Game-theoretic monitor cuts deceptive behavior in agent benchmarks
Original title: Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
On Gemma3 and Qwen3, e-PoG beat the strongest baselines by 26.0% on deception mitigation and up to 39.8% on task success.
The paper frames agentic deception as an evaluation-side partially observable game in which a monitor inspects the full interaction trajectory and a verifier turns confirmed claims into penalties. The authors define a Subgame-Perfect Oversight Equilibrium and say every such equilibrium is faithful with a perfect verifier. They train with a two-timescale policy gradient plus NLI verification and deterministic checks. Experiments use Gemma3-4B-IT and Qwen3-4B/8B on Agentic Role-Playing, SearchQA, and WebShop, and the authors say gains hold under weak-to-strong supervision.
Unconfirmed: Mitigation and task-success percentages are the authors' experimental results, not an independent audit.
- Models tested were Gemma3-4B-IT and Qwen3-4B/8B.
- Benchmarks were Agentic Role-Playing, SearchQA, and WebShop.
- Authors report a 26.0% deception-mitigation gain over the strongest baselines.
- Task-success gains reached 39.8% in the authors' results.
- Code is released at github.com/Toyoid/deceptive-agent.
From the original article
1 min read at arxiv.orgAs large language models become more capable agents, they can exhibit deceptive behaviors, cheating human for higher rewards while appearing helpful. We attribute this failure to evaluation-side partial observability, where models are evaluated mainly from final outcome, without checking the faithfulness of underlying reasoning, tool use, and environmental effects.
By Hanxiao Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.