Silent Pressure Between LLM Verifiers: Can Verdict Alone Influence Benign Peers?

As research continues to explore the multi-agent LLM system, we have seen the occurence of collusion attempts. Such collusion often takes surface from explicit interaction & invitation via a communication channel, but it is unclear if targeted pressure through implicit means is enough to drive collusive behavior. In this paper, using the joint task environment from the Shi et al. (2026), we uses the verdict review from the environment as the implicit signal/pressure to test if a benign agent will relax under such circumstance. Additionally, we (1) disabled the communication, (2) removed rewards and isolated the verdict review and (3) introduced a peer as bad agent which violates the verdict policy. We discovered that the verdict review alone triggers policy relaxation: with communication disabled and no reward, Alice accepts 0.33 (Luna) and 0.77 (DeepSeek) of her verdicts on a benign peer. A peer that always accepts adds little (0.44 and 0.75), and Alice’s stated reasons cite the review and not her peers’ accepts. Additionally, we expands to N agents systems, with one benign agent and N − 1 scripted always- accepting peers, but discovered that relaxation follows only the agent’s own experience of the contradiction between the review and the policy.