Reproduce “Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Liangjie Huang, Henry Peng Zou, Wei-Chieh Huang, Wooseong Yang, Yuwei Han, Philip S. Yu
The authors test whether multi-agent discussions contain hidden acceptance of false claims by comparing agents’ public messages with separate private answer probes. Across six models and three factual-question datasets, they find that under three attackers, benign agents give the false answer in 51.6% of exposed turns but publicly endorse it in 19.7%. The probes measure context-dependent answers, not internal beliefs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.