MALIBU Benchmark: Multi-Agent LLM Implicit Bias Uncovered

Abstract

We introduce MALIBU (Multi-Agent LLM Implicit Bias Uncovered), a benchmark for evaluating implicit biases in multi-agent LLM systems. MALIBU systematically probes how biases emerge, amplify, and propagate when multiple LLM agents interact in collaborative decision-making scenarios. Our framework reveals that multi-agent configurations can amplify individual model biases by 15-40% compared to single-agent baselines.

Citation