a-cb6f-4a83-b935-bc1f7f516aad","body_text":"Dynamic competition between selective attention and \nspatial prediction during visual search \nFloortje G. Bouwkamp, Jorie J.G. van Haren, Floris P. de Lange*, and Eelke Spaak* \n \n*) These authors contributed equally. \nDonders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen \nWe have no conflict of interest to declare. \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nAbstract  \nDuring visual search, we rely on both selective attention and spatial predictions to guide our \nbehavior. However, whether and how these mechanisms interact is largely unclear. Using a \ncontextual cueing paradigm, we investigated whether learning and exploitation of spatial predictive \ncontext can occur outside attentional focus. Repeating search scenes enabled distractor context to \nserve as a contextual cue predicting target location. Participants searched for a target among \ndistractors in two colors: the same as the target-to-be-searched (attended context) or a different \ncolor (ignored context). Halfway through the experiments, we changed the target color, thereby \naltering the attentional status of distractor contexts while maintaining their spatial predictiveness. In \nExperiment 1, where participants regularly switched between target colors, we found exploitation of \nspatial predictive context both within and outside attentional focus. However, in Experiment 2, where \nattention was more stably focused on one target color, only the attended predictive context was \nexploited before transfer. Intriguingly, after transfer, previously ignored predictive context showed \nimmediate benefits, revealing latent learning. These findings demonstrate a dynamic competition \nbetween selective attention and spatial predictions: while learning occurs independently of attention, \nexploitation may require attentional selection. Our results suggest that selective attention gates the \ninfluence of spatial predictions on behavior, with gating strength determined by the stability of \nattentional control. \nKeywords: Visual search, selective attention, spatial prediction, contextual cueing \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nIntroduction \nOur visual environment is usually crowded, making it impossible to process everything at once. Being \nable to locate a relevant item in such crowded scenes, something we refer to as visual search, is a \nubiquitous skill needed for goal-driven perception in everyday life.  \nWe know that selective attention is a crucial factor for efficient processing of our (visual) \nenvironment. By filtering out irrelevant items, we can devote cognitive resources to what matters for \nthe goal at hand. For example, when searching for a phone, only items similar to the shape and color \nof the phone will be highlighted, speeding up the search process. \nAdditionally, we can make use of the predictable structure in our environment, as items tend to co-\noccur. These co-occurrence relations can generate predictions about what to expect and where to \nexpect it (Chun & Turk-Browne, 2007; Peelen et al., 2023; Spaak et al., 2022; Theeuwes et al., \n2022; Võ et al., 2019). To return to the previous example: when searching for a phone, specifically \ncounter- and table tops are likely to be searched, as phones are most often found there. The \ndetection of these types of regularities is generally thought to be implicit and the exploitation of such \nregularities a highly automatic process. This type of learning is called visual statistical learning (VSL).  \nWithin visual search, statistical learning of environmental regularities has been demonstrated to \nboost performance in several ways. For example, spatial regularities may cue the most likely location \nof a target, a.k.a location probability cueing (Geng & Behrmann, 2005; Geyer et al., 2024; Y. V. Jiang \net al., 2013), or set up expectations about where distracting information will be (Beesley et al., 2016; \nFerrante et al., 2023; Richter et al., 2024). These types of cueing create a fixed attentional bias in \nspace. Another type of cueing within visual search is contextual cueing (Chun & Jiang, 1998a). In this \ntype of statistical learning, the spatial context within visual search scenes predicts where the target \nwill be, thereby setting up an implicit spatial relationship (Y. V. Jiang, 2018).  \nIn a typical contextual cueing experiment, participants need to search for a target (often a letter T) \nthat is embedded in a scene filled with similar-looking shapes (often rotated Ls). Some of these \nscenes are repeated over the course of the experiment, and although participants demonstrate little \nor no awareness of these repetitions (but see Meyen et al., 2024; Vadillo et al., 2016), they become \nmarkedly faster in finding the target in these repeated ‘old’ scenes compared to ‘new’ scenes. The \nconsensus is that after learning the spatial relationship between distractors and target, this \nknowledge is exploited, which leads to more efficient guidance of attention when searching these \nscenes (Bouwkamp et al., 2025; Goujon et al., 2015; Y. V. Jiang et al., 2019; Sisk et al., 2019; \nSpaak & de Lange, 2020). In other words, because the spatial context is predictive of target location, \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nparticipants become more efficient at visual search. The underlying mechanisms of attention and \nprediction are unified in the concept of an attentional priority map. Bottom-up salience, goal-driven \nrelevance, and prior experience together shape these priority maps (Duncan et al., 2023; Fecteau & \nMunoz, 2006; Ferrante et al., 2018; Sisk et al., 2019; Wolfe, 2021). These maps help us allocate \nattentional resources to whatever needs to be prioritized. But if, and how selective attention and \npredictions interact is largely unclear.  \nOne big question that has occupied the field is how automatic visual statistical learning is. More \nspecifically, is attention is necessary for statistical learning to occur? When J.R. Saffran (1996) \nshowed that children as young as 8 months old could extract regularities while passively listening to \nspeech streams, this mechanism was thought to be highly automatic and attention was assumed to \nbe unnecessary for this type of learning to occur. However, in the visual domain, it has been \nproposed that statistical learning, and contextual cueing specifically, though implicit, requires \nattentional selection (Y. Jiang & Chun, 2001; Turk-Browne et al., 2005), limiting the scope of \nlearning. If, instead, we could learn implicitly even from things we are ignoring, this would be \nincredibly useful, as it would be a less costly operation that could potentially run ‘in the background’. \nMoreover, the acquired knowledge might become relevant and thus beneficial at some future point \nin time. This attractive idea has led to multiple efforts to further investigate the role of selective \nattention in statistical learning. Both in the auditory and visual domain, statistical learning has been \ndemonstrated to still occur when a secondary demanding task is requiring attention (Batterink & \nPaller, 2019; Musz et al., 2015), contradicting previous findings. Similarly, contextual cueing \nappears to be robust to distracting secondary tasks (Vicente-Conesa et al., 2022; Vickery et al., \n2010).  \nA third stance is that selective attention is not required for statistical learning in visual search, but it \nis for the exploitation of what is learned. This has been suggested for contextual cueing when loading \n(visual) working memory with a secondary task (Goujon et al., 2015; Pollmann, 2019; Sisk et al., \n2019), but also for contextual cueing of ignored visual context (Jiang & Leung, 2005). In the latter \nstudy, distractors in search scenes were present in two colors; however, the target was always in one \nand the same color. This narrowed down the search space as participants could safely ignore \ndistractors in the non-target color (Wolfe, 2021). Contextual cueing did not occur for context that was \nbeing ignored, replicating the previous finding of Jiang & Chun (2001). However, when the task \nrelevance of the distractor context was reversed by a change of color, there was a behavioral benefit \nwhen the previously ignored context was predictive. Jiang & Leung concluded that spatial predictive \ncontext was in fact learned, but that this learning was exploited only when the context became task-\nrelevant and thus attended. This is thought to be evidence of ‘latent learning’. This latent learning \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nphenomenon however failed to replicate recently (Vadillo et al., 2020), potentially calling into \nquestion its robustness.  \nIn daily life, we change our search goals all the time (e.g., where is my phone, where is the traffic \nlight, where is my friend etc.), while contextual regularities tend to stay stable over time (a room will \nnot suddenly change color or shape). Though a feature change is inevitable in a transfer-paradigm, \nchanging the color of the context might be more impactful than changing the color of the target. \nMoreover, for the association between context and target to become apparent after a change of \ncontext color, the association must be (latently) learned independently from its color feature. \nAlthough it seems that mainly global configuration is learned, contextual cueing is sensitive to \nchanges of perceptual identity of distractor items (Chun & Jiang, 1998a; Jiang & Wagner, 2004; \nSpaak & de Lange, 2020). If not only location but also color of distractor context is encoded during \ncontextual learning (Jiang & Song, 2005), a color change of the context could be disruptive to the \nassociation between context and target, potentially erasing any contextual cueing.  \nIn our current study, we investigated the role of selective attention on the learning and exploitation of \nspatial predictive context: can we learn from spatial context that is outside attentional focus? We \nused a contextual cueing paradigm very similar to the seminal studies by Jiang & Chun (2001) and \nJiang & Leung (2005). The crucial difference is that after initial context learning, we applied a target \ncolor change in the transfer phase, while distractor context stayed exactly the same. Before this \ntarget color change we expected only spatial predictive context that is attended to be exploited. If, \nhowever, ignored spatial predictive context is latently learned, we expected to see the expression of \nthis learning (or ‘exploitation’) to be uncovered directly after transfer when it becomes attended. \nTo foreshadow our results, we found the expected exploitation of predictive visual context that was \nattended. Surprisingly, we also found an effect of predictive visual context that could be ignored, \nalready before transfer. Learning was therefore not 'latent', and its expression not dependent on \nattention. We reasoned that this exploitation of ignored spatial context may be due to a dynamic \ncompetition between selective attention and spatial predictions. Attention was regularly switching \nbetween colors due to alternating target color blocks, and this switching likely enabled spatial \npredictive context to influence behavior even from to-be-ignored locations. We confirmed this \nhypothesis in our follow-up experiment, where maintaining a consistent search goal (i.e., removing \nthe target color switching) eliminated the exploitation of to-be-ignored predictive context before \ntransfer. Interestingly, after transfer, this previously ignored context could immediately be exploited, \nsuggesting that while learning occurs independently of attention, exploitation may require attentional \nselection. This pattern reveals a fundamental tension: while attentional filtering consistently reduces \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nnoise by restricting the search space, it may simultaneously block access to potentially beneficial \npredictive information. Together, these findings demonstrate that selective attention ultimately gates \nthe influence of spatial predictions on behavior, with the strength of this gating is determined by the \nstability of attentional control. \nTransparency and Openness. All data and code used for stimulus presentation and analysis are \nfreely available on the Donders Repository at https://doi.org/10.34973/mqqe-jb24 This dataset has \nbeen subjected to a FAIR review protocol.  \nWe report how we determined our sample size, all data exclusions (if any), all manipulations, and all \nmeasures in the study. A priori power analysis using G*Power were used to calculate sample sizes. \nThese studies were not preregistered. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nExperiment 1 \nMaterial and methods  \nParticipants. A total of 109 participants were1 recruited in 2021 via the prolific platform \n(http://www.prolific.co) to participate in an online experiment. Participants that performed with an \naccuracy below 50% during any of the blocks were automatically excluded from the dataset. \nFurthermore we excluded seven participants with an overall accuracy score that fell below the 25th \npercentile minus 1.5 x interquartile range. This resulted in a final sample size of 102 participants \n(Age: 31.3±6.5 years, 35 males1) with an average performance of 94.87% (SD= 3.14) on the task. \nThis yielded a power of 85% to detect a difference with effect size d= .30 at alpha level .05, which is \napproximately the effect size of Ignored context found by Vadillo et al. (2020, experiment 3). All \nparticipants gave written informed consent beforehand and were paid for their participation. To \nmotivate participants, an overall performance level >90% correct was rewarded with an additional \nbonus pay-out. \nStimuli & Apparatus. Search scenes consisted of 17 stimuli, 1 target letter T and 16 L-shaped \ndistractors, measuring 1.2° × 1.2° in size. Distractors L shapes were rotated with a random multiple \nof 90°and had a 10% offset in the line junction to increase search difficulty (Y . Jiang & Chun, 2001). \nStimuli were placed on a regular grid spanning from −9° to +9°in 9 steps horizontal and −6° to +6° \nin 6 steps vertical from the center of the screen. Random jitter of ±0.4° was added to stimuli \nlocations to prevent collinearities with other stimuli (Chun & Jiang, 1998a). To control difficulty of the \nscenes, the target always appeared between 6° and 8° of eccentricity, and the mean distance \nbetween the target and all distractors was kept between 9° and 11°. The target was tilted either to \nthe left (-90°) or to the right (+90°). Stimuli were presented using the Gorilla platform \n(https://gorilla.sc/) and custom-written JavaScript. Participants performed the task online on their \nown device (tablet and phone excluded). Size of stimuli was controlled with a standard visual degree \ncheck implemented in Gorilla and the task could only be done in full screen mode.  \nProcedure. Participants were instructed to find a target letter T that was hidden amongst L-shaped \ndistractors, and report how this target was tilted with a key press on their keyboard; the left ‘C’ key \nfor tilted leftward and the right ‘M’ key for tilted rightward (Figure 1). Participants searched for a \nblack target for an entire block, after which they had to search for a white target for an entire block, \nand these blocks kept alternating until the end of the experiment. The target color of the first block \n \n1 Demographics of 98 of 102 participants, demographics of 4 participants were unknown \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nwas randomized across participants. Half of the distractors were the same color as the target, the \nother half of the distractors were in the other color, and all were presented on a grey background \n(see also experimental design). To remind people of the upcoming target color, each block started \nwith the target letter T shown upright in the correct color, serving as a cue (2.5 s). Throughout the \nexperiment there was a fixation dot presented at the center of the screen in same color as the \ntarget, as a continuous reminder of the current target color. Participants were asked to fixate on this \ndot between search trials, but were allowed to freely move their eyes when searching. Each trial \nstarted with a brief fixation period (1 sec), then search scenes where displayed until response or up \nto 3 s, after which the response on trial was registered as too late. After this, participants received \nfeedback on the trial for 500 ms: the fixation dot was replaced with a green ‘+’ indicating the answer \nwas correct, a red ‘x’ when the response was incorrect, and a blue ‘o’ if they were too late. The \nexperiment started with task instructions and a standard gorilla screen calibration, followed by the \nmain task consisting of 2 practice blocks (one white target color block and one black target color \nblock) and 28 experimental blocks (14 white target color blocks and 14 black target color blocks). \nAfter each block participants received feedback on their performance, indicating average response \nspeed and % correct of that block, as well as overall % correct as they would be rewarded if the latter \nwould be above 90%. They could take a short break after each block. At the end of the experiment \nthere was a short questionnaire on their experience. \n \n \nFigure 1. Task. Participants performed a visual search task where they had to locate a target letter T embedded \namongst distractor L-shapes and report its orientation with a left (‘C’) or right (‘M’) key press. A block started with \na target letter T cue in the color assigned to that block. Feedback was given at the end of each trial (correct, \nincorrect, or too late). All figures use example scenes with a black target color. \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nExperimental design. Participants searched for either a black or a white target, depending on the \nblock. On all trials half of the distractors were white, the other half were black. This resulted in one \ndistractor set that was the same color as the target, and one set which was of a different color. As it \nis typically assumed that participants can limit their (serial) search to one attended color, we label \nthe target-color-matching distractors the attended context. The other distractor set we label the \nignored context (Figure 2A). We manipulated predictability by repeating or not repeating distractor \ncontext along with the target location. If spatial context was predictive in a scene, both the location \nand orientation of the distractor set was repeated in subsequent blocks (‘old’ in contextual cueing \nparlance). The target location was also repeated if spatial context was predictive, but the target \norientation was always randomized, to prevent motor response learning. If not repeated in \nsubsequent blocks, the locations of the distractor set would be generated randomly each time, \nmaking them ‘new’ and thus nonpredictive (Figure 2B). These two factors, attentional status and \npredictiveness, were independently manipulated. Combined, they led to all scenes having attended \ncontext, that could be predictive (ATT+) or nonpredictive (ATT-), and ignored context that could be \npredictive (IGN+) or nonpredictive (IGN-). This generated four conditions (Figure 2B): both the \nattended and the ignored spatial context was predictive, labeled as all context predictive (ALL CP = \nATT+IGN+), only the attended spatial context was predictive, labeled as attended context predictive \n(ATT CP = ATT+IGN-), only the ignored spatial context was predictive, labeled as ignored context \npredictive (IGN CP = ATT-IGN+), or neither the attended, nor the ignored spatial context was \npredictive, labeled as no context predictive (NO CP = ATT-IGN-).  \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\n \nFigure 2. Experimental manipulations. A) The color of the distractor context determined the attentional status. In \nthis example the target is black, therefore black distractor context is attended and white distractor context is \nignored. Furthermore, the location and orientation of the distractors was either repeated over blocks or created \nrandomly. This determined the predictiveness of the distractor context, reflected additionally in the pictograms on \nthe right side: when connected with dashed lines to the T, the context is predictive. B) Attentional status (ATT or \nIGN) and Predictiveness (+/-)  combined generated our four conditions: all context predictive (ALL CP), attended \ncontext (ATT CP) predictive, ignored context predictive (IGN CP) and no context predictive (NO CP). \n \nCrucially, approximately half way the experiment we reversed the attentional status of the spatial \ncontext by changing the target color, a moment we call transfer (Figure 3). In doing so the spatial \ncontext remains identical but task relevance swaps: the distractor set that was previously attended \nnow became ignored, and the distractor set that was ignored became attended (Figure 3A). Because \nparticipants were always searching black and white targets in alternating blocks, this change \nremained entirely implicit. The transfer has no effect on the NO CP condition as there is no \npredictiveness in the scenes (ATT-IGN- → IGN-ATT-). Both contexts in the ALL CP condition stay \npredictive after the transfer, but the attended predictive context becomes ignored, and the ignored \npredictive context becomes attended (ATT+IGN+ → IGN+ATT+). The ATT CP condition and IGN CP \ncondition essentially trade places due to the transfer: scenes where the attended context was \npredictive (ATT CP) change into scenes where the ignored context is predictive (ATT+IGN- → \nIGN+ATT-) . The reverse is true for scenes where the ignored context was predictive (IGN CP: ATT-\nIGN+ → IGN-ATT+), they becomes scenes where the attended context is predictive (Figure 3B). \nFinally, we mimicked ‘stay’ and ‘switch’ trials from previous literature (Y. Jiang & Chun, 2001; Vadillo \net al., 2020) by adding a fifth condition: another all context predictive condition was generated, but \nwithout a transfer (ALL CP-nt: ATT+IGN+ → ATT+IGN+). \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nEach of the 28 blocks consisted of 4 scenes per condition with a target in each quadrant of the \nvisual field to prevent target probability learning (Jiang et al., 2013). This adds up to 20 trials per \ntarget color block, and in total 40 unique scenes, 8 per condition. These scenes were repeated (or \ncreated newly in the NO CP condition) for 14 times in total. The transfer was applied after 8 \nrepetitions, leaving 6 repetitions for ‘new’ learning. As commonly done, we averaged over two \nrepetitions, generating 7 epochs and these were then assigned to four phases. During the initial \nlearning phase we expect no exploitation of spatial predictive context as a minimum of two \nrepetitions is needed to learn (Tseng & Lleras, 2013). Subsequently this learning can continue, but \ncan also be visible as an behavioral advantage, which we label as the exploitation phase 1. The \nepoch directly after transfer is the crucial (post) transfer phase where we can see how the transfer \nimpacts exploitation behavior. And lastly, to assess exploitation of newly acquired spatial predictive \ncontext we have the final exploitation phase 2 (figure 3C). \n \nFigure 3. The transfer. A) Example of the transfer for two of our four conditions, and only for when the \ntarget letter T was black. Left is a scene where the attended context is predictive of target location. After \ntransfer the black distractor context is exactly the same and thus repeated, however, the target color \nchanged from black to white. Effectively, the predictive context becomes ignored. On the right an example \nwhere the predictive context was ignored, and the target color change at transfer had the opposite effect. \nB) Transfer for all four conditions depicted in pictograms. C) At the top the search goal structure over the \ntime course of the experiment is depicted: for experiment 1 this was alternating per block, for experiment 2 \nthis was consistent until transfer. Below the aggregation scheme for our different phases used for \nanalyses. Transfer occurred after 4 out of 7 epochs, indicated with the red dotted line. \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nData Analysis. Data were both analyzed and visualized using R (R Core Team, n.d.) using packages \nggplot2 (Wickham, 2016), ggrain (Allen et al., 2021), ez (Lawrence, n.d.).Reaction time was our \nprimary, and accuracy our secondary dependent variable of interest. Only trials with a response given \nwithin the time limit were included (98.18% of trials), and trials with an response given before 400 \nmsec were regarded as accidental presses and excluded (5 trials). Reaction time analysis was \nperformed on correct responses only. We first assessed whether our experimental manipulation was \nsuccessful with a 2x2 repeated measures ANOVA with time (all 7 epochs) and condition (all 5 levels) \nas factors, expecting both main effects and an interaction. We then analyzed the data in a priori \ndefined phases of the experiment: after initial learning (epoch 1, 2 repetitions), the first exploitation \nphase started (epoch 2-3-4). Between epoch 4 and 5 we applied the transfer and switched the target \ncolor in the scenes. The epoch directly after this is called the post transfer phase, and this where we \nexpect to see the effect of the transfer. Lastly, after there has been an opportunity for new learning, \nthere is the second exploitation phase (epoch 6 and 7). The exploitation phases (1 and 2) were \nanalyzed with a 2x2 repeated measures ANOVA with Attended context (predictive versus \nnonpredicitive) and Ignored context (predictive versus nonpredicitive) as factors. The impact of \ntransfer was assessed by adding transfer (pre/post) as a factor, resulting in a 2x2x2 repeated \nmeasures ANOVA. We additionally contrasted conditions both pre and post transfer using two sided t-\ntests with a Holm-Bonferroni correction for multiple comparisons (planned contrasts), which is the \nsame as the interaction between conditional difference and transfer. This allowed us to test the \neffect of transfer while removing the general task improvement over time, assuming this to be equal \nacross conditions. The ALL CP-nt (no transfer) condition was analyzed separately. First, we \nhypothesized that if changing the attentional status of spatial predictive context altered exploitation \nof the ALL CP condition, performance in the ALL CP condition should be worse compared to the ALL \nCP-nt (no transfer) condition, but only post transfer. We tested this with a one-sided t test on the \ndifference between the ALL CP conditions (transfer/no transfer) pre versus post transfer. Secondly, \nwe explored how a difference, if any, between these conditions post transfer, subsequently behaved \nover time with a 2x2 ANOVA, with factors condition (ALL CP and ALL CP-nt) and epoch (5-6-7). \nFor our ANOVA’s, if Mauchly’s test was significant, and thus the assumption of sphericity was \nviolated, we report the corrected p-value and degrees of freedom. Only when the Greenhouse– \nGeisser ε values were above 0.75 did we report the more liberal Huyn–Feldt corrected values (Field \net al., 2012). For F-tests, the generalized eta-squared measure of effect size (η2G ) is reported \n(Bakeman, 2005). For T-tests we report Cohen’s d as effect size, using the approach of Gibbons et \nal., (n.d.) for paired samples, including a suggested correction by Borenstein (2009). \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nResults \nParticipants searched for either a black or a white target, depending on the block. Half of all \ndistractors shared the color with the target (the attended distractor set), while the other half had a \ndifferent color (the ignored distractor set). Additionally, distractor context was either repeated across \nblocks (predictive) or created randomly (nonpredictive).  \nAccuracy \nOverall accuracy was near ceiling levels, indicating participants performed well on the visual search \ntask (96.62 ± 2.41). There were no significant differences between white target and black target \nblocks (t101 = .26, p= .792, d=.015).  Importantly, although the accuracy of participants improved \nover time (F3.7,373 = 20.02, pGG <.001, η2G= .048), there were no condition differences in the \naccuracy rates (F3,303 = 1.83, p= .142, η2G= .001), nor the improvement over time (F13.5,1363 = 1.33, \npGG=.184, η2G= .006).  \nReaction times \nGeneral. Average response speed was well beneath the time limit (1421 ± 537 ms). There was no \nsignificant difference in response speed between black target and white target blocks (t101= 1.663, \np= .099, d= .069). Participants improved over time, becoming faster at finding and reporting target \norientation (F4.6,468 = 105.73, pGG <.001, η2G= .138). There was a difference between our \nexperimental conditions (F2.89,292 = 7.10, pHF <.001, η2G= .005), and the improvement over time \ndiffered per condition (F16.9,1703 = 1.87, pHF =.014, η2G= .004). Our experimental manipulation thus \naffected response speed (Figure 4). \nExploitation phase 1. We examined whether participants, after the initial learning blocks, could \nexploit the predictive context, for both the attended and ignored context. As expected, participants \nsuccessfully exploited the attended spatial predictive context (F1,101= 10.32, p =.002, η2G= .006; \npredictive 1434 ± 530 ms vs nonpredictive 1461 ± 543 ms, Figure 4 & 5), confirming previous \nfindings. However, in contrast to previous reports, and to our surprise, participants could also exploit \nthe ignored spatial predictive context (F1,101= 4.07, p =.046, η2G= .002; predictive 1439 ± 532 vs \nnonpredictive 1456 ± 541 ms, Figure 4 & 5). There was no interaction between the effects of \nattended and ignored context (F1,101= 2.67, p =.105, η2G= .001).  \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\n \nFigure 4. Results experiment 1. Top: Reaction time per condition plotted over the time course of the experiment. \nShaded areas represent within subject corrected 95% confidence interval. Vertical dotted line between epoch 4 \nand 5 indicates target color change at transfer. Note how the line colors change after transfer, indicating a change \nof attentional status. Three panels below reflect the interaction analyses of reaction time per phase. From left to \nright: Exploitation phase1, post transfer, exploitation phase 2. Predictive status of ignored context is plotted on x-\naxis, predictive status of attended context is plotted in the two different colors. Note that for post-transfer and \nexploitation phase 2, the then-current (i.e., after transfer) predictive status is used to label conditions. Bars indicate \nwithin subject corrected 95% confidence interval. \n \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\n \nFigure 5. Distribution of the main effect (predictive versus \nnonpredictive) for both Attended context (left in purple) and Ignored \ncontext (right, in yellow) of experiment 1, exploitation phase 1. \nNegative values represent faster reaction times when the context is \npredictive versus when it is not. \n \nPre/Post Transfer: before and after target color change. Little over half way through the experiment, \nwe introduced the target color transfer: scenes that had a white target color before, now had a black \ntarget and scenes that had a black target before, now had a white target. The rest, i.e. the repetitions \nand the color of the distractor contexts, remained the same. This effectively changed the attended \ncontext to ignored, and the ignored spatial context to attended. When comparing the first exploitation \nphase (epoch 2-4) to search behavior directly after the transfer (epoch 5), we see that both the \nexploitation of the spatial predictive context that was attended before transfer, and the exploitation \nof spatial predictive context that was ignored before transfer, persisted, even though the target color \nchanged (Attended context: F1,101= 6.63, p=.015, η2G= .003; Ignored context: : F1,101= 5.71, p=.019, \nη2G= .003). Adding the post transfer phase additionally revealed an interaction between the effect of \nAttended context and Ignored context (F1,101= 4.22, p=.042, η2G= .002). Across phases, Ignored \npredictive context was exploited specifically when the Attended context was not predictive of target \nlocation (IGN CP condition). As expected, there is a general increase in performance over time \n(F1,101= 27.00, p< .001, η2G= .018), but none of the patterns changed from the exploitation phase to \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\npost transfer (no interaction with phase pre/post, all p-values >.05). We thus find no evidence of a \nsudden change in exploitative behavior, neither in the negative (loss of advantage) nor positive \ndirection (gain of advantage). Our planned contrasts support these results: none of the differences \nbetween conditions before the transfer was impacted by the target color change (all p values >.05). \nImportantly, the exploitation of spatial predictive context that was ignored (but in our case, exploited \nnonetheless), did not improve significantly when it became attended due to the target color change \nat transfer. We thus find no evidence that pre transfer learning leads to a sudden exploitation when \nspatial predictive context becomes task-relevant after transfer, a.k.a. ‘latent learning’. This is \nperhaps not surprising, considering we found, in contrast to previous reports, that ignored spatial \npredictive context is already exploited before transfer. Taken together, we can conclude that spatial \ncontext that is predictive of target location is exploited, both before and after target color change at \ntransfer. Does that mean the transfer had no impact at all? We can directly compare what happens \nwhen both attended and ignored predictive context are swapped due to the transfer (ALL CP) to \nwhen there is no swap at all (ALL CP-nt; both contexts predictive but no color swap at transfer, see \nMethods for details). We tested whether the transfer negatively impacted performance, and this is \nindeed what we find (t101 = 1.70, p=0.046, d= 0.22). After transfer this difference remained for the \nrest of the experiment (time: F2,202= 25.03, p< .001, η2G= .032; transfer/no transfer: F1,101= 5.13, \np= .026, η2G= .005; no interaction: F2,202=.89, p=.413, η2G< .001 ). While the former results \ndemonstrate that spatial predictive relations are learned and exploited independently of overall \nattention, this comparison indicates that the attentional set under which the relations were learned \nis nonetheless a relevant feature. \nExploitation phase 2. After post transfer exposure to the new target color, we found that participants \nlearned to exploit the ‘new’ spatial predictive relations only for the attended context (F1,101= 14.67, \np< .001, η2G= .010). After transfer we no longer see evidence of the exploitation of spatial predictive \ncontext that is ignored (Ignored context: F1,101= 1.47, p= .228, η2G= .001; no interaction F1,101=.16, \np= .686, η2G< .001). We thus conclude that both spatial predictive context that is attended and that \nis ignored can be exploited during visual search, and this exploitative behavior was neither suddenly \nnor drastically impacted when the target color changes at transfer, making the previously ignored \nspatial context attended and vice versa. Instead, we see a slow increase in the exploitation of \npreviously ignored spatial predictive context when it becomes attended and a slow decrease of \nexploitation of previously attended spatial predictive context when it becomes ignored. \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nDiscussion  \nThe fact that we see exploitation of predictive spatial context in ignored-color distractors specifically \nin the initial exploitation phase is in contrast with previous findings (Y. Jiang & Chun, 2001; Vadillo et \nal., 2020). We wondered whether this transient nature was due to the switching between white and \nblack targets, which occurred after every block of trials. It could be that the requirement to switch \nbetween black target color and white target color blocks generated a more ‘open’ attentional filter \ncompared to when participants would have to solely focus on one target color all the time. This would \nalso explain why the target color change at transfer was less impactful than expected. Moreover, if \npeople not only improve in the task in general, but also improve in the task of ‘focusing on the \ncorrect color’, this would explain why we do not see this exploitation of ignored spatial predictive \ncontext in the latter part of the experiment. Put differently, our results may indicate a tug of war \nbetween the filtering operation of attention and the exploitation of spatial predictive context in the \ntask-irrelevant color.  \nTo test this hypothesis, we executed a follow up experiment, which was identical to the first \nexperiment, except for one critical difference: participants no longer alternated in searching black \nand white targets throughout the experiment. Instead, they searched for one target color until \ntransfer. Then, at transfer, they were given explicit instructions that the target color was changed. If \nthe alternating nature of the search during the first experiment resulted in an ‘open’ attentional filter, \nand therefore to the exploitation of spatial predictive context even when it is ignored, this effect \nshould be absent in this non-alternating version of the experiment. \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nExperiment 2 \nMaterial & Methods \nParticipants. All procedures were exactly the same as the first experiment. We recruited 108 \nparticipants in 2024 via the prolific platform (http://www.prolific.co) to participate in the online \nexperiment. To ensure people were naïve, participation in our first experiment was an exclusion \ncriterion in the recruitment for the second experiment. Six participants with an overall accuracy score \nthat fell below the 25th percentile minus the 1.5 x interquartile range, were excluded. This resulted in \na final sample size of 102 participants (Age: 30.3±7.5 years, 50 males) with an average \nperformance of 93.94% (SD= 3.52%) on the task.  \nProcedure. The procedure was identical to the first experiment except for one aspect: participants \ndid not search for white and black targets alternating in blocks. Instead, they searched for a target of \none color (black target: N=51, white target: N=51) until transfer, when the target color changed. \nParticipants received short instructions at the transfer point indicating that from now on the target \ncolor was changed. \nExperimental design. As in Experiment 1, forty unique scenes were generated, 8 per condition, but \nthis time all with the same target color. At transfer the target color changed, changing the attentional \nstatus of the distractor sets, except for the ALL CP-nt condition. As we were now forced to change the \ntarget color also in this condition, we additionally changed the color of the distractor context. For \nthese scenes, since both target and distractor context change color, predictive context that is \nattended stays attended and the predictive context that is ignored, stays ignored. \nData Analysis. We excluded trials that were too late (2.22 %) and trials with a response time under \n400 ms (6 trials). After cleaning the data, we used one and the same pipeline for data analyses of \nboth the first and second experiment.  \n \nResults \nAccuracy \nOverall accuracy on the target orientation task was again near ceiling (96.06 ± 2.72%). There was no \ndifference between participants who started with a white color target and those who started with a \nblack colored target (t98.347 = 1.12, p= 0.264, d= .22). Mirroring the results of experiment 1, \nparticipants improved over time (F3.7,373 = 20.02, pGG <.001, η2G= .048), but conditions did not differ \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nin accuracy (F3,303 = 0.33, p= .801, η2G< .001), nor did the improvement over time depend on \ncondition (F16.5,1669 = .95, pHF=.517, η2G= .004).  \nReaction times \nGeneral.  Average response speed was highly similar to experiment 1 (1436 ± 545 ms). There was \nno difference between participants that started with a black target compared to those that started \nwith a white target (t98,685= 0.36, p= 0.717, d= .07). Participants improved over time, becoming \nfaster at finding and reporting target orientation (F4.5,452 = 96.99, pGG <.001, η2G= .115). There was a \ndifference between our experimental conditions (F3,303 = 3.90, p= .009, η2G= .003), and the \nimprovement over time also depended on condition (F17.5,1765 = 1.86, pHF= .016, η2G= .004). Again, \nour experimental manipulation had an effect on response speed of the task (Figure 6, top). \nExploitation phase 1. After initial learning, participants exploit spatial predictive context that is \nattended (F1,101 = 6.96, p= .010, η2G= .003, predictive 1436 ± 543 ms vs nonpredictive 1456 ± 543 \nms, Figure 6 & 7), similar to what we found in experiment 1. However, this time we find no evidence \nof exploitation of spatial predictive context that was being ignored (F1,101 = .13, p= .72, η2G< .001, \npredictive 1444 ± 543 ms vs nonpredictive 1447 ± 543 ms, Figure 6 & 7). We thus find, as \nhypothesized, that when target color search is stable, allowing for a more efficient filtering of color, \npeople no longer use spatial predictive context in the ignored color. The question that remains is \nwhether this context is also not latently learned. \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\n \nFigure 6. Results experiment 2. Top: Reaction time per condition plotted over the time course of the experiment. \nShaded areas represent within subject corrected 95% confidence interval. Vertical dotted line between epoch 4 \nand 5 indicates target color change at transfer. Note how the line colors change after transfer, indicating a change \nof attentional status. Three panels below reflect the interaction analyses of reaction time per phase. From left to \nright: Exploitation phase 1, post transfer, exploitation phase 2. Predictive status of ignored context is plotted on x-\naxis, predictive status of attended context is plotted in the two different colors. Bars indicate within subject \ncorrected 95% confidence interval. \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\n \nFigure 7. Distribution of the main effect for both Attended context (left \nin purple) and Ignored context (right, in yellow) of experiment 2, \nexploitation phase 1. Negative values represent faster reaction times \nwhen the context is predictive versus when it is not. \n \nPre/Post Transfer: before and after target color change. If only spatial predictive context that is \nattended can be learned and exploited, we should see the behavioral advantage of spatial predictive \ncontext that is attended disappear from pre to post transfer phase, as it can no longer be exploited \nnow that it is ignored. If, however, spatial predictive context that was ignored before transfer was \nlatently learned but requires attention to be exploited, we should see a behavioral advantage after \ntransfer of previously ignored, but now attended spatial predictive context. This is, in fact, what we \nfind. Besides a general improvement over time (F1,101 = 4.55, p= .03, η2G= .003) we find an effect of \nexploitation of spatial predictive context that was attended, both pre and post transfer (F1,101 = 8.12, \np= .005, η2G= .003). We again, find no evidence of exploitation of spatial predictive context that was \nignored (main effect Ignored context: F1,101 = .41, p= .521, η2G< .001; interaction Attended x Ignored \ncontext: F1,101 = .83, p= .364, η2G< .001 ), and none of the patterns changed from the exploitation \nphase to post transfer (no interaction with phase pre/post, all p-values >.05). These results indicate \nthat immediately after the target color change at transfer, spatial predictive context that is attended \nis exploited. Importantly, this spatial predictive context was still ignored pre-transfer. This would \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nimply this spatial predictive context was, in fact, latently learned, but not exploited, during the initial \nexploitation phase. Instead this learning is uncovered when the spatial predictive context became \nattended post-transfer, leading to exploitation. Similar to our first experiment, the target color change \nat transfer had little impact on the exploitation of either predictive spatial context that was attended \nand now ignored or vice versa: all planned contrasts on conditional differences pre versus post \ntransfer were non-significant. Contrary to our findings in our first experiment, we also find no \n(negative) impact of transfer when directly comparing the all context predictive conditions (transfer < \nno transfer, t101= 1.55, p= .062, d= .20). Participants still improved behaviorally in both ALL CP \nconditions during the latter part of the experiment (F2,202 = 19.59, p< .001, η2G< .023), but the lack \nof a difference between transfer and no transfer persisted (F1,101 = 1.51, p= .222, η2G= .002) and \ndid not depend on time (F2,202 = .36, p= .695, η2G< .001). \nExploitation phase 2. After the opportunity to learn after transfer, we see the that spatial predictive \ncontext is exploited, but only when this spatial predictive context is attended (F1,101 = 13.01, p< .001, \nη2G= .009) and not when it was ignored (F1,101 = 2.64, p= .107, η2G= .002), and there was also no \ninteraction (F1,101 = .74, p= .392, η2G< .001). \n \nDiscussion \nThese results show that when the task allows for a stable attentional filter, spatial predictive context \nthat was ignored can no longer be exploited. We do see, however, post transfer exploitation of spatial \npredictive context that was ignored pre transfer, when it becomes attended due to the target color \nchange. This is evidence of latent learning of ignored spatial predictive context and in line with the \nfindings of Jiang & Leung (2005). Our conclusion is that while learning spatial predictive seems to be \npossible independently of selective attention, the expression of that learning, a.k.a exploitation, of \nspatial predictive context is dependent on selective attention.  \nOur first experiment demonstrates that flexible attention enables both learning and exploitation of \nspatial predictive context, regardless of whether that context falls within or outside the attentional \nfocus. Our second experiment reveals that while learning remains intact, the exploitation of these \nspatial regularities critically depends on an 'open' attentional filter: eliminating task switching (as \npresent in Experiment 1) abolishes the exploitation of predictive spatial context in the ignored color, \nwith learned regularities only benefiting performance once they become attended. This pattern \nreveals a dynamic competition between selective attention and spatial predictions: when attention is \nflexible due to task switching, spatial predictions can influence behavior even from ignored locations, \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nbut when attentional selection is stable and efficient, it gates the exploitation of predictive \ninformation. Together, these experiments illuminate how selective attention and spatial predictions \nengage in a complex interplay during visual search, with their relative influence determined by the \nstability of attentional control. \n \nGeneral Discussion \n \nHere, we investigated whether learning and exploitation of spatial predictive context in visual search \ncan occur without attention. Levering a contextual cueing paradigm, we generated spatial \npredictions during visual search: repeated distractor context becomes predictive of target location \nand this enhances attentional guidance through these scenes (Chun, 2000; Chun & Jiang, 1998b; \nGoujon et al., 2015; Sisk et al., 2019). Additionally we altered the attentional status of parts of the \nscene by manipulating their color. Distractors of the same color as the to-be-searched target need to \nbe inspected, and are therefore attended. Conversely, distractors in the other color could be ignored. \nCrucially, approximately halfway through the experiments, we changed the color of the target \n(transfer) within the scenes, changing the attentional status of the distractor contexts. Combining \nspatial predictiveness and attentional status enabled us to investigate how goal-directed selective \nattention and predictions interact. As expected, we found robust learning and exploitation of spatial \npredictive context that was task-relevant and therefore attended. Surprisingly, and counter to earlier \nwork (Jiang & Chun, 2001; Jiang & Leung, 2005), we also observed both learning and exploitation of \nignored spatial predictive context. The exploitation of ignored spatial context however only occurred \nwhen participants regularly switched in terms of the color of the searched-for target, potentially \nleading to a broader attentional filter. When participants’ attention was more stable (Experiment 2), \nparticipants still learnt but no longer exploited the ignored spatial context. This latent learning \nbecame visible when the spatial context became attended after transfer. We discuss these findings \nbelow. \nAs expected, spatial predictive context that was task-relevant and thus attended was learned and \nexploited: in both experiments participants quickly learned and were able to exploit attended spatial \ncontext that was predictive of target location to improve visual search performance. This adds to a \nlarge body of evidence that indicates that when spatial predictive context is attended, it can be \nlearned and exploited (Chun, 2000; Chun & Turk-Browne, 2007; Goujon et al., 2015; Jiang & Chun, \n2001; Sisk et al., 2019). However, in our first experiment, we additionally found learning and \nexploitation of spatial predictive context that was task-irrelevant and thus could be ignored. \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nImportantly, we found this effect already before the change of attentional status at transfer. Learning \nfrom ignored spatial predictive context was therefore not latent but manifest: exploitation was not \ndependent on selective attention. Already in the classic study by Chun & Jiang (1998) evidence of \nthe exploitation of ignored predictive context was found. However, when they increased the difficulty \nin the task by making the distractors more similar to the target, this effect disappeared. This finding \nwas used to argue that when the task was too easy, attentional resources automatically spilled over \nto the irrelevant context. This was confirmed in a later study (Jiang & Leung, 2005). However, in this \nstudy they found that when the attentional status of ignored spatial predictive context changed, this \ncontext suddenly could be exploited. This was taken as evidence that the spatial relations in the \nignored context were latently learned. In an extensive and high-powered study by Vadillo et al. \n(2020), this latent learning was not replicated. Vadillo et al. did, however, find evidence of the \nexploitation of ignored predictive context in one of their experiments, and argued this is due to \nparticipants inability to truly ignore the context in the task-irrelevant color. This dovetails with our \nfindings. After observing the results from our first experiment, we hypothesized that the switching \nbetween target colors might have demanded more flexible attention, and as a consequence, created \na more open attentional filter. This would lead to less efficient suppression of task irrelevant context, \ngiving rise to an opportunity to learn and exploit this context when it was predictive of target location. \nTo investigate the influence of the flexibility of attention on the exploitation of spatial predictive \ncontext, we removed the task switching and instead gave participants a stable goal in a follow-up \nexperiment. As expected, learning and exploitation from the attended spatial predictive context was \nunaffected while we no longer found evidence of exploitation of ignored spatial predictive context. \nWe did however, find evidence for exploitation of ignored spatial predictive context when a transfer in \ntask relevance changed its attentional status to attended. This suggests that this spatial predictive \ncontext was latently learned when it was still ignored, but required attention to be exploited. Taken \ntogether, these experiments demonstrate two things. Firstly, selective attention is crucial for \nexploitation of spatial predictive context in visual search. People consistently learn and exploit \nspatial predictive context that is task relevant and thus attended. Moreover, if the irrelevant spatial \ncontext is efficiently filtered out, people can no longer exploit this context, even though it is predictive \nof target location. Secondly, learning spatial regularities is not dependent on selective attention. This \nis consistent with the early findings by Jiang & Leung (2005), but not with more recent work (Vadillo \net al., 2020). One important distinction was that the change of attentional status of the spatial \ncontext at transfer was instantiated via a target color change, in both our experiments. This was a \ndeliberate deviation from previous work investigating the role of selective attention and contextual \nlearning (Jiang & Leung, 2005; Vadillo et al., 2020, 2024). What is typically done, is changing the \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\ncolor of all distractor stimuli at transfer. While this approach maintains consistent search goals and \nminimizes explicit awareness of the change, it may introduce unintended consequences. Specifically, \nthe spatial context-target association is consistently paired with color and thus likely encoded with \ncolor information (Turk-Browne et al., 2008).  Altering the distractor colors could disrupt the learned \nassociations more severely than a simple target color change. Our results support this notion. \nInstead of losing all behavioral advantage, as is commonly seen, the change at transfer was less \nimpactful in both our experiments. For our first experiment, it can be argued that post transfer \nexploitation can be attributed to the open attentional filter that explains our results before transfer. \nHowever, this is less likely in our second experiment with stable and efficient selective attentional \nfiltering. We believe our target color change allowed for more exploitation of the learned spatial \ncontext-target location association compared to changing the color of all distractors and this \nrevealed latent learning. This is therefore an important consideration for future work on this topic. \nThe concept of an attentional filter stems from the well-established finding that selective attention \nenables people to restrict their visual search to a specific color-defined subset of elements. (Egeth et \nal., 1984; Kaptein et al., 1995; Palmer, 1994; Wolfe, 2021). Recently Duncan et al. (2024) \ndemonstrated that limiting the search space via an explicit color cue abolished the learning of the \nspatial distribution of distractors in the to-be-ignored color. In their design, however, color-based \nignoring prevented further processing of the shapes and with it access to predictive information, as it \nwas a certain shape that occurred more often at one location. In our visual search scenes, \nsuppression of distractors in one color automatically encompasses brief processing of the spatial \ncontext in that color, and it is this spatial context that can be predictive of target location. Another \nexample of guided visual search is a recent study by Duecker et al. (2024), investigating the \ndifference between guided (i.e., color-cued) and unguided search of scenes consisting of stimuli in \ntwo colors. They found that guided search behavioral performance was indistinguishable from \nunguided search of half the set size. This reduction of the search space was accompanied by neural \nenhancement of the task relevant context and neural suppression of the task irrelevant context. We \nhad a very similar set-up (amount of trials, stimuli and blocked target color cue), yet do not see such \nefficient suppression of the to be ignored spatial context. Again, an important distinction is that in \nour experiments the ignored spatial context was a potential cue that could be learned and exploited. \nInterestingly, a recent study by Vadillo et al. (2024) found that the set size of the irrelevant context \nimpacts visual search performance, demonstrating that this context was not perfectly ignored. \nHowever, this set size effect did not depend on attentional resources and did not interact with the \ncontextual cueing effect: more attention to the irrelevant spatial context did not lead to more \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nlearning nor exploitation when it was predictive of target location. More research is necessary to fully \nunderstand when and how selective attention enables learning from spatial regularities. \nThe exploitation of spatial predictive context in ignored locations appears to follow a dynamic pattern \nacross our experiments. In Experiment 1, this exploitation was primarily observed before transfer, \nwith later phases showing exploitation only for attended contexts. This pattern suggests that \nattentional filtering mechanisms strengthen over time, eventually preventing the utilization of \npredictive information outside the attentional focus, despite its initial benefit to performance. This \neffect was even more pronounced in Experiment 2, where optimized filtering completely eliminated \nthe exploitation of ignored predictive context until it became attended. These findings reveal a \nfundamental tension between spatial predictions and selective attention: while perfect filtering of \ntask-irrelevant context optimizes processing efficiency, it simultaneously blocks access to potentially \nbeneficial predictive information. This creates an interesting paradox where context that is predictive \nof target location could be considered 'task-relevant' by definition, yet attending to it requires the \nprocessing of more distractors than strictly necessary based on target color alone. While previous \nresearch has suggested that selective attention and predictions jointly modulate priority maps in the \nbrain (Fecteau & Munoz, 2006; Ferrante et al., 2018; Sisk et al., 2019; Wolfe, 2021), our results \ndemonstrate that these mechanisms may actually compete dynamically, with selective attention \nultimately constraining prediction's influence on behavior. This competition can be understood \nthrough a signal-to-noise framework: attentional filtering consistently reduces noise by restricting the \nsearch space, whereas the benefits of processing ignored distractors only materialize when they \ncontain predictive information. This asymmetry in reliability may explain why attention ultimately \ngates the influence of spatial predictions, advancing our understanding of how these fundamental \ncognitive mechanisms interact under competition. \nConstraints on Generality. This study made use of large sample sizes and online recruitment enabled \na diverse sample. This likely implies a greater generalization of our findings to the general \npopulation, compared to typical laboratory experiments conducted on small samples of \nundergraduate students. Our use of simplified visual search task might limit applicability to real \nworld experience. While naturally occurring searches can be complex, these simplified paradigms \nnonetheless capture key aspects of real-world visual search (Botch et al., 2023). \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nReferences \n \nAllen, M., Poggiali, D., Whitaker, K., Marshall, T. R., Van Langen, J., & Kievit, R. A. (2021). Raincloud \nplots: A multi-platform tool for robust data visualization. Wellcome Open Research, 4, 63. \nhttps://doi.org/10.12688/wellcomeopenres.15191.2 \nBakeman, R. (2005). Recommended effect size statistics for repeated measures designs. Behavior \nResearch Methods, 37(3), 379–384. https://doi.org/10.3758/BF03192707 \nBatterink, L. J., & Paller, K. A. (2019). Statistical learning of speech regularities can occur outside the \nfocus of attention. Cortex, 115, 56–71. https://doi.org/10.1016/j.cortex.2019.01.013 \nBeesley, T., Vadillo, M. A., Pearson, D., & Shanks, D. R. (2016). Configural learning in contextual \ncuing of visual search. Journal of Experimental Psychology: Human Perception and \nPerformance, 42(8), 1173–1185. https://doi.org/10.1037/xhp0000185 \nBorenstein, M. (Ed.). (2011). Introduction to meta-analysis. John Wiley & Sons. \nBotch, T. L., Garcia, B. D., Choi, Y . B., Feffer, N., & Robertson, C. E. (2023). Active visual search in \nnaturalistic environments reflects individual differences in classic visual search \nperformance. Scientific Reports, 13(1), 631. https://doi.org/10.1038/s41598-023-27896-7 \nBouwkamp, F . G., De Lange, F . P ., & Spaak, E. (2025). Spatial Predictive Context Speeds Up Visual \nSearch by Biasing Local Attentional Competition. Journal of Cognitive Neuroscience, 37(1), \n28–42. https://doi.org/10.1162/jocn_a_02254 \nChun, M. M. (2000). Contextual cueing of visual attention. Trends in Cognitive Sciences, 4(5), 170–\n178. https://doi.org/10.1016/S1364-6613(00)01476-5 \nChun, M. M., & Jiang, Y . (1998a). Contextual Cueing: Implicit Learning and Memory of Visual \nContext Guides Spatial Attention. Cognitive Psychology, 36(1), 28–71. \nhttps://doi.org/10.1006/cogp.1998.0681 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nChun, M. M., & Jiang, Y . (1998b). Contextual Cueing: Implicit Learning and Memory of Visual \nContext Guides Spatial Attention. Cognitive Psychology, 36(1), 28–71. \nhttps://doi.org/10.1006/cogp.1998.0681 \nChun, M. M., & Turk-Browne, N. B. (2007). Interactions between attention and memory. Current \nOpinion in Neurobiology, 17(2), 177–184. https://doi.org/10.1016/j.conb.2007.03.005 \nDuecker, K., Shapiro, K. L., Hanslmayr, S., Griffiths, B. J., Pan, Y ., Wolfe, J., & Jensen, O. (2024). \nGuided Visual Search is associated with a feature-based priority map in early visual cortex. \nhttps://doi.org/10.1101/2024.10.03.616568 \nDuncan, D. H., Van Moorselaar, D., & Theeuwes, J. (2023). Pinging the brain to reveal the hidden \nattentional priority map using encephalography. Nature Communications, 14(1), 4749. \nhttps://doi.org/10.1038/s41467-023-40405-8 \nDuncan, D. H., Van Moorselaar, D., & Theeuwes, J. (2024). Visual statistical learning requires \nattention. Psychonomic Bulletin & Review. https://doi.org/10.3758/s13423-024-02605-1 \nEgeth, H. E., Virzi, R. A., & Garbart, H. (1984). Searching for conjunctively defined targets. Journal of \nExperimental Psychology: Human Perception and Performance, 10(1), 32–39. \nhttps://doi.org/10.1037/0096-1523.10.1.32 \nFecteau, J., & Munoz, D. (2006). Salience, relevance, and firing: A priority map for target selection. \nTrends in Cognitive Sciences, 10(8), 382–390. https://doi.org/10.1016/j.tics.2006.06.011 \nFerrante, O., Patacca, A., Di Caro, V ., Della Libera, C., Santandrea, E., & Chelazzi, L. (2018). Altering \nspatial priority maps via statistical learning of target selection and distractor filtering. \nCortex, 102, 67–95. https://doi.org/10.1016/j.cortex.2017.09.027 \nFerrante, O., Zhigalov, A., Hickey, C., & Jensen, O. (2023). Statistical Learning of Distractor \nSuppression Downregulates Prestimulus Neural Excitability in Early Visual Cortex. The \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nJournal of Neuroscience, 43(12), 2190–2198. https://doi.org/10.1523/JNEUROSCI.1703-\n22.2022 \nField, A., Miles, J., & Field, Z. (2012). Discovering statistics using R. Sage Publications Ltd.,. \nhttp://repository.fue.edu.eg/xmlui/handle/123456789/2902 \nGeng, J. J., & Behrmann, M. (2005). Spatial probability as an attentional cue in visual search. \nPerception & Psychophysics, 67(7), 1252–1268. https://doi.org/10.3758/BF03193557 \nGeyer, T., Zinchenko, A., Seitz, W., Balik, M., Müller, H. J., & Conci, M. (2024). Mission impossible? \nSpatial context relearning following a target relocation event depends on cue \npredictiveness. Psychonomic Bulletin & Review, 31(1), 148–155. \nhttps://doi.org/10.3758/s13423-023-02328-9 \nGibbons, R. D., Hedeker, D. R., & Davis, J. M. (1993). Estimation of Effect Size From a Series of \nExperiments Involving Paired Comparisons. 18(3), 271–279. \nGoujon, A., Didierjean, A., & Thorpe, S. (2015). Investigating implicit statistical learning \nmechanisms through contextual cueing. Trends in Cognitive Sciences, 19(9), 524–533. \nhttps://doi.org/10.1016/j.tics.2015.07.009 \nJiang, Y ., & Chun, M. M. (2001). Selective attention modulates implicit learning. The Quarterly \nJournal of Experimental Psychology Section A, 54(4), 1105–1124. \nhttps://doi.org/10.1080/713756001 \nJiang, Y ., & Leung, A. W. (2005). Implicit learning of ignored visual context. Psychonomic Bulletin & \nReview, 1(12), 7. \nJiang, Y ., & Song, J.-H. (2005). Hyperspecificity in Visual Implicit Learning: Learning of Spatial Layout \nIs Contingent on Item Identity. Journal of Experimental Psychology: Human Perception and \nPerformance, 31(6), 1439–1448. https://doi.org/10.1037/0096-1523.31.6.1439 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nJiang, Y . V . (2018). Habitual versus goal-driven attention. Cortex, 102, 107–120. \nhttps://doi.org/10.1016/j.cortex.2017.06.018 \nJiang, Y . V ., Sisk, C. A., & Toh, Y . N. (2019). Implicit guidance of attention in contextual cueing: \nNeuropsychological and developmental evidence. Neuroscience & Biobehavioral Reviews, \n105, 115–125. https://doi.org/10.1016/j.neubiorev.2019.07.002 \nJiang, Y . V ., Swallow, K. M., Rosenbaum, G. M., & Herzig, C. (2013). Rapid acquisition but slow \nextinction of an attentional bias in space. Journal of Experimental Psychology: Human \nPerception and Performance, 39(1), 87–99. https://doi.org/10.1037/a0027611 \nKaptein, N. A., Theeuwes, J., & Van Der Heijden, A. H. C. (1995). Search for a conjunctively defined \ntarget can be selectively limited to a color-defined subset of elements. Journal of \nExperimental Psychology: Human Perception and Performance, 21(5), 1053–1069. \nhttps://doi.org/10.1037/0096-1523.21.5.1053 \nLawrence, M. (n.d.). Lawrence, MA. (2011) ez: Easy analysis and visualization of factorial \nexperiments. R package version 3.0-0. [Computer software]. http://CRAN.R-\nproject.org/package=ez \nMeyen, S., Vadillo, M. A., Von Luxburg, U., & Franz, V . H. (2024). No evidence for contextual cueing \nbeyond explicit recognition. Psychonomic Bulletin & Review, 31(3), 907–930. \nhttps://doi.org/10.3758/s13423-023-02358-3 \nMusz, E., Weber, M. J., & Thompson-Schill, S. L. (2015). Visual statistical learning is not reliably \nmodulated by selective attention to isolated events. Attention, Perception, & \nPsychophysics, 77(1), 78–96. https://doi.org/10.3758/s13414-014-0757-5 \nPalmer, J. (1994). Set-size effects in visual search: The effect of attention is independent of the \nstimulus for simple tasks. Vision Research, 34(13), 1703–1721. \nhttps://doi.org/10.1016/0042-6989(94)90128-7 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nPeelen, M., Berlot, E., & De Lange, F . (2023). Predictive processing of scenes and objects [Preprint]. \nPsyArXiv. https://doi.org/10.31234/osf.io/nuf59 \nPollmann, S. (2019). Working memory dependence of spatial contextual cueing for visual search. \nBritish Journal of Psychology, 110(2), 372–380. https://doi.org/10.1111/bjop.12311 \nR Core Team. (n.d.). R: A language and environment for statistical computing [Computer software]. \nR Foundation for Statistical Computing. http://www.R-project.org/ \nRichter, D., Van Moorselaar, D., & Theeuwes, J. (2024). Proactive distractor suppression in early \nvisual cortex. https://doi.org/10.7554/eLife.101733.1 \nSaffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical Learning by 8-Month-Old Infants. \nScience, 274(5294), 1926–1928. https://doi.org/10.1126/science.274.5294.1926 \nSisk, C. A., Remington, R. W., & Jiang, Y . V . (2019). Mechanisms of contextual cueing: A tutorial \nreview. Attention, Perception, & Psychophysics, 81(8), 2571–2589. \nhttps://doi.org/10.3758/s13414-019-01832-2 \nSpaak, E., & de Lange, F . P . (2020). Hippocampal and prefrontal theta-band mechanisms underpin \nimplicit spatial context learning. The Journal of Neuroscience, 1660–19. \nhttps://doi.org/10.1523/JNEUROSCI.1660-19.2019 \nSpaak, E., Peelen, M. V ., & De Lange, F . P . (2022). Scene Context Impairs Perception of Semantically \nCongruent Objects. Psychological Science, 33(2), 299–313. \nhttps://doi.org/10.1177/09567976211032676 \nTheeuwes, J., Bogaerts, L., & Van Moorselaar, D. (2022). What to expect where and when: How \nstatistical learning drives visual selection. Trends in Cognitive Sciences, 26(10), 860–872. \nhttps://doi.org/10.1016/j.tics.2022.06.001 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nTseng, Y .-C., & Lleras, A. (2013). Rewarding context accelerates implicit guidance in visual search. \nAttention, Perception, & Psychophysics, 75(2), 287–298. https://doi.org/10.3758/s13414-\n012-0400-2 \nTurk-Browne, N. B., Isola, P . J., Scholl, B. J., & Treat, T. A. (2008). Multidimensional visual statistical \nlearning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 34(2), 399–\n407. https://doi.org/10.1037/0278-7393.34.2.399 \nTurk-Browne, N. B., Jungé, J. A., & Scholl, B. J. (2005). The Automaticity of Visual Statistical \nLearning. Journal of Experimental Psychology: General, 134(4), 552–564. \nhttps://doi.org/10.1037/0096-3445.134.4.552 \nVadillo, M. A., Aniento, P ., Hernández-Gutiérrez, D., Saini, L., & Aivar, M. P . (2024). Measuring \nlearning and attention to irrelevant distractors in contextual cueing. Journal of Experimental \nPsychology: Human Perception and Performance, 50(9), 952–970. \nhttps://doi.org/10.1037/xhp0001230 \nVadillo, M. A., Giménez-Fernández, T., Aivar, M. P ., & Cubillas, C. P . (2020). Ignored visual context \ndoes not induce latent learning. Psychonomic Bulletin & Review, 27(3), 512–519. \nhttps://doi.org/10.3758/s13423-020-01722-x \nVadillo, M. A., Konstantinidis, E., & Shanks, D. R. (2016). Underpowered samples, false negatives, \nand unconscious learning. Psychonomic Bulletin & Review, 23(1), 87–102. \nhttps://doi.org/10.3758/s13423-015-0892-6 \nVicente-Conesa, F ., Giménez-Fernández, T., Shanks, D. R., & Vadillo, M. A. (2022). The role of \nworking memory in contextual cueing of visual attention. Cortex, 154, 287–298. \nhttps://doi.org/10.1016/j.cortex.2022.05.019 \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint \n\nVickery, T. J., Sussman, R. S., & Jiang, Y . V . (2010). Spatial context learning survives interference from \nworking memory load. Journal of Experimental Psychology: Human Perception and \nPerformance, 36(6), 1358–1371. https://doi.org/10.1037/a0020558 \nVõ, M. L.-H., Boettcher, S. E., & Draschkow, D. (2019). Reading scenes: How scene grammar guides \nattention and aids perception in real-world environments. Current Opinion in Psychology, \n29, 205–210. https://doi.org/10.1016/j.copsyc.2019.03.009 \nWickham, H. (2016). ggplot2: Elegant Graphics for Data Analysis (2nd ed. 2016). Springer \nInternational Publishing : Imprint: Springer. https://doi.org/10.1007/978-3-319-24277-4 \nWolfe, J. M. (2021). Guided Search 6.0: An updated model of visual search. Psychonomic Bulletin & \nReview, 28(4), 1060–1092. https://doi.org/10.3758/s13423-020-01859-9 \n \n \n.CC-BY 4.0 International licenseavailable under a \nwas not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made \nThe copyright holder for this preprint (whichthis version posted July 15, 2025. ; https://doi.org/10.1101/2025.07.10.664103doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}