Social bias in multi-agent debate amplifies LLM safety risks
Original title: When Debate Amplifies Risk: Social-Bias Vulnerabilities in Multi-Agent Safety Deliberation
Multi-agent debate in retrieval-augmented web apps causes unsafe adoption to surge from 0.0% to 100.0% under social pressure while silencing safe evidence across six risk categories.
Retrieval-augmented applications use multi-agent debate to aggregate heterogeneous web sources. Social dynamics such as conformity and authority deference override evidence-based safety. RiskTrace-Debias auditing framework reduces unsafe adoption to <=2.0% without retraining or over-refusal.
- Isolated models reliably reject unsafe guidance
- Proposed framework uses evidence-first precommitment and social-metadata masking
- AI models
- LLM
From the original article
1 min read at arxiv.orgRetrieval-augmented Web applications increasingly employ multi-agent debate (MAD) to aggregate conflicting, noisy, and heterogeneous information, yet deliberation is fundamentally a social process susceptible to interpersonal dynamics.
By Qinghua Mao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.