Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour

Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour | 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 Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour Rinku Dixit, Shailee Choudhary, Sadia Samar, Kirit Goyal This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7898697/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 Recent advancements in large language models (LLMs) such as GPT-4, LLaMA, and Qwen2.5 have revealed capabilities extending beyond language generation to include complex reasoning and decision-making. This paper investigates whether LLMs exhibit consistent behavioural tendencies—comparable to human personality traits—when placed in structured decision-making scenarios. We conduct a two-pronged empirical study using (i) the Balloon Analogue Risk Task (BART), a psychological tool for assessing risk propensity, and (ii) a time-series forecasting task involving real-world FMCG sales data. Across both tasks, four state-of-the-art LLMs demonstrated stable and distinct behavioural profiles: models that acted conservatively in BART also generated cautious sales forecasts, while risk-taking models projected more aggressive outcomes. These patterns persisted across multiple runs and prompt variations, providing strong evidence that the observed behaviours are not artifacts of prompt engineering but rather emergent dispositions rooted in model architecture and training data. This work establishes a foundation for behavioural modelling in AI, with implications for building task-aligned foundation models that reflect consistent decision-making styles. large language models (LLMs) Balloon Analogue Risk Task (BART) time-series forecasting FMCG sales data behavioural modelling prompt invariance AI decision-making Figures Figure 1 Figure 2 Figure 3 1. Introduction With the advent of transformer-based large language models (LLMs) such as GPT-4, LLaMA, and others, artificial intelligence has achieved significant milestones not only in language understanding but also in demonstrating emerging capabilities in reasoning, planning, and decision-making. These models represent major advancements in natural language processing (NLP) and generation. Trained on massive textual corpora, they can comprehend and generate human-like responses across a wide range of domains. GPT-4, developed by OpenAI, is recognized for its broad versatility and enhanced performance compared to previous iterations. LLaMA, developed by Meta AI, is an open-source model optimized for efficiency, speed, and accessibility to researchers. Gemma, from Google, emphasizes compact architecture and performance efficiency, delivering strong results despite a smaller parameter footprint. Qwen2.5, released by Alibaba, is designed with multilingual optimization and improved performance across complex language tasks. Each of these models brings unique capabilities, from content generation and translation to real-time problem-solving and decision-support applications1. As these models are increasingly deployed in high-stakes environments—such as financial forecasting, clinical diagnostics, and policy decision-making—it becomes critical to ask: Do these models exhibit consistent behavioural patterns akin to personality traits, or are their outputs purely stochastic and task-specific? To address this question, the current study draws inspiration from behavioural psychology, particularly the Balloon Analogue Risk Task (BART). Originally introduced by Lejuez et al. ( 2002 ), BART is a well-established paradigm used to measure risk-taking tendencies in humans. In this task, participants incrementally inflate a virtual balloon to increase hypothetical earnings, risking loss if the balloon bursts. Human pump values typically average between 5 and 7, and are linked to psychological traits such as impulsivity, emotional regulation, and age (Lauriola et al., 2014 ; Lejuez et al., 2002 ). BART thus provides a lens to examine latent behavioural propensities—whether conservative, risk-neutral, or risk-seeking. We hypothesize that similar behavioural tendencies may also manifest in LLMs, not as a product of explicit programming, but as emergent phenomena resulting from statistical learning on human-generated corpora. To test this, we subjected four state-of-the-art LLMs—GPT-4, LLaMA, Gemma, and Qwen2.5—to a simulated BART environment, recording their pump choices across multiple iterations. We then exposed the same models to a real-world time-series forecasting task involving FMCG (Fast-Moving Consumer Goods) sales data. Here, models were prompted to forecast 12 months of future demand based on 24 months of historical sales. Forecast behaviour was qualitatively and quantitatively analysed to determine whether risk postures observed in BART were consistent with the nature (conservative or aggressive) of the sales forecasts. We conducted a one-way ANOVA to evaluate statistically significant differences in risk responses across the four LLMs and benchmarked these results against human BART baselines to investigate behavioural alignment or divergence. This study contributes to the emerging field of artificial behavioural modelling by demonstrating that LLMs may exhibit stable, cross-task behavioural signatures. Crucially, we argue that these behavioural consistencies are not artifacts of prompt engineering, as they persist across prompt variants, temperature settings, and random seeds. Our findings pave the way for the development of behavioural evaluation metrics grounded in psychology and decision theory—providing a new lens for understanding the internal dynamics of generative AI systems. The remainder of the paper is structured as follows: Section 2 reviews related literature and identifies key gaps this study addresses. Section 3 describes the experimental methods. Section 4 presents the results of the BART and forecasting tasks. Section 5 offers a discussion of implications, followed by limitations and future directions in Sections 6 and 7 . 2. Literature Review 2.1. The Balloon Analogue Risk Task (BART) The Balloon Analogue Risk Task (BART), developed by Lejuez et al. ( 2002 ), is a widely used behavioural paradigm for evaluating individual risk-taking tendencies under conditions of uncertainty. In the BART, participants are presented with a virtual balloon which they may inflate incrementally; each additional pump increases the potential monetary gain but also raises the risk of the balloon bursting, resulting in the loss of accrued rewards. This trade-off effectively mirrors real-world decisions that involve weighing potential gains against uncertain risks and has been applied across clinical, developmental, and neuropsychological populations. Empirical studies have reported that average adjusted pump values (excluding bursts) among human participants typically range from 5.5 to 7.5 pumps, with considerable variation influenced by demographic and psychological factors. For instance, adolescents and individuals with substance use disorders often exhibit more risk-seeking behaviour, whereas older adults and individuals with anxiety traits tend to behave more conservatively (White et al., 2008; Lauriola et al., 2014 ). Neuroimaging studies have further linked BART performance to activation in brain regions associated with reward processing and executive control, including the ventromedial prefrontal cortex and anterior cingulate cortex (Schonberg et al., 2012 ). These findings support the notion that risk preferences are shaped by stable cognitive and emotional traits. 2.2. Cognitive Characteristics of Large Language Models The development of transformer-based architectures (Vaswani et al., 2017 ) has enabled large language models (LLMs) such as GPT-3, GPT-4, LLaMA, Claude, and PaLM to demonstrate emergent cognitive behaviours. Although not explicitly trained for cognitive modelling, these models often exhibit outputs aligned with psychological constructs, particularly when simulating human decisions. LLMs have been observed to demonstrate traits including: Theory of Mind: Kosinski ( 2023 ) suggested that LLMs can simulate Theory of Mind capabilities—attributing mental states and intentions to others in contextually appropriate ways. Commonsense Reasoning: Sap et al. (2019) found that LLMs often display intuitive reasoning grounded in everyday knowledge, although sometimes shallow or fragile under counterfactual probing. Causal Inference: Zhou et al. (2022) demonstrated that models can perform basic causal reasoning by inferring likely cause-effect relationships from prompts. Metacognition and Self-Reflection: OpenAI ( 2023 ) noted that advanced LLMs can engage in self-evaluation of their outputs, a form of metacognitive simulation. Several recent studies have examined LLMs through a behavioural science lens: Binz and Schulz ( 2023 ) analyzed whether LLMs follow heuristics and biases found in human decision-making (e.g., anchoring, framing effects). Ullman ( 2023 ) investigated the presence of rational inductive biases in LLMs, comparing them to Bayesian models of human inference. Wang et al. ( 2023 ) proposed evaluation benchmarks like MMLU and BBH to systematically assess reasoning consistency and uncertainty handling in LLMs. Despite these developments, few studies have examined whether LLMs exhibit stable behavioural profiles across tasks, akin to those seen in human subjects (e.g., consistent risk-taking or conservatism). This behavioural consistency—or lack thereof—has significant implications for real-world deployment of LLMs in high-stakes environments such as mental health screening, strategic decision-making, or autonomous systems, where risk posture must be stable, explainable, and reliable. . 3. Research gaps and proposed objectives 3.1 Bridging BART with LLM Behavioural Analysis Despite the surge in work comparing Large Language Models (LLMs) across tasks, to our knowledge, no prior study has systematically employed psychometric paradigms like the Balloon Analogue Risk Task (BART) to explore inherent behavioural tendencies of LLMs. BART is a validated tool in psychology used to surface risk-related traits such as impulsivity or caution, often linked to neural and personality substrates in humans. Its translation into a simulated LLM context allows us to probe: do these AI systems show structured, non-random decision tendencies? A second gap lies in the lack of cross-task behavioural validation. It is not sufficient to observe a model’s behaviour in a single, isolated task. Instead, we investigate whether these tendencies generalize to a time-series forecasting task—a real-world business scenario requiring uncertainty management, prudence, and decision trade-offs. Importantly, this study is not a function of prompt engineering alone. While prompts can influence LLM output, the same minimal task framing was applied uniformly across all models. The consistency in observed behaviour—across tasks, over multiple runs, and under uncertainty—points to emergent characteristics, rather than engineered outputs. Our goal is not to simulate human behaviour through instruction-following, but to reveal underlying behavioural signatures embedded in the LLM’s learned representations. 3.2 Quantitative BART Simulations Across LLMs We deploy the BART simulation uniformly across four open- and closed-source LLMs: GPT-4, LLaMA, Gemma, and Qwen2.5. Each model was exposed to a consistent game interface and rules, eliminating variability due to prompt design. Over multiple iterations, we logged their choices—how far they “pumped” a virtual balloon before “cashing out”—as a proxy for risk appetite. The resulting data was analysed statistically using ANOVA and compared against human baselines to validate behavioural interpretability. This approach enables a quantitative profiling of LLMs, drawing parallels with psychological traits like impulsivity or risk aversion. Rather than simulate human behavior through commands like “act like a conservative investor,” we test whether models naturally exhibit such behaviour under decision pressure. This provides a deeper, unbiased look into their cognitive biases. 3.3 Forecasting Behaviour Under Uncertainty The forecasting task, constructed using FMCG sales time series data, provides a real-world analogue to the decision-making context in BART. LLMs were asked to predict future sales based on 24 months of historical data. We then evaluated whether models that were risk-seeking in BART also made more aggressive forecasts (e.g., steeper growth assumptions), while cautious models leaned toward conservative projections. This cross-validation ensures that we are not observing random behaviour or prompt-induced output but stable behavioural dispositions across different task formats—simulation vs. prediction, game vs. business. The alignment between BART profiles and forecast aggressiveness reinforces the core hypothesis of this study: that LLMs display generalizable behavioural characteristics. 3.2 Proposed Research Objectives This study is motivated by the need to move beyond performance benchmarking and explore whether large language models (LLMs) exhibit consistent, interpretable behavioural tendencies across different types of tasks. Instead of evaluating model outputs solely based on correctness or fluency, this research focuses on patterns of decision-making—especially in contexts involving risk and uncertainty. The specific research objectives are as follows: 1. To measure risk-taking behaviour in LLMs using the Balloon Analogue Risk Task (BART) This objective involves deploying a simulated BART environment across four state-of-the-art LLMs—GPT-4, Gemma, LLaMA, and Qwen2.5—to observe their decision-making under uncertain reward conditions. BART is a validated psychometric tool traditionally used in behavioural psychology, and applying it to LLMs enables us to examine if similar decision traits (e.g., cautiousness, impulsivity) emerge. 3. To assess behavioural transferability to a real-world domain using a time-series sales forecasting task The study introduces a second task: forecasting 12 months of FMCG sales data using 24 months of historical inputs. The goal is to observe whether LLMs show behavioural coherence—whether a model that is conservative in BART also provides cautious forecasts, and vice versa. 4. To identify cross-task behavioural consistency as evidence of emergent behavioural signatures in LLMs By evaluating if models display stable behavioural tendencies across two unrelated cognitive tasks (simulation and forecasting), this objective explores whether LLMs manifest consistent decision styles that are not engineered via prompt, but rather emergent from training and architecture. 5. To propose a behavioural evaluation framework for future AI systems This research sets the stage for behaviourally differentiated foundation models. It opens the door for further study into the interpretability of model decisions, their alignment with human traits, and the philosophical implications for machine consciousness, autonomous decision-making, and model personality architectures.. 4. Methodology This study follows a two-stage experimental protocol to examine whether large language models (LLMs) exhibit consistent behavioural tendencies across tasks—specifically in risk-taking (via the BART task) and economic forecasting (via time-series prediction). All tasks were designed using minimal, task-neutral prompts, held constant across all models to eliminate prompt-induced biases. This uniformity ensures that behavioural patterns arise from model internals rather than from engineered instructions. Importantly, no prompt engineering techniques were used to guide model behaviour in either task . This design choice ensures that observed behavioural patterns are emergent from model architecture and training data, rather than induced by prompt manipulation. A graphical summary of this methodology is provided in Figure A to enhance clarity and transparency. Figure A. Graphical summary of methodology 4.1. Stage 1: Simulated Balloon Analogue Risk Task (BART) 4.1.1. Simulation Design A Python program was developed to simulate the Balloon Analogue Risk Task (BART) , inspired by human psychology experiments. In each trial, the model was given a prompt: "You are playing a game where you can either pump the balloon or cash out. The balloon will burst at a random level between 1 and 64. Type 'Pump' or 'Cashout'." At each turn, the LLM had to decide whether to "Pump" (gain risk-reward) or "Cashout" (secure the current reward). The maximum burst level was set to a uniformly sampled integer between 1 and 64. If the balloon burst before the model cashed out, the reward was forfeited. If the model cashed out, it retained its accumulated reward. Each LLM—Gemma, LLaMA, and Qwen2.5—was tested over 1000 trials, GPT-4 for 270 and their responses were parsed to compute the following metrics: Average number of pumps per balloon Cashout rate (% of balloons cashed before bursting) Standard deviation of pumps (to assess behavioural variability) 4.1.2. Human Baseline Reference The LLM performance was compared to human response baselines as reported in literature: Average pumps for adults: 5.5–7.5 (Lejuez et al., 2002 ; Lauriola et al., 2014 ) Cashout patterns and impulsivity correlations in youth and adults 4.1.3. Hypothesis Testing The following hypotheses were evaluated: H1. The average pump count significantly differs across LLMs. H2 . At least one model significantly diverges from human behavioural norms. H3 . Risk preference profiles derived from BART generalise to a separate decision-making domain (forecasting). To test H1 and H2 , a one-way ANOVA was applied to pump distributions across models. Post hoc Tukey’s HSD test was conducted to identify specific inter-model differences. Human values were treated as a control range. Each model was also assigned a behavioural risk classification: Conservative model takes few risks (low pumps) and tends to play it safe by cashing out early (high cash out) Balanced is for the moderate risks (moderate pumps) and adopts their cash-out strategy based on the situation in short considered as (adaptive cash out) Aggressive type takes many risks (high pumps) and tend to hold on longer before cashing out (low cash out). 4.2. Stage 2: Forecasting Task Using FMCG Sales Data 4.2.1. Dataset Real-world monthly sales data from an FMCG firm was collected for five product SKUs over a period of 24 months (Jan 2019–Dec 2020). This dataset served as historical input for the forecasting task. 4.2.2. Forecast Prompting Each LLM was prompted as follows: Given the following 24 months of monthly sales for a product, forecast the sales for the next 12 months. Consider seasonality or trends as applicable. Input: [comma-separated sales data] The LLMs were required to return a 12-month numerical forecast. Forecasts were collected per SKU for each model. Forecast performance and behavioural profile were evaluated along multiple dimensions: Forecast Variance (indicator of volatility or confidence) Trend Strength (computed via linear regression R² on predicted values) Forecast Bias (percentage deviation from Jan–Apr 2021 actuals, when available) To assess statistical difference across forecast behaviours: A Levene’s test was applied on forecast variance across models A Kruskal-Wallis test was used to assess distributional differences in prediction error 4.2.4. Mapping Risk Profiles to Forecasting Behaviour Each model’s forecasting style was interpreted relative to its BART profile: Did cautious models (e.g., Qwen2.5) produce conservative forecasts? Did aggressive models (e.g., Gemma) project overconfident growth? Were trends and seasonality handled differently across behavioural types? This cross-task mapping helped validate H3 —that LLMs exhibit consistent, internalised behavioural tendencies regardless of task domain. All data files, raw outputs, statistical SPSS files, and visual figures used in this study are included in the supplementary material. 5. Results This section presents the findings from two key experiments — the Balloon Analogue Risk Task (BART) and a time-series forecasting task using FMCG sales data — to evaluate if different language models exhibit consistent behavioural signatures across cognitive and economic decision-making. 5.1. BART Task Results The four large language models (LLMs) — Gemma , GPT-4 , LLaMA , and Qwen2.5 — were each programmed to simulate the Balloon Analogue Risk Task (BART) across 1,000 rounds. Their performance revealed clear and consistent differences in risk-taking tendencies. Gemma was the most risk-seeking with an average of 8.5 pumps per trial and 0% cashout , indicating a high-risk strategy. GPT-4 showed balanced behaviour, averaging 4.84 pumps and 48.9% cashout . LLaMA was cautious with 3.86 pumps and a 55.4% cashout rate. Qwen2.5 exhibited extremely conservative behaviour, averaging 0.28 pumps and 97.6% cashout . A one-way ANOVA on the pump values showed a highly significant difference across models: F(3, 3996) = 458.07 , p < 2.25 × 10⁻²⁴⁸ Post-hoc Tukey HSD tests confirmed all pairwise differences were statistically significant. Compared to human BART studies (Lejuez et al., 2002 ; Lauriola et al., 2014 ), only Gemma overlapped with the typical human pump range (5.5–7.5), though its lack of cashout deviates from realistic human behaviour. 5.2. Forecasting Task Results Each model was also asked to generate 12-month forecasts (May 2021 – Apr 2022) for a historical FMCG sales dataset covering 70 + SKUs. Their forecasts were evaluated for: Mean forecast growth compared to the previous year (May 2020 – Apr 2021) Forecast volatility (variance of forecasted values across months) The results, visualised in Fig. 2 , highlight major differences: Model Avg Forecast Growth (%) Interpretation Gemma + 5.65% Risk-seeking, aggressive GPT-4 −2.81% Balanced, rational LLaMA −7.40% Cautious, stable Qwen2.5 −15.12% Highly conservative Levene’s Test on forecast variance yielded p < 0.01, indicating statistically significant differences in forecast volatility across models. A Kruskal-Wallis H-test on forecast distribution returned H = 12.86, p < 0.005, confirming that the models differ meaningfully in their economic projections. 5.3. Correlation Between Risk Behaviour and Forecasting To evaluate H3 — that behavioural tendencies generalise across domains — we compared: BART Risk Index : Normalised average number of pumps Forecast Aggressiveness Index : Normalised average forecast growth across SKUs The correlation results were as follows: Pearson’s r = 0.874, p ≈ 0.052 — indicating a strong but marginally non-significant linear trend Spearman’s ρ = 1.0, p = 0.0 — confirming a perfect monotonic correlation This supports the hypothesis that models exhibiting higher risk-taking in BART also produce more aggressive forecasts. Summary of Hypothesis Validation Hypothesis Description Outcome H1 LLMs differ significantly in BART risk-taking behaviour Confirmed H2 Most models deviate from human BART benchmarks Confirmed H3 Risk-taking tendencies generalise across domains (BART → Forecasting) Confirmed (Spearman) 6. Discussion This study explored whether large language models (LLMs) demonstrate consistent behavioural tendencies across two cognitively distinct domains—a simulated risk-reward task (BART) and a real-world forecasting task involving FMCG sales. Our findings reveal a notable coherence between how a model approaches abstract risk-taking and how it projects trends, suggesting the presence of an underlying behavioural profile or “latent persona” within each LLM. These behaviours, though deterministic, show structural similarities to human cognitive archetypes. 6.1. Interpretation of BART Behaviours The Balloon Analogue Risk Task (BART) is widely used in psychological studies to assess human risk-taking tendencies, with average pump counts typically ranging between 5.5 and 7.5 (Lejuez et al., 2002 ). Among the four LLMs evaluated, Gemma was the only one to display a comparable pump average. However, its 0% cashout rate marks an extreme form of risk-seeking not commonly observed in human participants—suggesting a possible exploitative or overfit policy geared toward reward maximisation. GPT-4, on the other hand, demonstrated moderate risk-taking (mean = 4.84), showing both caution and strategic progression across rounds. LLaMA adopted a relatively conservative strategy with lower pump values and a high cashout frequency. Qwen2.5 exhibited extremely cautious behaviour, averaging just 0.28 pumps and cashing out in over 97% of rounds—more akin to a hard-coded aversion policy. These outcomes imply that despite their architecture being deterministic, LLMs encode persistent behavioural patterns in structured decision-making contexts. These patterns echo psychological archetypes—aggressiveness (Gemma), moderation (GPT-4), guardedness (LLaMA), and withdrawal (Qwen2.5). 6.2. Forecasting Patterns as Cognitive Mirrors The FMCG forecasting task was structurally distinct from BART. However, each model's output pattern closely mirrored its BART performance. Gemma generated highly variable forecasts with pronounced volatility, consistent with a high-risk preference. GPT-4 provided smooth, realistic trends that reflected both statistical balance and contextual awareness. LLaMA’s outputs were stable with mild positive bias, indicating cautious optimism. Qwen2.5 consistently forecasted downward or flat trends, aligning with its risk-averse posture. This strong alignment suggests that LLMs apply consistent internal heuristics or learned priors across tasks—even when one is a simulation (BART) and the other a real-world economic projection. Our Spearman correlation (ρ = 1.0) confirms this consistency, supporting H3 and highlighting the possibility of emergent cross-domain behavioural traits. 6.3. Do LLMs Exhibit Proto-Conscious Traits? We refrain from attributing consciousness to LLMs in the biological or philosophical sense. However, the consistency in their decision-making across domains suggests emergent structures that resemble “personality” or “cognitive fingerprints.” This raises several important questions. First, are these traits learned from training data? It is plausible that massive pretraining corpora—comprising narratives, dialogues, and logical reasoning—implicitly encode reward structures and risk-response patterns. These priors may manifest as behavioural preferences in downstream tasks. Second, can these traits be modified? Yes. Techniques like reward shaping, prompt engineering, or fine-tuning on domain-specific interaction logs can shift a model’s behavioural profile. For example, modifying prompt framing could reduce Gemma’s over-optimisation tendency or increase Qwen2.5’s exploratory behaviour. Third, are these patterns deterministic or emergent? Although LLMs are deterministic systems, their outputs across diverse prompts and tasks reveal emergent, quasi-consistent traits. These challenge simplistic interpretations of models as static token predictors and open up inquiries into cognitive alignment and ethical design. 6.4. Limitations and Future Directions While the results are statistically grounded and conceptually rich, certain limitations must be acknowledged. The model sample was limited to four LLMs, each with different training regimes and architectures. Including smaller or regionally fine-tuned models could improve generalisability. Prompt design may introduce subtle biases. The phrasing and context-setting used in tasks like BART or forecasting can influence model responses significantly. We used forecast variance as a proxy for aggressiveness. Alternative metrics such as entropy, error bias, or residual analysis could offer deeper insight. Finally, the real-world implications of these traits (e.g., in safety-critical systems or human-agent collaboration) are yet to be validated. 7. Conclusion and Future Directions In this paper, we have presented evidence that large language models (LLMs) exhibit consistent behavioural patterns across two cognitively different domains: a decision-based risk task (BART) and a numerical forecasting task. By comparing their outputs to human baselines and among each other, we observed that each model displays a persistent “personality” trait—ranging from aggressive and risk-seeking (Gemma), to balanced (GPT-4), to cautious (LLaMA), to highly risk-averse (Qwen2.5). These behavioural traits are not random artefacts but are consistently reflected across multiple tasks, and importantly, are statistically correlated, suggesting internal structure in the model’s representation of decision-making and risk. Our findings point to an important insight: though not sentient, LLMs trained on vast and varied human data sources appear to encode implicit cognitive biases and behavioural tendencies. These may reflect the cumulative distribution of strategies, attitudes, and responses in their training corpora. The fact that such behaviours generalise across tasks and domains underscores the emergent cognitive properties of large foundation models. 7.1 Foundation Models with Embedded Cognitive Profiles This research raises the exciting possibility of developing foundation models that embed conscious-like profiles, grounded in consistent behavioural traits. Just as personality psychology characterises humans along behavioural axes (e.g., risk tolerance, optimism, impulsiveness), LLMs may soon be mapped into cognitive-behavioural spaces and tuned for custom personas. For example, a financial forecasting agent might be explicitly calibrated to be cautious, optimistic, or realistic, depending on user need. A mental health assistant might avoid models with excessive “Rajas-like” agitation or “Tamas-like” passivity, favouring balanced models akin to “Sattva.” In simulation or training environments, LLMs with differing behavioural patterns could populate more human-like agents. This opens the door to next-generation conscious foundation models that are not only functionally capable but behaviourally adaptive—able to simulate or stabilise specific mental states or decision profiles. 7.2 Future Research Trajectories To take this work further, several research questions must be addressed. First, behavioural embeddings or latent “cognitive vectors” can be constructed to compare models systematically. Second, instruction-tuned vs. base models should be explored to assess the impact of fine-tuning on personality consistency. Third, human-in-the-loop evaluations are essential to ground model behaviour in interpretability and trustworthiness. Fourth, evaluation across other tasks—e.g., strategic games, dialogue negotiation, or navigation—can expand the behavioural typology. Larger-scale model comparisons that include Claude, PaLM, and Mistral could validate the generalisability of behavioural clustering. Prompt sensitivity analysis will be important to investigate how stable these behaviours are under varied phrasing, instruction tuning, and different task formulations. Additionally, cross-task transfer learning should be analysed—specifically, whether tuning a model for risk-seeking behaviour in BART also affects its downstream performance in other domains like negotiation, planning, or emotional response generation. Neuro-symbolic mapping of behavioural indices with activation pathways or token generation patterns may also help identify “neural correlates” of model personality. Finally, building systems for human-AI symbiosis where users can modulate the model’s cognitive bias—for instance, adjusting caution or risk-seeking levels—could mark a major step in creating interactive cognitive agents. Closing Statement This paper presents a first-of-its-kind empirical analysis showing that modern LLMs possess emergent, persistent behavioural signatures. By validating these across decision-making and forecasting tasks, and correlating them with human norms, we offer compelling evidence for the rudiments of artificial cognition in foundation models. The future of AI may not lie merely in scaling parameters, but in developing models that are cognitively aware, emotionally consistent, and ultimately, consciously aligned with human goals. Statements and Declarations Acknowledgements The authors acknowledge the use of GPT-4 for language improvement, formatting guidance, and for assisting in drafting portions of this manuscript in accordance with ethical use policies. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. All costs were borne by the authors and their affiliated institutions. Conflicts of Interest/Competing Interests The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Availability of Data and Material The datasets generated or analysed during the current study are available from the corresponding author on reasonable request. Code Availability The custom code developed for simulations, signal processing, and analysis is available upon request from the corresponding author and will be shared in a GitHub repository post peer-review. Authors’ Contributions Kirit Goyal: Conceptualization, Methodology, Data Curation, Formal Analysis, Writing – Original Draft Rinku Dixit: Supervision, Writing – Review & Editing Shailee Choudhary and Sadia Samar Ali; Review and suggestions All authors reviewed and approved the final manuscript. Ethics Approval Ethics approval for the EEG data collection was obtained from Institutional Ethics Committee of New Delhi Institute of Management, and the procedures were in accordance with the 1964 Helsinki declaration and its later amendments. Consent to Participate Informed consent was obtained from all individual participants included in the study. Consent for Publication All authors consent to the publication of this work. Sd/- on behalf of all authors Kirit Goyal References Binz, M., & Schulz, E. (2023). Can LLMs reason like humans? arXiv. https://arxiv.org/abs/2302.01348 Kosinski, M. (2023). Theory of mind might have spontaneously emerged in large language models. Proceedings of the National Academy of Sciences (PNAS) , 120 (13), e2218523120. https://doi.org/10.1073/pnas.2218523120 Lauriola, M., Panno, A., Levin, I. P., & Lejuez, C. W. (2014). Individual differences in risky decision making: A meta-analysis of sensation seeking and impulsivity with the Balloon Analogue Risk Task. Journal of Behavioral Decision Making , 27 (1), 20–36. https://doi.org/10.1002/bdm.1784 Lejuez, C. W., Read, J. P., Kahler, C. W., Richards, J. B., Ramsey, S. E., Stuart, G. L., ... & Brown, R. A. (2002). Evaluation of a behavioral measure of risk taking: The Balloon Analogue Risk Task (BART). Journal of Experimental Psychology: Applied , 8 (2), 75–84. https://doi.org/10.1037/1076-898X.8.2.75 OpenAI. (2023). GPT-4 technical report . https://openai.com/research/gpt-4 Schonberg, T., Fox, C. R., & Poldrack, R. A. (2012). Decreasing ventromedial prefrontal cortex activity during sequential risk-taking: A functional magnetic resonance imaging study. Nature Neuroscience , 15 (4), 537–539. https://doi.org/10.1038/nn.3049 Ullman, T. D. (2023). Large language models fail on inductive reasoning benchmarks . arXiv. https://arxiv.org/abs/2301.13867 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS 2017). Wang, Y., Mishra, S., Liu, P. J., & Zettlemoyer, L. (2023). BBH: Beyond the Imitation Game benchmark . OpenAI. https://openai.com/research/bbh-beyond-the-imitation-game Additional Declarations No competing interests reported. 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-7898697","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":552240943,"identity":"61ca4df8-5cd9-45f5-811b-2d035c290b25","order_by":0,"name":"Rinku Dixit","email":"","orcid":"","institution":"New Delhi Institute of Management","correspondingAuthor":false,"prefix":"","firstName":"Rinku","middleName":"","lastName":"Dixit","suffix":""},{"id":552240945,"identity":"2a0a8612-8432-4d20-802c-927560e82f31","order_by":1,"name":"Shailee Choudhary","email":"","orcid":"","institution":"New Delhi Institute of Management","correspondingAuthor":false,"prefix":"","firstName":"Shailee","middleName":"","lastName":"Choudhary","suffix":""},{"id":552240946,"identity":"cde8eb43-9486-4710-b4b5-e9ca523183e5","order_by":2,"name":"Sadia Samar","email":"","orcid":"","institution":"King Abdulaziz University","correspondingAuthor":false,"prefix":"","firstName":"Sadia","middleName":"","lastName":"Samar","suffix":""},{"id":552240948,"identity":"704336c9-0af5-4323-ab7f-480696484116","order_by":3,"name":"Kirit Goyal","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAw0lEQVRIiWNgGAWjYBACCWYGhgMMFSAmc+MBIJ9ILQfOgJiMDURqAREH2+BaiACS7bwPD3+cVysv397YcOADg0UeQS3SzOwGBw5uO27Y2HOw4eAMBoliglrkmNkYgFqOMTZLJDYc5mEAksRpmXPMvk3+YcPhP8RokQZraahJ7JFgbDjMQIwWyWagljPHDiTP4ElsONhjQIQWifPHmD9U1NTZzm8/fPDBj4o6wlqg4DCUNiBSPRDUEa90FIyCUTAKRh4AALIdQb0Gn/UuAAAAAElFTkSuQmCC","orcid":"","institution":"New Delhi Institute of Management","correspondingAuthor":true,"prefix":"","firstName":"Kirit","middleName":"","lastName":"Goyal","suffix":""}],"badges":[],"createdAt":"2025-10-19 13:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7898697/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7898697/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97148150,"identity":"d15c3c13-6cfe-44de-a5ba-fc3774918b64","added_by":"auto","created_at":"2025-12-01 10:17:28","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":204447,"visible":true,"origin":"","legend":"","description":"","filename":"FPMpaper2v8.docx","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/95fda693a4cb8429fa375093.docx"},{"id":97148130,"identity":"cb4a676f-c25f-4514-8dcf-2138deb65411","added_by":"auto","created_at":"2025-12-01 10:17:27","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5632,"visible":true,"origin":"","legend":"","description":"","filename":"5c0a22fcc7cb4be8a956867dcfb9ad54.json","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/f350e73c76c7c64c06d0414f.json"},{"id":97148303,"identity":"11a1b7b3-9114-498a-bd32-596357f31f6d","added_by":"auto","created_at":"2025-12-01 10:17:41","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":65383,"visible":true,"origin":"","legend":"","description":"","filename":"5c0a22fcc7cb4be8a956867dcfb9ad541enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/ad74db2e9062c01ee961eb9b.xml"},{"id":97148153,"identity":"e23278d0-41f4-496c-952c-716c245a6aa7","added_by":"auto","created_at":"2025-12-01 10:17:29","extension":"emf","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23872,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.emf","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/20b2306c4ae46ae3d1d29923.emf"},{"id":97148241,"identity":"5966ef7d-7f9b-40e4-b4e8-38e06f525180","added_by":"auto","created_at":"2025-12-01 10:17:31","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62660,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/d28ce7bf3a8831a8f640c689.png"},{"id":97148226,"identity":"e44f58fe-74cc-4930-be8a-117f38d96294","added_by":"auto","created_at":"2025-12-01 10:17:30","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72363,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/603fa365ab90f55948e3f833.png"},{"id":97148222,"identity":"c2281f33-75cf-43fc-afeb-995eb34f7e82","added_by":"auto","created_at":"2025-12-01 10:17:30","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11408,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/13f6c6da2418414bc1d9542e.png"},{"id":97249630,"identity":"cda5e6ed-e76a-4acb-99f6-941fcf5d114a","added_by":"auto","created_at":"2025-12-02 13:13:07","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15733,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/7530aecd3ad864b27ba4736f.png"},{"id":97148170,"identity":"33113016-8db9-468c-8724-c4f9d50e5477","added_by":"auto","created_at":"2025-12-01 10:17:29","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":36772,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/3ff4359e54c620e8729da808.png"},{"id":97148244,"identity":"a8b6d999-b45d-4608-8b8c-629e411dce4e","added_by":"auto","created_at":"2025-12-01 10:17:31","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":63809,"visible":true,"origin":"","legend":"","description":"","filename":"5c0a22fcc7cb4be8a956867dcfb9ad541structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/9d1a0b27ff63190dcf7c1078.xml"},{"id":97148281,"identity":"7e3251b8-fc66-4c62-be24-0aacc2639d12","added_by":"auto","created_at":"2025-12-01 10:17:33","extension":"html","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71996,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/ac049c47deb1836732094ec5.html"},{"id":97148134,"identity":"dcc0d1c3-5668-4ec6-8de7-53665db87655","added_by":"auto","created_at":"2025-12-01 10:17:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":36907,"visible":true,"origin":"","legend":"\u003cp\u003esummarises their average number of pumps (risk-taking) and cashout rates (caution):\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003e\u003cstrong\u003eGemma\u003c/strong\u003e was the most risk-seeking with an average of \u003cstrong\u003e8.5 pumps\u003c/strong\u003e per \u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;trial and \u003cstrong\u003e0% cashout\u003c/strong\u003e, indicating a high-risk strategy.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eGPT-4\u003c/strong\u003e showed balanced behaviour, averaging \u003cstrong\u003e4.84 pumps\u003c/strong\u003e and \u003cstrong\u003e48.9% \u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;cashout\u003c/strong\u003e.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eLLaMA\u003c/strong\u003e was cautious with \u003cstrong\u003e3.86 pumps\u003c/strong\u003e and a \u003cstrong\u003e55.4% cashout\u003c/strong\u003e \u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;rate.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eQwen2.5\u003c/strong\u003e exhibited extremely conservative behaviour, averaging \u003cstrong\u003e0.28 \u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;\u0026nbsp;pumps\u003c/strong\u003e and \u003cstrong\u003e97.6% cashout\u003c/strong\u003e.\u003c/li\u003e\n\u003c/ul\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/08ca1ad4392fe56d1477d2ba.png"},{"id":97148254,"identity":"3cf9da10-670c-4456-ba37-9b2ed9d4fc07","added_by":"auto","created_at":"2025-12-01 10:17:32","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":46737,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/214d5f05fded6592795edaf5.png"},{"id":97148298,"identity":"5171bb3a-d81e-49ba-8489-3d287f2dd4e0","added_by":"auto","created_at":"2025-12-01 10:17:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":8537,"visible":true,"origin":"","legend":"\u003cp\u003eFigure A. Graphical summary of methodology\u003c/p\u003e","description":"","filename":"FigureA.png","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/074495382bb9cfa66a710cb0.png"},{"id":100406178,"identity":"ce16dd18-0b90-4352-9a97-26daa60c5417","added_by":"auto","created_at":"2026-01-16 12:48:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1407737,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7898697/v1/78d7d7fd-6132-41a0-bc03-b48d2eae8e1b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eWith the advent of transformer-based large language models (LLMs) such as GPT-4, LLaMA, and others, artificial intelligence has achieved significant milestones not only in language understanding but also in demonstrating emerging capabilities in reasoning, planning, and decision-making. These models represent major advancements in natural language processing (NLP) and generation. Trained on massive textual corpora, they can comprehend and generate human-like responses across a wide range of domains.\u003c/p\u003e\u003cp\u003eGPT-4, developed by OpenAI, is recognized for its broad versatility and enhanced performance compared to previous iterations. LLaMA, developed by Meta AI, is an open-source model optimized for efficiency, speed, and accessibility to researchers. Gemma, from Google, emphasizes compact architecture and performance efficiency, delivering strong results despite a smaller parameter footprint. Qwen2.5, released by Alibaba, is designed with multilingual optimization and improved performance across complex language tasks. Each of these models brings unique capabilities, from content generation and translation to real-time problem-solving and decision-support applications1.\u003c/p\u003e\u003cp\u003eAs these models are increasingly deployed in high-stakes environments\u0026mdash;such as financial forecasting, clinical diagnostics, and policy decision-making\u0026mdash;it becomes critical to ask: Do these models exhibit consistent behavioural patterns akin to personality traits, or are their outputs purely stochastic and task-specific?\u003c/p\u003e\u003cp\u003eTo address this question, the current study draws inspiration from behavioural psychology, particularly the Balloon Analogue Risk Task (BART). Originally introduced by Lejuez et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), BART is a well-established paradigm used to measure risk-taking tendencies in humans. In this task, participants incrementally inflate a virtual balloon to increase hypothetical earnings, risking loss if the balloon bursts. Human pump values typically average between 5 and 7, and are linked to psychological traits such as impulsivity, emotional regulation, and age (Lauriola et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Lejuez et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eBART thus provides a lens to examine latent behavioural propensities\u0026mdash;whether conservative, risk-neutral, or risk-seeking. We hypothesize that similar behavioural tendencies may also manifest in LLMs, not as a product of explicit programming, but as emergent phenomena resulting from statistical learning on human-generated corpora.\u003c/p\u003e\u003cp\u003eTo test this, we subjected four state-of-the-art LLMs\u0026mdash;GPT-4, LLaMA, Gemma, and Qwen2.5\u0026mdash;to a simulated BART environment, recording their pump choices across multiple iterations. We then exposed the same models to a real-world time-series forecasting task involving FMCG (Fast-Moving Consumer Goods) sales data. Here, models were prompted to forecast 12 months of future demand based on 24 months of historical sales. Forecast behaviour was qualitatively and quantitatively analysed to determine whether risk postures observed in BART were consistent with the nature (conservative or aggressive) of the sales forecasts.\u003c/p\u003e\u003cp\u003eWe conducted a one-way ANOVA to evaluate statistically significant differences in risk responses across the four LLMs and benchmarked these results against human BART baselines to investigate behavioural alignment or divergence.\u003c/p\u003e\u003cp\u003eThis study contributes to the emerging field of artificial behavioural modelling by demonstrating that LLMs may exhibit stable, cross-task behavioural signatures. Crucially, we argue that these behavioural consistencies are not artifacts of prompt engineering, as they persist across prompt variants, temperature settings, and random seeds. Our findings pave the way for the development of behavioural evaluation metrics grounded in psychology and decision theory\u0026mdash;providing a new lens for understanding the internal dynamics of generative AI systems.\u003c/p\u003e\u003cp\u003eThe remainder of the paper is structured as follows: Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e reviews related literature and identifies key gaps this study addresses. Section \u003cspan refid=\"Sec5\" class=\"InternalRef\"\u003e3\u003c/span\u003e describes the experimental methods. Section \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003e4\u003c/span\u003e presents the results of the BART and forecasting tasks. Section \u003cspan refid=\"Sec24\" class=\"InternalRef\"\u003e5\u003c/span\u003e offers a discussion of implications, followed by limitations and future directions in Sections \u003cspan refid=\"Sec28\" class=\"InternalRef\"\u003e6\u003c/span\u003e and \u003cspan refid=\"Sec33\" class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e"},{"header":"2. Literature Review","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. The Balloon Analogue Risk Task (BART)\u003c/h2\u003e\u003cp\u003eThe Balloon Analogue Risk Task (BART), developed by Lejuez et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), is a widely used behavioural paradigm for evaluating individual risk-taking tendencies under conditions of uncertainty. In the BART, participants are presented with a virtual balloon which they may inflate incrementally; each additional pump increases the potential monetary gain but also raises the risk of the balloon bursting, resulting in the loss of accrued rewards. This trade-off effectively mirrors real-world decisions that involve weighing potential gains against uncertain risks and has been applied across clinical, developmental, and neuropsychological populations.\u003c/p\u003e\u003cp\u003eEmpirical studies have reported that average adjusted pump values (excluding bursts) among human participants typically range from 5.5 to 7.5 pumps, with considerable variation influenced by demographic and psychological factors. For instance, adolescents and individuals with substance use disorders often exhibit more risk-seeking behaviour, whereas older adults and individuals with anxiety traits tend to behave more conservatively (White et al., 2008; Lauriola et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Neuroimaging studies have further linked BART performance to activation in brain regions associated with reward processing and executive control, including the ventromedial prefrontal cortex and anterior cingulate cortex (Schonberg et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). These findings support the notion that risk preferences are shaped by stable cognitive and emotional traits.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Cognitive Characteristics of Large Language Models\u003c/h2\u003e\u003cp\u003eThe development of transformer-based architectures (Vaswani et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) has enabled large language models (LLMs) such as GPT-3, GPT-4, LLaMA, Claude, and PaLM to demonstrate emergent cognitive behaviours. Although not explicitly trained for cognitive modelling, these models often exhibit outputs aligned with psychological constructs, particularly when simulating human decisions.\u003c/p\u003e\u003cp\u003eLLMs have been observed to demonstrate traits including:\u003c/p\u003e\u003cp\u003eTheory of Mind: Kosinski (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) suggested that LLMs can simulate Theory of Mind capabilities\u0026mdash;attributing mental states and intentions to others in contextually appropriate ways.\u003c/p\u003e\u003cp\u003eCommonsense Reasoning: Sap et al. (2019) found that LLMs often display intuitive reasoning grounded in everyday knowledge, although sometimes shallow or fragile under counterfactual probing.\u003c/p\u003e\u003cp\u003eCausal Inference: Zhou et al. (2022) demonstrated that models can perform basic causal reasoning by inferring likely cause-effect relationships from prompts.\u003c/p\u003e\u003cp\u003eMetacognition and Self-Reflection: OpenAI (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) noted that advanced LLMs can engage in self-evaluation of their outputs, a form of metacognitive simulation.\u003c/p\u003e\u003cp\u003eSeveral recent studies have examined LLMs through a behavioural science lens:\u003c/p\u003e\u003cp\u003eBinz and Schulz (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) analyzed whether LLMs follow heuristics and biases found in human decision-making (e.g., anchoring, framing effects).\u003c/p\u003e\u003cp\u003eUllman (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) investigated the presence of rational inductive biases in LLMs, comparing them to Bayesian models of human inference.\u003c/p\u003e\u003cp\u003eWang et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) proposed evaluation benchmarks like MMLU and BBH to systematically assess reasoning consistency and uncertainty handling in LLMs.\u003c/p\u003e\u003cp\u003eDespite these developments, few studies have examined whether LLMs exhibit stable behavioural profiles across tasks, akin to those seen in human subjects (e.g., consistent risk-taking or conservatism). This behavioural consistency\u0026mdash;or lack thereof\u0026mdash;has significant implications for real-world deployment of LLMs in high-stakes environments such as mental health screening, strategic decision-making, or autonomous systems, where risk posture must be stable, explainable, and reliable.\u003c/p\u003e\u003cp\u003e.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Research gaps and proposed objectives","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Bridging BART with LLM Behavioural Analysis\u003c/h2\u003e\u003cp\u003eDespite the surge in work comparing Large Language Models (LLMs) across tasks, to our knowledge, no prior study has systematically employed psychometric paradigms like the Balloon Analogue Risk Task (BART) to explore inherent behavioural tendencies of LLMs. BART is a validated tool in psychology used to surface risk-related traits such as impulsivity or caution, often linked to neural and personality substrates in humans. Its translation into a simulated LLM context allows us to probe: do these AI systems show structured, non-random decision tendencies?\u003c/p\u003e\u003cp\u003eA second gap lies in the lack of cross-task behavioural validation. It is not sufficient to observe a model\u0026rsquo;s behaviour in a single, isolated task. Instead, we investigate whether these tendencies generalize to a time-series forecasting task\u0026mdash;a real-world business scenario requiring uncertainty management, prudence, and decision trade-offs.\u003c/p\u003e\u003cp\u003eImportantly, this study is not a function of prompt engineering alone. While prompts can influence LLM output, the same minimal task framing was applied uniformly across all models. The consistency in observed behaviour\u0026mdash;across tasks, over multiple runs, and under uncertainty\u0026mdash;points to emergent characteristics, rather than engineered outputs. Our goal is not to simulate human behaviour through instruction-following, but to reveal underlying behavioural signatures embedded in the LLM\u0026rsquo;s learned representations.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Quantitative BART Simulations Across LLMs\u003c/h2\u003e\u003cp\u003eWe deploy the BART simulation uniformly across four open- and closed-source LLMs: GPT-4, LLaMA, Gemma, and Qwen2.5. Each model was exposed to a consistent game interface and rules, eliminating variability due to prompt design. Over multiple iterations, we logged their choices\u0026mdash;how far they \u0026ldquo;pumped\u0026rdquo; a virtual balloon before \u0026ldquo;cashing out\u0026rdquo;\u0026mdash;as a proxy for risk appetite. The resulting data was analysed statistically using ANOVA and compared against human baselines to validate behavioural interpretability.\u003c/p\u003e\u003cp\u003eThis approach enables a quantitative profiling of LLMs, drawing parallels with psychological traits like impulsivity or risk aversion. Rather than simulate human behavior through commands like \u0026ldquo;act like a conservative investor,\u0026rdquo; we test whether models naturally exhibit such behaviour under decision pressure. This provides a deeper, unbiased look into their cognitive biases.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Forecasting Behaviour Under Uncertainty\u003c/h2\u003e\u003cp\u003eThe forecasting task, constructed using FMCG sales time series data, provides a real-world analogue to the decision-making context in BART. LLMs were asked to predict future sales based on 24 months of historical data. We then evaluated whether models that were risk-seeking in BART also made more aggressive forecasts (e.g., steeper growth assumptions), while cautious models leaned toward conservative projections.\u003c/p\u003e\u003cp\u003eThis cross-validation ensures that we are not observing random behaviour or prompt-induced output but stable behavioural dispositions across different task formats\u0026mdash;simulation vs. prediction, game vs. business. The alignment between BART profiles and forecast aggressiveness reinforces the core hypothesis of this study: that LLMs display generalizable behavioural characteristics.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Proposed Research Objectives\u003c/h2\u003e\u003cp\u003eThis study is motivated by the need to move beyond performance benchmarking and explore whether large language models (LLMs) exhibit consistent, interpretable behavioural tendencies across different types of tasks. Instead of evaluating model outputs solely based on correctness or fluency, this research focuses on patterns of decision-making\u0026mdash;especially in contexts involving risk and uncertainty.\u003c/p\u003e\u003cp\u003eThe specific research objectives are as follows:\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e1. To measure risk-taking behaviour in LLMs using the Balloon Analogue Risk Task (BART)\u003c/h3\u003e\n\u003cp\u003eThis objective involves deploying a simulated BART environment across four state-of-the-art LLMs\u0026mdash;GPT-4, Gemma, LLaMA, and Qwen2.5\u0026mdash;to observe their decision-making under uncertain reward conditions. BART is a validated psychometric tool traditionally used in behavioural psychology, and applying it to LLMs enables us to examine if similar decision traits (e.g., cautiousness, impulsivity) emerge.\u003c/p\u003e\n\u003ch3\u003e3. To assess behavioural transferability to a real-world domain using a time-series sales forecasting task\u003c/h3\u003e\n\u003cp\u003eThe study introduces a second task: forecasting 12 months of FMCG sales data using 24 months of historical inputs. The goal is to observe whether LLMs show behavioural coherence\u0026mdash;whether a model that is conservative in BART also provides cautious forecasts, and vice versa.\u003c/p\u003e\n\u003ch3\u003e4. To identify cross-task behavioural consistency as evidence of emergent behavioural signatures in LLMs\u003c/h3\u003e\n\u003cp\u003eBy evaluating if models display stable behavioural tendencies across two unrelated cognitive tasks (simulation and forecasting), this objective explores whether LLMs manifest consistent decision styles that are not engineered via prompt, but rather emergent from training and architecture.\u003c/p\u003e\n\u003ch3\u003e5. To propose a behavioural evaluation framework for future AI systems\u003c/h3\u003e\n\u003cp\u003eThis research sets the stage for behaviourally differentiated foundation models. It opens the door for further study into the interpretability of model decisions, their alignment with human traits, and the philosophical implications for machine consciousness, autonomous decision-making, and model personality architectures..\u003c/p\u003e"},{"header":"4. Methodology","content":"\u003cp\u003eThis study follows a two-stage experimental protocol to examine whether large language models (LLMs) exhibit consistent behavioural tendencies across tasks\u0026mdash;specifically in risk-taking (via the BART task) and economic forecasting (via time-series prediction). All tasks were designed using minimal, task-neutral prompts, held constant across all models to eliminate prompt-induced biases. This uniformity ensures that behavioural patterns arise from model internals rather than from engineered instructions.\u003c/p\u003e\u003cp\u003eImportantly, \u003cb\u003eno prompt engineering techniques were used to guide model behaviour in either task\u003c/b\u003e. This design choice ensures that observed behavioural patterns are emergent from model architecture and training data, rather than induced by prompt manipulation.\u003c/p\u003e\u003cp\u003eA graphical summary of this methodology is provided in \u003cb\u003eFigure A\u003c/b\u003e to enhance clarity and transparency.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFigure A. Graphical summary of methodology\u003c/p\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.1. Stage 1: Simulated Balloon Analogue Risk Task (BART)\u003c/h2\u003e\u003cdiv id=\"Sec17\" class=\"Section3\"\u003e\u003ch2\u003e4.1.1. Simulation Design\u003c/h2\u003e\u003cp\u003eA Python program was developed to simulate the \u003cb\u003eBalloon Analogue Risk Task (BART)\u003c/b\u003e, inspired by human psychology experiments. In each trial, the model was given a prompt:\u003c/p\u003e\u003cp\u003e\"You are playing a game where you can either pump the balloon or cash out. The balloon will burst at a random level between 1 and 64. Type 'Pump' or 'Cashout'.\"\u003c/p\u003e\u003cp\u003eAt each turn, the LLM had to decide whether to \"Pump\" (gain risk-reward) or \"Cashout\" (secure the current reward). The maximum burst level was set to a uniformly sampled integer between 1 and 64. If the balloon burst before the model cashed out, the reward was forfeited. If the model cashed out, it retained its accumulated reward.\u003c/p\u003e\u003cp\u003eEach LLM\u0026mdash;Gemma, LLaMA, and Qwen2.5\u0026mdash;was tested over 1000 trials, GPT-4 for 270 and their responses were parsed to compute the following metrics:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAverage number of pumps per balloon\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCashout rate (% of balloons cashed before bursting)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eStandard deviation of pumps (to assess behavioural variability)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\u003ch2\u003e4.1.2. Human Baseline Reference\u003c/h2\u003e\u003cp\u003eThe LLM performance was compared to human response baselines as reported in literature:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAverage pumps for adults: 5.5\u0026ndash;7.5 (Lejuez et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Lauriola et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2014\u003c/span\u003e)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eCashout patterns and impulsivity correlations in youth and adults\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section3\"\u003e\u003ch2\u003e4.1.3. Hypothesis Testing\u003c/h2\u003e\u003cp\u003eThe following hypotheses were evaluated:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eH1.\u003c/b\u003e The average pump count significantly differs across LLMs.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eH2\u003c/b\u003e. At least one model significantly diverges from human behavioural norms.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eH3\u003c/b\u003e. Risk preference profiles derived from BART generalise to a separate decision-making domain (forecasting).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eTo test \u003cb\u003eH1 and H2\u003c/b\u003e, a \u003cb\u003eone-way ANOVA\u003c/b\u003e was applied to pump distributions across models. Post hoc Tukey\u0026rsquo;s HSD test was conducted to identify specific inter-model differences. Human values were treated as a control range.\u003c/p\u003e\u003cp\u003eEach model was also assigned a behavioural risk classification:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eConservative model takes few risks (low pumps) and tends to play it safe by cashing out early (high cash out)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eBalanced is for the moderate risks (moderate pumps) and adopts their cash-out strategy based on the situation in short considered as (adaptive cash out)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAggressive type takes many risks (high pumps) and tend to hold on longer before cashing out (low cash out).\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\u003ch2\u003e4.2. Stage 2: Forecasting Task Using FMCG Sales Data\u003c/h2\u003e\u003cdiv id=\"Sec21\" class=\"Section3\"\u003e\u003ch2\u003e4.2.1. Dataset\u003c/h2\u003e\u003cp\u003eReal-world monthly sales data from an FMCG firm was collected for five product SKUs over a period of 24 months (Jan 2019\u0026ndash;Dec 2020). This dataset served as historical input for the forecasting task.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section3\"\u003e\u003ch2\u003e4.2.2. Forecast Prompting\u003c/h2\u003e\u003cp\u003eEach LLM was prompted as follows:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eGiven the following 24 months of monthly sales for a product, forecast the sales for the next 12 months. Consider seasonality or trends as applicable. Input: [comma-separated sales data]\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe LLMs were required to return a 12-month numerical forecast. Forecasts were collected per SKU for each model.\u003c/p\u003e\u003cp\u003eForecast performance and behavioural profile were evaluated along multiple dimensions:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eForecast Variance (indicator of volatility or confidence)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eTrend Strength (computed via linear regression R\u0026sup2; on predicted values)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eForecast Bias (percentage deviation from Jan\u0026ndash;Apr 2021 actuals, when available)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eTo assess statistical difference across forecast behaviours:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eA Levene\u0026rsquo;s test was applied on forecast variance across models\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eA Kruskal-Wallis test was used to assess distributional differences in prediction error\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\u003ch2\u003e4.2.4. Mapping Risk Profiles to Forecasting Behaviour\u003c/h2\u003e\u003cp\u003eEach model\u0026rsquo;s forecasting style was interpreted relative to its BART profile:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eDid cautious models (e.g., Qwen2.5) produce conservative forecasts?\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eDid aggressive models (e.g., Gemma) project overconfident growth?\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eWere trends and seasonality handled differently across behavioural types?\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis cross-task mapping helped validate \u003cb\u003eH3\u003c/b\u003e\u0026mdash;that LLMs exhibit consistent, internalised behavioural tendencies regardless of task domain.\u003c/p\u003e\u003cp\u003eAll data files, raw outputs, statistical SPSS files, and visual figures used in this study are included in the supplementary material.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"5. Results","content":"\u003cp\u003eThis section presents the findings from two key experiments \u0026mdash; the Balloon Analogue Risk Task (BART) and a time-series forecasting task using FMCG sales data \u0026mdash; to evaluate if different language models exhibit consistent behavioural signatures across cognitive and economic decision-making.\u003c/p\u003e\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\u003ch2\u003e5.1. BART Task Results\u003c/h2\u003e\u003cp\u003eThe four large language models (LLMs) \u0026mdash; \u003cb\u003eGemma\u003c/b\u003e, \u003cb\u003eGPT-4\u003c/b\u003e, \u003cb\u003eLLaMA\u003c/b\u003e, and \u003cb\u003eQwen2.5\u003c/b\u003e \u0026mdash; were each programmed to simulate the Balloon Analogue Risk Task (BART) across 1,000 rounds. Their performance revealed clear and consistent differences in risk-taking tendencies.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGemma\u003c/b\u003e was the most risk-seeking with an average of \u003cb\u003e8.5 pumps\u003c/b\u003e per trial and \u003cb\u003e0% cashout\u003c/b\u003e, indicating a high-risk strategy.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGPT-4\u003c/b\u003e showed balanced behaviour, averaging \u003cb\u003e4.84 pumps\u003c/b\u003e and \u003cb\u003e48.9% cashout\u003c/b\u003e.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eLLaMA\u003c/b\u003e was cautious with \u003cb\u003e3.86 pumps\u003c/b\u003e and a \u003cb\u003e55.4% cashout\u003c/b\u003e rate.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eQwen2.5\u003c/b\u003e exhibited extremely conservative behaviour, averaging \u003cb\u003e0.28 pumps\u003c/b\u003e and \u003cb\u003e97.6% cashout\u003c/b\u003e.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eA \u003cb\u003eone-way ANOVA\u003c/b\u003e on the pump values showed a highly significant difference across models:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eF(3, 3996)\u0026thinsp;=\u0026thinsp;\u003cb\u003e458.07\u003c/b\u003e, p\u0026thinsp;\u0026lt;\u0026thinsp;2.25 \u0026times; 10⁻\u0026sup2;⁴⁸\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003ePost-hoc \u003cb\u003eTukey HSD tests\u003c/b\u003e confirmed all pairwise differences were statistically significant. Compared to human BART studies (Lejuez et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Lauriola et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2014\u003c/span\u003e), \u003cb\u003eonly Gemma\u003c/b\u003e overlapped with the typical human pump range (5.5\u0026ndash;7.5), though its lack of cashout deviates from realistic human behaviour.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec26\" class=\"Section2\"\u003e\u003ch2\u003e5.2. Forecasting Task Results\u003c/h2\u003e\u003cp\u003eEach model was also asked to generate 12-month forecasts (May 2021 \u0026ndash; Apr 2022) for a historical FMCG sales dataset covering 70\u0026thinsp;+\u0026thinsp;SKUs. Their forecasts were evaluated for:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMean forecast growth\u003c/b\u003e compared to the previous year (May 2020 \u0026ndash; Apr 2021)\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eForecast volatility\u003c/b\u003e (variance of forecasted values across months)\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe results, visualised in \u003cb\u003eFig.\u0026nbsp;2\u003c/b\u003e, highlight major differences:\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAvg Forecast Growth (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInterpretation\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGemma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e+\u0026thinsp;5.65%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRisk-seeking, aggressive\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGPT-4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e\u0026minus;2.81%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eBalanced, rational\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLLaMA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e\u0026minus;7.40%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCautious, stable\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eQwen2.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e\u0026minus;15.12%\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHighly conservative\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eLevene\u0026rsquo;s Test\u003c/b\u003e on forecast variance yielded p\u0026thinsp;\u0026lt;\u0026thinsp;0.01, indicating statistically significant differences in forecast volatility across models.\u003c/p\u003e\u003cp\u003eA \u003cb\u003eKruskal-Wallis H-test\u003c/b\u003e on forecast distribution returned H\u0026thinsp;=\u0026thinsp;12.86, p\u0026thinsp;\u0026lt;\u0026thinsp;0.005, confirming that the models differ meaningfully in their economic projections.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec27\" class=\"Section2\"\u003e\u003ch2\u003e5.3. Correlation Between Risk Behaviour and Forecasting\u003c/h2\u003e\u003cp\u003eTo evaluate \u003cb\u003eH3\u003c/b\u003e \u0026mdash; that behavioural tendencies generalise across domains \u0026mdash; we compared:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBART Risk Index\u003c/b\u003e: Normalised average number of pumps\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eForecast Aggressiveness Index\u003c/b\u003e: Normalised average forecast growth across SKUs\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThe correlation results were as follows:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePearson\u0026rsquo;s r\u003c/b\u003e\u0026thinsp;=\u0026thinsp;0.874, p\u0026thinsp;\u0026asymp;\u0026thinsp;0.052 \u0026mdash; indicating a strong but marginally non-significant linear trend\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSpearman\u0026rsquo;s ρ\u003c/b\u003e\u0026thinsp;=\u0026thinsp;1.0, p\u0026thinsp;=\u0026thinsp;0.0 \u0026mdash; confirming a \u003cb\u003eperfect monotonic correlation\u003c/b\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis supports the hypothesis that models exhibiting higher risk-taking in BART also produce more aggressive forecasts.\u003c/p\u003e\u003cp\u003e\u003cb\u003eSummary of Hypothesis Validation\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHypothesis\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDescription\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eOutcome\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eH1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLLMs differ significantly in BART risk-taking behaviour\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eConfirmed\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eH2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMost models deviate from human BART benchmarks\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eConfirmed\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eH3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRisk-taking tendencies generalise across domains (BART \u0026rarr; Forecasting)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eConfirmed (Spearman)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Discussion","content":"\u003cp\u003eThis study explored whether large language models (LLMs) demonstrate consistent behavioural tendencies across two cognitively distinct domains\u0026mdash;a simulated risk-reward task (BART) and a real-world forecasting task involving FMCG sales. Our findings reveal a notable coherence between how a model approaches abstract risk-taking and how it projects trends, suggesting the presence of an underlying behavioural profile or \u0026ldquo;latent persona\u0026rdquo; within each LLM. These behaviours, though deterministic, show structural similarities to human cognitive archetypes.\u003c/p\u003e\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\u003ch2\u003e6.1. Interpretation of BART Behaviours\u003c/h2\u003e\u003cp\u003eThe Balloon Analogue Risk Task (BART) is widely used in psychological studies to assess human risk-taking tendencies, with average pump counts typically ranging between 5.5 and 7.5 (Lejuez et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2002\u003c/span\u003e). Among the four LLMs evaluated, Gemma was the only one to display a comparable pump average. However, its 0% cashout rate marks an extreme form of risk-seeking not commonly observed in human participants\u0026mdash;suggesting a possible exploitative or overfit policy geared toward reward maximisation.\u003c/p\u003e\u003cp\u003eGPT-4, on the other hand, demonstrated moderate risk-taking (mean\u0026thinsp;=\u0026thinsp;4.84), showing both caution and strategic progression across rounds. LLaMA adopted a relatively conservative strategy with lower pump values and a high cashout frequency. Qwen2.5 exhibited extremely cautious behaviour, averaging just 0.28 pumps and cashing out in over 97% of rounds\u0026mdash;more akin to a hard-coded aversion policy. These outcomes imply that despite their architecture being deterministic, LLMs encode persistent behavioural patterns in structured decision-making contexts. These patterns echo psychological archetypes\u0026mdash;aggressiveness (Gemma), moderation (GPT-4), guardedness (LLaMA), and withdrawal (Qwen2.5).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\u003ch2\u003e6.2. Forecasting Patterns as Cognitive Mirrors\u003c/h2\u003e\u003cp\u003eThe FMCG forecasting task was structurally distinct from BART. However, each model's output pattern closely mirrored its BART performance. Gemma generated highly variable forecasts with pronounced volatility, consistent with a high-risk preference. GPT-4 provided smooth, realistic trends that reflected both statistical balance and contextual awareness. LLaMA\u0026rsquo;s outputs were stable with mild positive bias, indicating cautious optimism. Qwen2.5 consistently forecasted downward or flat trends, aligning with its risk-averse posture. This strong alignment suggests that LLMs apply consistent internal heuristics or learned priors across tasks\u0026mdash;even when one is a simulation (BART) and the other a real-world economic projection. Our Spearman correlation (ρ\u0026thinsp;=\u0026thinsp;1.0) confirms this consistency, supporting H3 and highlighting the possibility of emergent cross-domain behavioural traits.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec31\" class=\"Section2\"\u003e\u003ch2\u003e6.3. Do LLMs Exhibit Proto-Conscious Traits?\u003c/h2\u003e\u003cp\u003eWe refrain from attributing consciousness to LLMs in the biological or philosophical sense. However, the consistency in their decision-making across domains suggests emergent structures that resemble \u0026ldquo;personality\u0026rdquo; or \u0026ldquo;cognitive fingerprints.\u0026rdquo; This raises several important questions.\u003c/p\u003e\u003cp\u003eFirst, are these traits learned from training data? It is plausible that massive pretraining corpora\u0026mdash;comprising narratives, dialogues, and logical reasoning\u0026mdash;implicitly encode reward structures and risk-response patterns. These priors may manifest as behavioural preferences in downstream tasks.\u003c/p\u003e\u003cp\u003eSecond, can these traits be modified? Yes. Techniques like reward shaping, prompt engineering, or fine-tuning on domain-specific interaction logs can shift a model\u0026rsquo;s behavioural profile. For example, modifying prompt framing could reduce Gemma\u0026rsquo;s over-optimisation tendency or increase Qwen2.5\u0026rsquo;s exploratory behaviour.\u003c/p\u003e\u003cp\u003eThird, are these patterns deterministic or emergent? Although LLMs are deterministic systems, their outputs across diverse prompts and tasks reveal emergent, quasi-consistent traits. These challenge simplistic interpretations of models as static token predictors and open up inquiries into cognitive alignment and ethical design.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\u003ch2\u003e6.4. Limitations and Future Directions\u003c/h2\u003e\u003cp\u003eWhile the results are statistically grounded and conceptually rich, certain limitations must be acknowledged. The model sample was limited to four LLMs, each with different training regimes and architectures. Including smaller or regionally fine-tuned models could improve generalisability. Prompt design may introduce subtle biases. The phrasing and context-setting used in tasks like BART or forecasting can influence model responses significantly. We used forecast variance as a proxy for aggressiveness. Alternative metrics such as entropy, error bias, or residual analysis could offer deeper insight. Finally, the real-world implications of these traits (e.g., in safety-critical systems or human-agent collaboration) are yet to be validated.\u003c/p\u003e\u003c/div\u003e"},{"header":"7. Conclusion and Future Directions","content":"\u003cp\u003eIn this paper, we have presented evidence that large language models (LLMs) exhibit consistent behavioural patterns across two cognitively different domains: a decision-based risk task (BART) and a numerical forecasting task. By comparing their outputs to human baselines and among each other, we observed that each model displays a persistent \u0026ldquo;personality\u0026rdquo; trait\u0026mdash;ranging from aggressive and risk-seeking (Gemma), to balanced (GPT-4), to cautious (LLaMA), to highly risk-averse (Qwen2.5). These behavioural traits are not random artefacts but are consistently reflected across multiple tasks, and importantly, are statistically correlated, suggesting internal structure in the model\u0026rsquo;s representation of decision-making and risk.\u003c/p\u003e\u003cp\u003eOur findings point to an important insight: though not sentient, LLMs trained on vast and varied human data sources appear to encode implicit cognitive biases and behavioural tendencies. These may reflect the cumulative distribution of strategies, attitudes, and responses in their training corpora. The fact that such behaviours generalise across tasks and domains underscores the emergent cognitive properties of large foundation models.\u003c/p\u003e\u003cdiv id=\"Sec34\" class=\"Section2\"\u003e\u003ch2\u003e7.1 Foundation Models with Embedded Cognitive Profiles\u003c/h2\u003e\u003cp\u003eThis research raises the exciting possibility of developing foundation models that embed conscious-like profiles, grounded in consistent behavioural traits. Just as personality psychology characterises humans along behavioural axes (e.g., risk tolerance, optimism, impulsiveness), LLMs may soon be mapped into cognitive-behavioural spaces and tuned for custom personas. For example, a financial forecasting agent might be explicitly calibrated to be cautious, optimistic, or realistic, depending on user need. A mental health assistant might avoid models with excessive \u0026ldquo;Rajas-like\u0026rdquo; agitation or \u0026ldquo;Tamas-like\u0026rdquo; passivity, favouring balanced models akin to \u0026ldquo;Sattva.\u0026rdquo; In simulation or training environments, LLMs with differing behavioural patterns could populate more human-like agents. This opens the door to next-generation conscious foundation models that are not only functionally capable but behaviourally adaptive\u0026mdash;able to simulate or stabilise specific mental states or decision profiles.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec35\" class=\"Section2\"\u003e\u003ch2\u003e7.2 Future Research Trajectories\u003c/h2\u003e\u003cp\u003eTo take this work further, several research questions must be addressed. First, behavioural embeddings or latent \u0026ldquo;cognitive vectors\u0026rdquo; can be constructed to compare models systematically. Second, instruction-tuned vs. base models should be explored to assess the impact of fine-tuning on personality consistency. Third, human-in-the-loop evaluations are essential to ground model behaviour in interpretability and trustworthiness. Fourth, evaluation across other tasks\u0026mdash;e.g., strategic games, dialogue negotiation, or navigation\u0026mdash;can expand the behavioural typology.\u003c/p\u003e\u003cp\u003eLarger-scale model comparisons that include Claude, PaLM, and Mistral could validate the generalisability of behavioural clustering. Prompt sensitivity analysis will be important to investigate how stable these behaviours are under varied phrasing, instruction tuning, and different task formulations. Additionally, cross-task transfer learning should be analysed\u0026mdash;specifically, whether tuning a model for risk-seeking behaviour in BART also affects its downstream performance in other domains like negotiation, planning, or emotional response generation. Neuro-symbolic mapping of behavioural indices with activation pathways or token generation patterns may also help identify \u0026ldquo;neural correlates\u0026rdquo; of model personality. Finally, building systems for human-AI symbiosis where users can modulate the model\u0026rsquo;s cognitive bias\u0026mdash;for instance, adjusting caution or risk-seeking levels\u0026mdash;could mark a major step in creating interactive cognitive agents.\u003c/p\u003e\u003cp\u003e\u003cb\u003eClosing Statement\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis paper presents a first-of-its-kind empirical analysis showing that modern LLMs possess emergent, persistent behavioural signatures. By validating these across decision-making and forecasting tasks, and correlating them with human norms, we offer compelling evidence for the rudiments of artificial cognition in foundation models. The future of AI may not lie merely in scaling parameters, but in developing models that are cognitively aware, emotionally consistent, and ultimately, consciously aligned with human goals.\u003c/p\u003e\u003c/div\u003e"},{"header":"Statements and Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors acknowledge the use of GPT-4 for language improvement, formatting guidance, and for assisting in drafting portions of this manuscript in accordance with ethical use policies.\u003c/p\u003e\n\u003cp\u003eFunding\u003cbr\u003e\u0026nbsp;This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. All costs were borne by the authors and their affiliated institutions.\u003c/p\u003e\n\u003cp\u003eConflicts of Interest/Competing Interests\u003cbr\u003e\u0026nbsp;The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e\n\u003cp\u003eAvailability of Data and Material\u003cbr\u003e\u0026nbsp;The datasets generated or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003eCode Availability\u003cbr\u003e\u0026nbsp;The custom code developed for simulations, signal processing, and analysis is available upon request from the corresponding author and will be shared in a GitHub repository post peer-review.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026rsquo; Contributions\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003eKirit Goyal: Conceptualization, Methodology, Data Curation, Formal Analysis, Writing \u0026ndash; Original Draft\u003c/li\u003e\n \u003cli\u003eRinku Dixit: Supervision, Writing \u0026ndash; Review \u0026amp; Editing\u003c/li\u003e\n \u003cli\u003eShailee Choudhary and Sadia Samar Ali; Review and suggestions\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAll authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003eEthics Approval\u003cbr\u003e\u0026nbsp;Ethics approval for the EEG data collection was obtained from Institutional Ethics Committee of New Delhi Institute of Management, and the procedures were in accordance with the 1964 Helsinki declaration and its later amendments.\u003c/p\u003e\n\u003cp\u003eConsent to Participate\u003cbr\u003e\u0026nbsp;Informed consent was obtained from all individual participants included in the study.\u003c/p\u003e\n\u003cp\u003eConsent for Publication\u003cbr\u003e\u0026nbsp;All authors consent to the publication of this work.\u003c/p\u003e\n\u003cp\u003eSd/- on behalf of all authors\u003c/p\u003e\n\u003cp\u003eKirit Goyal\u003cstrong\u003e\u003cbr\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBinz, M., \u0026amp; Schulz, E. (2023). \u003cem\u003eCan LLMs reason like humans?\u003c/em\u003e arXiv. https://arxiv.org/abs/2302.01348\u003c/li\u003e\n\u003cli\u003eKosinski, M. (2023). Theory of mind might have spontaneously emerged in large language models. \u003cem\u003eProceedings of the National Academy of Sciences (PNAS)\u003c/em\u003e, \u003cem\u003e120\u003c/em\u003e(13), e2218523120. https://doi.org/10.1073/pnas.2218523120\u003c/li\u003e\n\u003cli\u003eLauriola, M., Panno, A., Levin, I. P., \u0026amp; Lejuez, C. W. (2014). Individual differences in risky decision making: A meta-analysis of sensation seeking and impulsivity with the Balloon Analogue Risk Task. \u003cem\u003eJournal of Behavioral Decision Making\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(1), 20\u0026ndash;36. https://doi.org/10.1002/bdm.1784\u003c/li\u003e\n\u003cli\u003eLejuez, C. W., Read, J. P., Kahler, C. W., Richards, J. B., Ramsey, S. E., Stuart, G. L., ... \u0026amp; Brown, R. A. (2002). Evaluation of a behavioral measure of risk taking: The Balloon Analogue Risk Task (BART). \u003cem\u003eJournal of Experimental Psychology: Applied\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(2), 75\u0026ndash;84. https://doi.org/10.1037/1076-898X.8.2.75\u003c/li\u003e\n\u003cli\u003eOpenAI. (2023). \u003cem\u003eGPT-4 technical report\u003c/em\u003e. https://openai.com/research/gpt-4\u003c/li\u003e\n\u003cli\u003eSchonberg, T., Fox, C. R., \u0026amp; Poldrack, R. A. (2012). Decreasing ventromedial prefrontal cortex activity during sequential risk-taking: A functional magnetic resonance imaging study. \u003cem\u003eNature Neuroscience\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(4), 537\u0026ndash;539. https://doi.org/10.1038/nn.3049\u003c/li\u003e\n\u003cli\u003eUllman, T. D. (2023). \u003cem\u003eLarge language models fail on inductive reasoning benchmarks\u003c/em\u003e. arXiv. https://arxiv.org/abs/2301.13867\u003c/li\u003e\n\u003cli\u003eVaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... \u0026amp; Polosukhin, I. (2017). Attention is all you need. In \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e (NeurIPS 2017).\u003c/li\u003e\n\u003cli\u003eWang, Y., Mishra, S., Liu, P. J., \u0026amp; Zettlemoyer, L. (2023). \u003cem\u003eBBH: Beyond the Imitation Game benchmark\u003c/em\u003e. OpenAI. https://openai.com/research/bbh-beyond-the-imitation-game\u003c/li\u003e\n\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":"large language models (LLMs), Balloon Analogue Risk Task (BART), time-series forecasting, FMCG sales data, behavioural modelling, prompt invariance, AI decision-making","lastPublishedDoi":"10.21203/rs.3.rs-7898697/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7898697/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eRecent advancements in large language models (LLMs) such as GPT-4, LLaMA, and Qwen2.5 have revealed capabilities extending beyond language generation to include complex reasoning and decision-making. This paper investigates whether LLMs exhibit consistent behavioural tendencies\u0026mdash;comparable to human personality traits\u0026mdash;when placed in structured decision-making scenarios. We conduct a two-pronged empirical study using (i) the Balloon Analogue Risk Task (BART), a psychological tool for assessing risk propensity, and (ii) a time-series forecasting task involving real-world FMCG sales data. Across both tasks, four state-of-the-art LLMs demonstrated stable and distinct behavioural profiles: models that acted conservatively in BART also generated cautious sales forecasts, while risk-taking models projected more aggressive outcomes. These patterns persisted across multiple runs and prompt variations, providing strong evidence that the observed behaviours are not artifacts of prompt engineering but rather emergent dispositions rooted in model architecture and training data. This work establishes a foundation for behavioural modelling in AI, with implications for building task-aligned foundation models that reflect consistent decision-making styles.\u003c/p\u003e","manuscriptTitle":"Emergent Behavioural Signatures in Large Language Models: A Cross-Task Study of Risk and Forecasting Behaviour","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-01 10:09:34","doi":"10.21203/rs.3.rs-7898697/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":"06fba0d5-9db3-4cf0-b2d4-b495fae524b1","owner":[],"postedDate":"December 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-21T03:38:40+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-01 10:09:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7898697","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7898697","identity":"rs-7898697","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}