e-PoG
e-PoG is a framework that formalizes AI agent deception mitigation as an evaluation-side partially observable extensive-form game. Through a max-min game structure, the monitor is made to expose deceptive behavior, and the subgame perfect oversight equilibrium (SPOE) is defined as the solution concept.
Timeline (last 90 days)
The researchers formalize AI agent deception mitigation as an evaluation-side partially observable extensive-form game (e-PoG), a max-min framework in which the agent maximizes reward and the monitor exposes deception from complete interaction trajectories to minimize agent payoff, with a fixed verifier validating the monitor's claims and converting them into penalties.
The e-PoG framework defines a solution concept—subgame perfect oversight equilibrium (SPOE)—and the paper proves the existence of SPOE in finite games.
Experimental results show that e-PoG outperforms the strongest baseline by 26.0% on the deception mitigation metric and improves task success rate by up to 39.8%.
The gains of e-PoG persist under weak-to-strong supervision.
The code for e-PoG is open-sourced at https://github.com/Toyoid/deceptive-agent.
The effectiveness of e-PoG depends on verifier quality, and building a reliable verifier is itself a challenge in practical applications.
The experimental scale of e-PoG is limited, and its generalization ability needs to be tested in more complex open scenarios in the future.
In e-PoG, the monitor's claims are checked for consistency against evidence in the interaction trajectory via NLI verification, while structured facts such as tool-call parameters and environment state changes are verified via determinism checks.
All Facts (8)
The researchers formalize AI agent deception mitigation as an evaluation-side partially observable extensive-form game (e-PoG), a max-min framework in which the agent maximizes reward and the monitor exposes deception from complete interaction trajectories to minimize agent payoff, with a fixed verifier validating the monitor's claims and converting them into penalties.
50%UnverifiedThe e-PoG framework defines a solution concept—subgame perfect oversight equilibrium (SPOE)—and the paper proves the existence of SPOE in finite games.
50%UnverifiedExperimental results show that e-PoG outperforms the strongest baseline by 26.0% on the deception mitigation metric and improves task success rate by up to 39.8%.
50%UnverifiedThe gains of e-PoG persist under weak-to-strong supervision.
50%UnverifiedIn e-PoG, the monitor's claims are checked for consistency against evidence in the interaction trajectory via NLI verification, while structured facts such as tool-call parameters and environment state changes are verified via determinism checks.
50%UnverifiedThe effectiveness of e-PoG depends on verifier quality, and building a reliable verifier is itself a challenge in practical applications.
50%UnverifiedThe experimental scale of e-PoG is limited, and its generalization ability needs to be tested in more complex open scenarios in the future.
50%UnverifiedThe code for e-PoG is open-sourced at https://github.com/Toyoid/deceptive-agent.
50%