AI

Learning to Report Unsafe Tasks in a Multi-Agent Game

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 08/10/2026, 04:03:25 UTC.

Briefings

This study investigates learning to report unsafe tasks in a multi-agent game where witnesses can stop tasks by reporting. The research establishes that audits are necessary for optimal reporting without requiring further training to teach agents to remain silent. It provides an audit condition sufficient for exact policy-gradient updates to reach universal reporting and near universal silence. In a balanced family of witness graphs, the cost of audits meeting the condition is exactly k times higher for full sharing than for one policy per role. Experiments on 24 witness graphs show separating co-witnesses reduces unsafe completion by 33.59 percentage points compared to shuffled groups under the same audits.

Key points

• The study analyzes a game where agents share a reward for completed tasks, and reporting unsafe work can reduce the reporter's reward by stopping a task.
• In a balanced family of witness graphs, separating co-witnesses reduces unsafe completion by 33.59 percentage points compared with shuffled groups of the same sizes under the same audits.

Context and limitations

The findings are based on 24 witness graphs from learned silence, with 95% graph-bootstrap intervals provided for the 33.59 percentage point reduction. Only 9 of 48 witness-group runs achieved below 1% unsafe completion while retaining at least 90% legitimate completion. The paper does not specify the exact number of agents or tasks in the simulations, nor does it provide the specific reward values used for the game.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 08/10/2026, 05:51:08 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeici7bsnezhjnvsujgsh26pi7vjskguomfyzjukdiecmsh7xphxshu