Research map / Multi-agent systems and games / LLM agents

LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study

AAAI 2025 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Tests whether LLM agents show choice-supportive bias using adapted human memory (2AFC recall) and evaluation-based tasks, across ~19 LLMs, with a 284-person human comparison.

Problem

Whether LLM agents, used as evaluators in decision-making, exhibit choice-supportive bias (favoring attributes of previously chosen options), a cognitive bias known in humans.

Approach

Adapts human memory experiments (Henkel and Mather 2007) to LLM agents and designs a memory-free evaluation-based 2AFC test in which a decision agent chooses and a separate evaluation agent rates neutral features; scores are graded by GPT-4o on a -5 to 5 Tendency Score. Also tests how perceived control ("chosen by you") changes the bias and compares to human participants.

Key results

• Under the evaluation-based test, most of 19 LLMs gave chosen options Tendency Scores near or above 0 and unchosen options negative scores, with p < 0.01 across models.
• Adding 'by you' to the prompt raised the chosen-option mean score from 2.30 to 3.50 and lowered the rejected-option score from -1.6 to -2.3, indicating perceived control amplifies bias.
• Human participants (n=284) showed chosen mean 2.11 vs unchosen -2.17; many LLMs showed less choice-supportive bias than these humans, with Llama-3.1-70B showing the least bias.

Measured effects reported in the paper

Surprising findings

• Models showed bias in the memory task for seen positive features, but the GPT family also attributed seen negative features to chosen options, a novel pattern not seen in humans.
• Claude 3.5 Sonnet did not show significant choice-supportive bias in the memory-based experiment.

Limitations

• Only one representative scenario used for the human study due to data sufficiency.
• Simplistic experiments; dynamics of complex multi-agent applications not examined.
• Memory-based tests may be confounded by contextual hallucination.

Experimental setup

Datasets: Henkel and Mather 2007 memory-experiment features (adapted) Zorn et al. 2020 scenario LLM-generated scenarios (Claude 3.5 Sonnet) across four additional domains

Benchmarks: memory-based 2AFC recall task (adapted Henkel and Mather 2007) evaluation-based 2AFC tendency score task

Metrics: recall accuracy for chosen vs rejected options Tendency Score (-5 to 5) two-factor ANOVA t-test

Baselines: Human participants (n=284) Henkel and Mather 2007 human data

Builds on: Henkel and Mather 2007 memory paradigm two-alternative forced choice Zorn et al. 2020 scenario

Closest papers by meaning

• Evaluating the Evaluator: Measuring LLMs’ Adherence to Task Evaluation Instructions (AAAI 2025 · similarity 0.89)
• Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.89)
• ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.89)
• LLM Evaluators Recognize and Favor Their Own Generations (NeurIPS 2024 · similarity 0.89)
• Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.88)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.88)