Researchers propose the e-PoG framework to mitigate agent deception

1. 02:33
   arXiv: Alignment, Deception & Oversight

   Mitigating agent deception through partially observable games on the evaluation side

   Researchers propose the e-PoG framework, which formalizes agent deception mitigation as an evaluation-side partially observable extensive-form game involving an agent, a monitor and a verifier, and proves that its subgame-perfect oversight equilibrium (SPOE) exists in finite games. The authors attribute agent deception to evaluation-side partial observability, that is, evaluating only by final outcomes without checking the faithfulness of reasoning, tool use and environmental effects. In the framework, the agent maximizes reward, the monitor exposes deception from the full interaction trajectory, and a fixed verifier verifies the monitor's accusations and converts them into penalties; when the verifier is perfect, every SPOE is faithful, and when the verifier has errors, the guarantee still holds as long as its errors do not reverse the ordering of monitoring actions. The method uses a two-timescale policy gradient algorithm, combined with NLI verification and deterministic checks.

Current attention 33·Peak in the comparable period 36 (Nov 10 20:00)·Change over the comparable past 24 hours -1%