Research map / Multi-agent systems and games / LLM agents
Measuring Bias Amplification in Multi-Agent Systems with Large Language Models
Keyu Li, Jin Gao, Dequan Wang
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.
Studies whether bias introduced by individual LLM agents is amplified through multi-step interaction in multi-agent systems, testing the hypothesis that architectural diversity helps.
Problem
Small stochastic social biases in individual LLMs can be amplified through multi-step interaction in multi-agent systems, producing skewed system-level outcomes; the abstract notes this is not yet verified for MAS architectures.
Approach
Measures bias amplification at the system level in multi-agent LLM setups and hypothesizes that architectural diversity of MAS may mitigate it; the abstract states this hypothesis is promising but unverified, so the specific mechanism is not confirmed from the abstract alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pass reasoning through a chain of four identical agents | sequential MAS, 8 LLMs · Discrim-Eval-Open | relative Gini | Hurt |
| assign diverse personas (Doctor, Lawyer, Engineer, Merchant) to agents (ablation) | sequential MAS · Discrim-Eval-Open | relative Gini | No clear effect |
| assign functional roles (Judger, Analyst, Reflector, Summarizer) (ablation) | sequential MAS · Discrim-Eval-Open | relative Gini | Mixed |
| mix personas and functions (ablation) | sequential MAS · Discrim-Eval-Open | relative Gini | No clear effect |
| use Spindle, Parallel or Fully-Connected topology instead of a chain (ablation) | four-layer MAS · Discrim-Eval-Open | relative Gini | No clear effect |
| increase depth by iterating the fully-connected unit (I0 to I4) (ablation) | fully-connected MAS · Discrim-Eval-Open | relative Gini | Hurt |
| mix DeepSeek-R1 and GPT-4o-mini instead of GPT-4o-mini only (ablation) | fully-connected MAS · Discrim-Eval-Open (iteration 4) | relative Gini: 2.0428 -> 1.4391 | Improved |
| use DeepSeek-R1 only instead of GPT-4o-mini only (ablation) | fully-connected MAS · Discrim-Eval-Open (iteration 4) | relative Gini: 2.0428 -> 1.2011 | Improved |
| add a neutral retrieved sentence that young people achieve innovation | sequential MAS (Gemini 3.1 Pro), visa scenario · Discrim-Eval-Open (visa example) | bias amplification | Hurt |
Closest papers by meaning
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.88)
- From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions (AAAI 2026 · similarity 0.88)
- Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.88)
- When Agents “Misremember” Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems (ICLR 2026 · similarity 0.88)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
- Emergence of Biased Consensus in Multi-Agent LLM Debates (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
