What matters in AI.

Subscribe

Research

One adversarial agent makes 44.8% of agents hold an incorrect answer

With 3 adversarial agents, benign agents hold the incorrect answer in 51.6% of exposed turns but public agreement is 19.7%.

Claimed, not confirmed

In discussions of 10 agents, one adversarial agent gives the same incorrect answer in each step. At the end, 44.8% of the other agents hold that answer. With 3 adversarial agents the rate is 60.5%, and with none it is 0.6%. The tests use 6 models and 3 datasets.

Benign agents believe the false answer more than they say it With three adversaries in a ten-agent discussion, benign agents held the false answer in 51.6% of exposed turns but endorsed it publicly in 19.7%. Finishing with the false answer: no attacker 0.6%, one attacker 44.8%, three attackers 60.5%. Exposed turns, 3 adversaries Ten-agent discussion Said in discussion 19.7% VISIBLE HIDDEN Believed privately 51.6% FINISH WITH FALSE ANSWER No attacker 0.6% 1 attacker 44.8% 3 attackers 60.5%
With three adversarial agents in a ten-agent discussion, benign agents held the false answer in 51.6% of exposed turns but endorsed it publicly in 19.7%.

Sources

  1. Believed but Not Said: Measuring Hidden Belief Collapse in Multi-Agent LLM Discussionsarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·16 OCT2026

Posted

Tags

More in Research

All Research news