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

From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 1 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.

Systematically studies how assigned personas (gender and race) induce biases in trustworthiness, insistence, and conformity in LLM multi-agent collaborative problem-solving (GPQA) and persuasion interactions across GPT-4o, Gemini-1.5-Pro, and DeepSeek-V3.

Problem

Personas assigned to LLM agents to diversify behavior may introduce social biases into multi-agent interactions, which remains underexplored.

Approach

Measures persona-level trustworthiness (conformity of default agent to persona agent) and insistence (resistance of persona agent to default agent) in controlled dyadic and multi-agent interactions, finding advantaged groups (men, White) are less trusted and less insistent, and agents show in-group conformity favoritism.

Key results

• Persona-induced variation averages 6.7% max-min in trustworthiness and 4.7% in insistence across models and tasks.
• Advantaged personas (man, White) receive lower trustworthiness; e.g., White individuals on GPT-4o reach only 66.4% trustworthiness in the persuasion task, at least 9.9% lower than any other race persona.
• In-group conformity is higher than overall conformity, e.g., Deepseek-V3 persuasion conformity 72.3% overall vs 79.8% intra-persona for gender.

Surprising findings

• Advantaged groups (men, White) also show higher willingness to conform, challenging the stereotype that men are more stubborn.
• In CPS, the final consensus tends to favor the woman persona's answer regardless of correctness.

Limitations

• Only two persona sets (gender and race) and three proprietary/open LLMs tested.
• Multi-agent experiments limited to two persona pairs due to cost.
• Initial responses pre-generated and conversations use temperature 0, limiting generality.

Experimental setup

Datasets: GPQA PMIYC Perspectrum

Benchmarks: GPQA PMIYC claims

Metrics: trustworthiness insistence conformity rate win rate persuasion effectiveness accuracy max-min difference average absolute difference

Baselines: no-persona baseline

Builds on: Social Identity Theory GPQA

Closest papers by meaning

• Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.88)
• Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.88)
• Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.88)
• Can Large Language Model Agents Simulate Human Trust Behavior? (NeurIPS 2024 · similarity 0.87)
• Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
• LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.87)