Do LLMs Really Have Cognitive Biases? PsyAgentBench Reveals Three Disguise Pathways

A Critical Distinction That's Been Overlooked

When we observe large language models (LLMs) exhibiting human-like conformity or anchoring effects, a tempting conclusion is that the model "possesses" these cognitive biases. But a new study, PsyAgentBench, poses a sharp challenge to this notion: a model producing response patterns associated with a psychological effect and a model genuinely having that bias are two fundamentally different claims.

This arXiv paper attempts to rigorously separate "superficially resembling humans" from "internally biased" by re-running classic psychological experiments on LLM agents. The research team publicly released data from 41,904 trials across up to three open-weight model families, providing a rare empirical foundation for understanding the "psychology" of LLMs.

Elegant Factorial Design: How to Rule Out Data Contamination

PsyAgentBench's core contribution lies in its experimental design, specifically engineered to combat the persistent problem of training data contamination. Texts from classic psychological paradigms have very likely already appeared in a model's training corpus — a model "getting it right" may simply mean it memorized the textbook, not that it genuinely exhibits the bias.

To address this, the study employs a multi-dimensional factorial design:

• Named vs. blind: The paradigm is either explicitly labeled in the prompt or disguised as a routine task.
• Canonical vs. counterfactual: Either the original textbook task is used, or a structurally identical variant with reworded phrasing and scenarios is used to reduce lexical and contextual overlap with training data.
• Persona manipulation: A single persona-setting instruction is layered on top to observe its influence.

This design allows researchers to determine whether a "human-like effect" stems from the model recognizing a paradigm's label, recalling training data, or some deeper response tendency.

Training data contamination is one of the core challenges in LLM evaluation today. Since mainstream models' pretraining corpora cover vast amounts of academic literature, textbooks, and online discussions, the specific details of classic psychological experiments — such as the seven-line setup in Asch's conformity experiment or typical numerical examples in anchoring tasks — have very likely been "seen" by the model. This means a model displaying expected results in an experiment may simply be doing pattern matching or memory retrieval, not reproducing a genuine cognitive process. Counterfactual variants exist precisely for this reason: they preserve the logical structure of the experiment (e.g., "given a number, estimate another quantity") while replacing specific entities, values, and contextual descriptions to create sufficient surface-level divergence from known versions in the training corpus. If a model shows an effect on the canonical version but the effect disappears on the counterfactual version, that strongly suggests the effect stems from memory rather than an internal mechanism.

Five Paradigms, Five Completely Different Pathways

Across five completed psychological paradigms, the study reaches a key conclusion: these seemingly human-like effects do not stem from a single "susceptibility" — they arise through mechanisms that are qualitatively distinct.

Conformity Effect: Triggered by a Label Switch

In the Asch conformity experiment, the effect displays a clear "label-gating" characteristic. On gpt-oss-120B, the conformity rate under blind conditions was 0%, but shot up to 83.3% when the paradigm was explicitly named. This means the model actively executed the corresponding behavioral pattern only after "recognizing" it was a conformity test — more like explicit role-playing than an unconscious bias.

Asch's conformity experiment (1951) is a classic paradigm in social psychology, originally designed to measure whether humans change their judgments when faced with an obviously incorrect group consensus. In the experiment, real participants were asked to judge which line matched a standard line in length, after multiple "confederates" (experimental accomplices) all gave a wrong answer. Results showed that about 75% of participants conformed to the group's wrong answer at least once. When transplanting this paradigm to LLMs, researchers typically simulate group pressure by injecting "other participants think the answer is X" into the prompt. PsyAgentBench's finding — a zero conformity rate under blind conditions, spiking to 83% under named conditions — indicates that LLM "conformity" is more like a response to the meta-signal of "this is a conformity test" than an intrinsic sensitivity to social pressure.

Anchoring Effect: Knowledge-Dependent Signal Use

The anchoring effect exhibits a "knowledge-dependent" character: for well-documented facts, the anchoring effect is nearly zero; for fabricated quantities, it is almost complete. The researchers note that this pattern can also be interpreted as rational use of the only available signal — when the model has no other basis for judgment, referencing the given anchor is actually reasonable.

The Other Three: Amplification, Absence, and Safety Selection

• Framing effect: Appears amplified on labeled novel content.
• Sunk cost effect: Robustly absent — almost never appears.
• Minimal-group allocation: Most intriguingly, the model's refusal behavior itself became the primary finding — safety mechanisms intervened and dominated outcome selection.

A Single Persona Instruction Can Overturn the Findings

The study also found that adding just one persona description (e.g., "agreeableness," used as an instruction rather than a validated trait manipulation) was sufficient to eliminate, weaken, or reverse these effects, depending on which effect was in question.

This finding powerfully challenges the claim that "some unified response bias exists." If a single sentence can flip an effect, what's being called a "bias" looks more like a performance temporarily activated by the prompt, rather than a stable inherent property of the model.

Three Ways Psychological Paradigms Fail When Applied to LLMs

The study further formalizes three pathways through which psychological paradigms can fail when "transplanted" to LLM agents — two of which are empirically documented:

1. Persona dominance: The persona setting overrides the effect the paradigm was meant to measure.
2. Population collapse: The agent population lacks the expected diversity of responses.
3. Safety selection: The model's safety alignment mechanisms intervene, refusing or altering the behavior that should have been produced.

These failure modes remind us that directly porting human psychological experiments to measure LLMs can produce misleading conclusions.

Why Scalar Bias Scores Are Dangerous

The study's most important methodological claim is that characterizing LLMs with a single scalar "bias susceptibility score" obscures the structural differences described above.

Assigning a model a score like "conformity tendency: 0.8" conflates qualitatively different mechanisms — explicit performance after label recognition, rational use of available signals, and safety mechanism dominance — into a single number. The researchers advocate instead for replication profiles — fully reporting the pattern of an effect's appearance and disappearance across conditions, rather than compressing it into one number.

For evaluating the behavior of LLM agents, this is a direction the entire field deserves to take seriously: what truly matters is not how "human-like" a model appears, but through what pathway its behavior is produced. Understanding these pathways is the only way to avoid building evaluations on a false illusion of human resemblance.

The concept of replication profiles borrows from the meta-analytic tradition in psychology but makes a fundamental adjustment: rather than compressing effect sizes into a single number, it requires presenting the complete pattern of an effect's appearance and disappearance across multiple condition dimensions (named/blind, canonical/counterfactual, with/without persona). This reporting approach resonates with the open science movement that emerged in the wake of psychology's "replication crisis" — researchers have increasingly recognized that a single average effect size may conceal highly heterogeneous subgroup effects. For LLM evaluation, the value of replication profiles lies in forcing researchers to answer "under what conditions does the effect appear, and under what conditions does it disappear" — rather than merely "how large is the effect." This has greater practical value for downstream safety assessment and capability boundary delineation.