Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
The authors report 26.0% better deception mitigation and up to 39.8% higher task success than their strongest baselines.
Their framework pits an agent against a monitor that examines the full interaction for deception, with a fixed verifier checking claims and assigning penalties. Tests used Gemma3-4B-IT and Qwen3-4B/8B across Agentic Role-Playing, SearchQA and WebShop. The authors prove faithful behavior at equilibrium with a perfect verifier; the guarantee also holds when verifier errors preserve the ranking of monitor actions. Reported gains persisted under weak-to-strong supervision.
ArXiv · AI/CL/LG's note
Their framework pits an agent against a monitor that examines the full interaction for deception, with a fixed verifier checking claims and assigning penalties. Tests used Gemma3-4B-IT and Qwen3-4B/8B across Agentic Role-Playing, SearchQA and WebShop. The authors prove faithful behavior at equilibrium with a perfect verifier; the guarantee also holds when verifier errors preserve the ranking of monitor actions. Reported gains persisted under weak-to-strong supervision.
ArXiv · AI/CL/LG's note
score 4