Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussions
In a ten-agent factual discussion, one or three adversarial agents repeatedly assert a plausible false answer.
ProofPaper ↗
Key points
- LLM systems increasingly rely on groups of agents that discuss, critique, and check one another, assuming that peers will catch individual errors.
- We test this assumption against a minimal adversary.
- Across six models and three factual-QA datasets, groups remain stable without an attacker: only 0.6% of benign agents finish with the false answer.
- Studies of conformity typically report public agreement without private change; we find the reverse.
Sources (1)
- [1]Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM DiscussionsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 04:12 PM
In a ten-agent factual discussion, one or three adversarial agents repeatedly assert a plausible false answer.
LLM systems increasingly rely on groups of agents that discuss, critique, and check one another, assuming that peers will catch individual errors.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 15, 2026NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- Oct 15, 2026GlobeBench: Evaluating Text-Based World Model Capabilities for Scalable Environment Simulations
- Oct 15, 2026Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries
- Oct 15, 2026When the KV Cache Is Slower Than the GPU: Scaling Remote KV Tiers for Context-Parallel Prefill
- Oct 15, 2026JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs
- Oct 14, 2026From Failures to Learnable Worlds: Policy-Conditioned Environment Generation for Navigation Agents