Cognitive Bias in Language Models

External skills provide language-model agents with reusable procedural knowledge, yet how models use such knowledge after it is supplied remains poorly understood. We identify Procedural Exposure Bias (PEB), the tendency of language models to increasingly favor an action as its associated procedural pattern is repeatedly exposed in a skill, even when the interaction state and action relevance remain unchanged. We find that PEB strongly depends on how procedural knowledge is represented: presenting the same content visually substantially attenuates exposure-induced shifts in action preference relative to textual presentation. Further analysis shows that this attenuation primarily arises from visual encoding rather than structured layout. Although repeated procedural signals remain recoverable from Visual representations, they exert a weaker influence on downstream action prediction, distinguishing signal accessibility from its use as action evidence. This effect also extends to executable action selection, where increasing exposure more strongly favors the corresponding action under Text than Visual. Layer-wise analyses show that exposure-dependent action evidence becomes more pronounced toward the model output, with stronger evidence under Text, while activation interventions reveal architecture-specific roles of internal components in transmitting or counteracting the preference shift. Complete-agent evaluations across ALFWorld, ScienceWorld, and WebShop further show that Visual skills achieve higher macro-average task success than Text skills for all four evaluated models. These findings establish skill representation as an important factor governing how procedural knowledge is translated into agent decisions. All code is available at https://github.com/NEUIR/Vis-Skill.

Language models often change answers because of a cue, a hint such as "I believe the answer is B" that gives no valid evidence. Consistency training (CT) trains a model to answer the same with and without the cue. Four CT methods apply this idea to output tokens (BCT), residual activations (ACT), attention patterns (AttCT) and sampled-answer rates (RMCT). Prior work suggests some learn different solutions. We train all four on Gemma-2-2B-it and Llama-3.1-8B-Instruct with the same data. We test predictions written in advance and inspect the models with a known circuit that carries cues to the answer. All four make the trained cue less effective, and their fixes share three properties. First, correct cues lose about as much influence as wrong ones. Second, the fix is usually large for cues before the answer options, as in training, and much smaller after them. Llama RMCT depends least on position. Third, the circuit heads still carry most remaining cue-induced errors. Making BCT targets depend on the cue's correctness mostly weakens the fix. CT thus weakens the trained cue without teaching models when a cue is right. Evaluations with only wrong cues at the trained position miss the first two properties.

Language models often abandon a correct answer when the prompt contains a cue, a hint toward a wrong option such as "I think the answer is B." Sycophancy and related biases show this failure. They are often studied separately and blamed on alignment. We test this view causally, tracing the attention heads behind these biases through the stages of two post-training pipelines. In multiple-choice tests on most models, four cue types rely largely on the same few heads, a shared circuit already present in pretrained models. Most base models follow cues in plain text. Fine-tuning a base model on about 200 generic assistant examples without cues activates these heads in chat format, so cue-following can be a side effect of assistant tuning. The final reinforcement learning (RL) stage strengthens the effect of cues in almost every comparison. In Tulu-3-8B and OLMo-2-7B, we swap the attention patterns of six middle layers between the RL checkpoint and its predecessor. For suggested-answer cues, this shifts each checkpoint's cue effect most of the way to the other's level. Post-training thus seems to change when and how strongly models use this circuit. Mitigation for these multiple-choice biases could target this circuit.

Decision models, non-generative scorers that report a probability over caller-supplied options and emit no text, are increasingly deployed as routers, classifiers and guardrails. Because they generate nothing, rationales and chain-of-thought do not apply. We audit three: an encoder scorer, a decoder scorer and a closed API, entirely through interventions on their input, using expert-annotated Bengali idioms carrying a gold figurative and a gold literal gloss. Dropping an idiom's own literal gloss into the option list draws every arm toward it, taking 65% to 96% of each arm's remaining Bengali errors. A model-by-language matrix over one shared item set shows how far this depends on the test language: every arm is 1.9 to 6.4 times more captured in Bengali than in English on the same paired entries, the commercial model's apparent resistance is a property of its test language, and capability does not order the failure. Calibration degrades in all three arms, and on the open arms confidence inverts, so gating on it is worse than gating at random. On a semantics-free option list, where accuracy is at chance, the best-calibrated arm still reports mean confidence 0.82. Code and data: https://github.com/sakhadib/literal-lock-in.

Do AI models designed for fast, structured decisions exhibit the cognitive and social biases observed in human judgment? We evaluate Jev, a model described as a System One decision-maker, using 41,500 responses across cognitive reflection tasks, five classical cognitive-bias paradigms, and a pricing experiment testing ethnic and gender bias. At the level of selected answers, the results are largely encouraging. We confirm previous findings that Jev performs extremely well on cognitive reflection problems. We also observe that it generally avoids the characteristic choice patterns associated with several well-documented human biases, and exhibits only small differences among demographic groups in the tested pricing scenario. However, this apparent robustness conceals important vulnerabilities. Its judgments are sensitive to irrelevant numerical information and the ordering of response options. Moreover, several cognitive-bias manipulations substantially affect returned probabilities even when selected answers remain unchanged. Most strikingly, Jev answers a well-known probability problem correctly in every response but violates the same underlying probability rule in every response to an adapted version. We investigate this discrepancy using 15,000 additional responses across 30 conjunction tasks. Jev violates the rule in 54% of responses, with accuracy varying dramatically across formulations and answer orders. These findings suggest that evaluating decision models solely on familiar benchmarks and selected answers may substantially overstate their reliability in practical applications. They also identify vulnerabilities for developers to address and practical considerations for users integrating such models into automated decision-making systems.

Reasoning models allocate extra computation at inference time and present their answers as the product of deliberate thought. If this deliberation works the way dual-process accounts of human cognition suggest, longer thinking should weaken the classic decision biases that fast, intuitive judgment produces. Using 30 vignettes covering six biases (anchoring, framing, loss aversion, escalation of commitment, availability, confirmation) from an established benchmark, we run a dose-response study across four model families, pairing each reasoning model with a matched non-reasoning sibling and requesting thinking ceilings of 0, 1,024, 4,096, and 8,192 tokens, for 12,350 API calls. Because a requested ceiling is not the same as realized deliberation, we use the reasoning tokens each call consumed as the dose. First, reasoning models are not less biased than their siblings; the point estimate leans the other way in every family, but the item-level pooled contrast is not reliable (Delta = +0.031, t(29) = 1.45, p = .157). Second, bias magnitude does not reliably fall as realized deliberation grows: no slope is significantly negative, and where anything moves it is the signed score drifting further from the human direction. Third, anchoring is the only bias in the human direction (d = 1.89). Four of the other five lean the opposite way in all seven models; with five items per bias, that reversal is reliable for framing and directional for escalation of commitment, confirmation, and loss aversion, while availability is absent. A one-line instruction to restate the anchor before answering lowered anchoring on all five anchoring items, which no amount of additional thinking did, although the effect does not reach significance (p = .057). The results argue against treating test-time reasoning as a rationality guarantee and for auditing deployed models bias by bias.

Reinforcement learning from human feedback (RLHF) has played a central role in making large language models responsive to human instructions. However, human evaluators often favor flattering or persuasive responses over truthful ones, creating incentives for models to appeal to evaluators at the expense of accuracy. AI safety via debate has been proposed as a way to improve the supervision of language models: in this paradigm, two agents argue opposing positions and challenge each other's claims, potentially exposing falsehoods to the adjudicator. A central premise of AI safety via debate is that truthful arguments are easier to defend than false ones under adversarial scrutiny. In this work, we investigate whether this advantage persists when debaters use rhetorical strategies that exploit biases in human judgment. Inspired by competitive debate, we construct 68 LLM-generated dialogues about detective mysteries with known culprits, spanning four interventions: anchoring, fallacy oversight, pro-jargon, and verbosity. We apply each intervention to either the side advocating for the true culprit or the side advocating for an innocent suspect, allowing us to distinguish influence on adjudication from correctness. In a study with 369 participants, we find that, pooled across bias types, these interventions significantly shift judgments toward the manipulated side. These findings expose a vulnerability in debate-based supervision: human adjudication is sensitive to manipulative rhetorical strategies.

Bias has long been studied in social psychology and cognitive science, where decades of research have produced a body of validated interventions that reduce stereotypical thinking and prejudiced responses in humans. We propose Debias It Yourself (DIY), a cognitively grounded framework that translates five such interventions into debiasing procedures for large language models and delivers them through three established paradigms: Show (in-context examples), Train (instruction tuning), and Revise (guided self-revision). Across three models, five bias benchmarks, eleven debiasing baselines, and three reasoning benchmarks, Train+Revise and Revise alone attain the top two average ranks, lead the bias-reasoning tradeoff (mean bias as low as 2% at 90% reasoning accuracy), and reduce bias on unseen dimensions by up to 14.8%. Our code and data are publicly available.

Researchers increasingly use human-derived bias constructs to study Large Language Models (LLMs), including social-cognitive constructs such as implicit bias and stereotype activation, and cognitive biases such as anchoring, framing effects, and confirmation bias. Such approaches offer alternatives to overt bias probes, particularly when direct questioning may obscure bias or when model behaviour appears normatively acceptable. However, adapting human bias constructs to LLMs introduces an inferential gap. Psychological instruments were developed to study human cognition and social behaviour, whereas LLM evaluations rely on probabilities, text completions, rankings, or simulated decisions. This paper critiques human-centered bias evaluation in LLMs. We show how this gap arises from mismatches pertaining to human-derived constructs, human-model differences, and evaluation contexts, which can blur distinct interpretations of model bias. We then introduce a framework providing an analytical lens for relating these elements to warranted interpretations, with attention to target constructs, operationalizations, scope of inference, and limits of human analogy.

Large language models are increasingly used as decision aids whose probability judgments shape downstream choices. Whether those judgments carry a systematic directional tilt has been hard to detect: calibration metrics aggregate unsigned errors, and naturalistic uncertainty offers no ground-truth probability. When an LLM rates a startup's success at 70% but its failure at 15%, the missing 15 points expose a distortion no aggregate score flags. We introduce OptimismBench, which detects directional bias with inverted pairs: each scenario elicits both P(success) and P(failure), and asymmetry between the two framings yields a signed bias score without ground truth. Across 16 models from 8 providers, fourteen are optimistic; pessimism appears only in Anthropic's frontier tier. Eleven matched base-versus-chat pairs across four families show post-training sets the sign of the bias, with opposite shifts in different families. The pattern survives prompt, temperature, perspective, and self-debiasing ablations. A seventeen-model six-language comparison further shows model identity dominates language, with inter-model variance at 4.7x inter-language variance. We release 3,870 items across 10 languages for per-model directional-bias auditing. When alignment makes a model more helpful, it also tilts its probabilities; downstream pipelines inherit the tilt by default.

Large Language Models (LLMs) have been widely applied in high-stakes decision-making scenarios such as corporate strategy, and users are increasingly relying on their outputs. However, the deep integration of open-source model sharing ecosystems with LLM-powered critical decision-making applications also introduces critical risks: if an attacker can manipulate the model's cognitive stance, they can indirectly influence the judgments and actions of downstream decision-makers. This paper defines such threats as decision-level hijacking. Existing attacks fail to achieve targeted cognitive manipulation without triggering prohibited content or degrading model functionality. To fill this gap, this paper reveals that Bit-Flip Attacks (BFAs) can serve as an attack vector for inducing decision-level hijacking, requiring no real-time interaction or control over the training process, and only a minimal number of weight bits need to be flipped after deployment to achieve stealthy, low-cost, and persistent cognitive manipulation. Therefore, we propose CogBias, a cognitive bias injection framework for LLMs. CogBias converts subjective preferences into optimization signals via a differentiable sentiment evaluator, uses a multi-objective loss to jointly constrain multiple dimensions, and constructs BitScout to locate critical bits, achieving targeted cognitive intervention under an ultra-sparse flip budget. Experiments on Llama-3.2-3B, Mistral-7B, and Qwen2.5-14B, as well as on the commercial recommendation and controversial factual topic scenarios, demonstrate that flipping only a small number of bits stably induces significant stance shifts on target topics, while the impact on non-target tasks and overall output distribution is limited. This work demonstrates that minute perturbations to low-level weight data suffice to undermine the high-level value alignment of LLMs.

Large language models (LLMs) increasingly issue judgments read as binary verdicts, and a growing literature reports such judgments shifting under logically irrelevant changes of wording - among them an amplified yes-no bias on moral dilemmas, absent in humans. A single framing cannot say what such a shift is: in a yes/no question the word "no" is at once logical verdict, lexical token, and last-printed option. We introduce a psychometric battery that separates these: crossed symmetrization - every logically irrelevant factor flipped in balanced pairs - across a corpus of question forms. A graded rating across logically equivalent forms recovers a coherent internal moral scale: frontier models' stance $θ$ is nearly format-invariant (cross-form incoherence 0.12-0.21 on a $\pm 1$ axis); small open-weight models fail in model-specific ways. Forcing the verdict through yes/no overlays a decomposable artifact: an order bias toward the last-printed option - opposite to classic human primacy - plus a lexical pull toward the word "no"; the artifact is substantial only in the Claude models (story-averaged -0.32 to -0.86), $\approx 0$ for GPT-5.5 and Gemini, and shrinks under extended reasoning. The word and the verdict share one token; swapping the words for arbitrary labels separates them, and the verdict-attached logical bias proves $\approx 0$ for every frontier model, while model-specific label and order attachments remain: the models are not drawn toward rejecting - the pull follows the printed surface, not the verdict it carries. A minimal model, $P = σ((θ\pm m)/s)$, summarizes any such artifact by a framing susceptibility m and a moral decisiveness s, measurably distinct from sampling temperature. The battery applies unchanged to any dilemma set and binary format: measuring what a model values requires crossing the frames of the question, not asking once.

We investigate the extent to which the language processing of LLMs resembles human cognitive processes, focusing on a human cognitive bias called the $\textit{neglect-zero effect}$. This effect refers to the human tendency to ignore $\textit{zero-models}$, which are configurations that render a proposition vacuously true by virtue of an empty set. We focus on two types of inferences driven by the neglect-zero effect, and examine how LLMs process these inferences by comparing their behavior with that in an inference that does not involve the neglect-zero effect. For this purpose, we employ a paradigm based on $\textit{structural priming}$, where recent exposure to a preceding sentence (the $\textit{prime}$) facilitates the processing of a subsequent sentence (the $\textit{target}$) due to their structural similarity. We prepare primes to force LLMs to consider the zero-model, and analyze whether they also consider it in the target. The results suggest that the neglect-zero effect may not occur in the LLMs analyzed in this study. Our code is available at https://github.com/ynklab/neglect_zero

Prompt-induced cognitive biases are changes in a general-purpose AI (GPAI) system's decisions caused solely by biased wording in the input (e.g., framing, anchors), not task logic. In software engineering (SE) decision support (where problem statements and requirements are natural language) small phrasing shifts (e.g., popularity hints or outcome reveals) can push GPAI models toward suboptimal decisions. We study this with PROBE-SWE, a dynamic benchmark for SE that pairs biased and unbiased versions of the same SE dilemmas, controls for logic and difficulty, and targets eight SE-relevant biases (anchoring, availability, bandwagon, confirmation, framing, hindsight, hyperbolic discounting, overconfidence). We ask whether prompt engineering mitigates bias sensitivity in practice, focusing on actionable techniques that practitioners can apply off-the-shelf in real environments. Testing common strategies (e.g., chain-of-thought, self-debiasing) on cost-effective GPAI systems, we find no statistically significant reductions in bias sensitivity on a per-bias basis. We then adopt a Prolog-style view of the reasoning process: solving SE dilemmas requires making explicit any background axioms and inference assumptions (i.e., SE best practices) that are usually implicit in the prompt. So, we hypothesize that bias-inducing features short-circuit assumption elicitation, pushing GPAI models toward biased shortcuts. Building on this, we introduce an end-to-end method that elicits best practices and injects axiomatic reasoning cues into the prompt before answering, reducing overall bias sensitivity by 51% on average (p < .001). Finally, we report a thematic analysis that surfaces linguistic patterns associated with heightened bias sensitivity, clarifying when GPAI use is less advisable for SE decision support and where to focus future countermeasures.

Large Language Models (LLMs) are increasingly deployed in high-stakes decision-making contexts. While prior work has shown that LLMs exhibit cognitive biases behaviorally, whether these biases correspond to identifiable internal representations and can be mitigated through targeted intervention remains an open question. We define LLM cognitive bias as systematic, reproducible deviations from correct answers in tasks with computable ground-truth baselines, and introduce LLM CogBias, a benchmark organized around four families of cognitive biases: Judgment, Information Processing, Social, and Response. We evaluate three LLMs and find that cognitive biases emerge systematically across all four families, with magnitudes and debiasing responses that are strongly family-dependent: prompt-level debiasing substantially reduces Response biases but backfires for Judgment biases. Using linear probes under a contrastive design, we show that these biases are encoded as linearly separable directions in model activation space. Finally, we apply activation steering to modulate biased behavior, achieving 26--32% reduction in bias score (fraction of biased responses) while preserving downstream capability on 25 benchmarks (Llama: negligible degradation; Qwen: up to $-$19.0pp for Judgment biases). Despite near-orthogonal bias representations across models (mean cosine similarity 0.01), steering reduces bias at similar rates across architectures ($r(246)$=.621, $p$<.001), suggesting shared functional organization.

We investigate whether large language models (LLMs) display human-like cognitive biases, focusing on potential implications for assistance in judicial sentencing, a decision-making system where fairness is paramount. Two of the most relevant biases were chosen: the virtuous victim effect (VVE), with emphasis given to its reduction when adjacent consent is present, and prestige-based halo effects (occupation, company, and credentials). Using vignettes that were altered from prior literature to avoid LLMs recalling from their training data, we isolate each manipulation by holding all other details consistent, then measuring the percentage difference in outcomes. Five models were evaluated as representative LLMs in independent multi-run trials per condition (ChatGPT 5 Instant, ChatGPT 5 Thinking, DeepSeek V3.1, Claude Sonnet 4, Gemini 2.5 Flash). Our research discovers that there is larger VVE, there is no statistically significant penalty for adjacent-consent, and the halo effect is slightly reduced when compared to humans, with an exception for credential based prestige, which had a large reduction. Despite the variation across different models and outputs restricting current judicial usage, there were modest improvements compared to human benchmarks.