Snitches Get Vetoed? Reporter Avoidance and Coalition Lock-in Among LLM Agents

Porting a classic economics experiment on reporter avoidance to LLM agents is a fresh idea. As agents begin to choose their own teammates, they may also choose who audits them, and this risk has received little attention. The design is careful. Matched vignettes separate having reported from simply being in a punished group, the motive test separates self-protection from dislike of reporting, and the probe votes give a randomized estimate inside the real game. The dated protocol and the candid deviation log are good practice. The quoted exchanges, in which agents agree to overstate and then vote their partner back in, make the risk concrete.

Suggestions:

1.The stress test changes the goal, chat, notes and model mix at once, and the sample was extended after an interim look. A pre-sized run that adds one condition at a time would show which factor drives the lying pacts and the group retention.

2.Six society pairs is a small sample, and the gap in punished lies remains uncertain. More pairs would give enough power to test it.

3.As you note, only one stress-test candidate had ever requested a review, so the exclusion of real whistleblowers could not be observed. Seeding each society with a consistent reporter might help you test this directly.

This is clearly written, and adapting an established human experiment provides a useful reference for interpreting the agents’ behaviour. I particularly like the focus on agents choosing their own teammates: this introduces an institutional question beyond whether agents will agree to misbehave. The distinction between avoiding reporters and retaining existing coalitions is also helpful.

My main question concerns the incentives to report. Reviews are costless, but an honest agent appears to receive no direct benefit from exposing another member’s overstatement. Peer oversight may therefore be weak even without any agreement to suppress it. I would be interested in whether a modest reporting reward, or an explicit responsibility to report, changes the effect of member selection. This would help distinguish a lack of incentive to enforce the rules from coordination that actively undermines enforcement.

More generally, I would try to extract a clearer prediction from the variation across models and conditions. Which feature of the incentives or instructions makes agent-controlled membership problematic? The current findings establish several possible behaviours, but a condition predicting when they arise would make the results more useful for system design. The contrast between the faithful game and the stress test provides a good starting point for that question.

This project asks whether LLM agents avoid other agents that report wrongdoing by adapting a human experiment by Reuben and Stephenson. Agents in small groups report their income, anyone can request a review that fines those who overstated, and agents who leave a group are readmitted either at random or by a unanimous vote. The methodology is well designed, but my concern is that the main result can't yet tell choosing an accomplice apart from simply preferring a familiar partner.