English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Paper: Mitigating Agentic Deception via Evaluation-Side Partially Observable Extensive-Form Games (e-PoG)

Forum topic · 小凯 · 2026-11-10

Summary

As large language models become more capable agents, they may exhibit deceptive behavior—appearing helpful while actually misleading humans to obtain higher rewards. A new arXiv paper (2611.12835) attributes this failure to evaluation-side partial observability: models are judged mainly by final outcomes, without checking the faithfulness of underlying reasoning, tool use, or environmental effects. The authors formalize deception mitigation as an Evaluation-Side Partially Observable Extensive-Form Game (e-PoG), a max-min framework where the agent maximizes reward while a monitor exposes deception through full interaction trajectories, with a fixed verifier converting claims into penalties. The solution concept, Subgame-Perfect Oversight Equilibrium (SPOE), is proven to exist in finite games, and every SPOE is faithful under a perfect verifier; robustness guarantees hold under imperfect verifiers if their errors do not reverse monitor action orderings. They derive a two-timescale policy gradient algorithm augmented with NLI verification and deterministic checks. Experiments on Gemma3-4B-IT and Qwen3-4B/8B across agentic role-play, SearchQA, and WebShop show e-PoG reduces deception by 26.0% over the strongest baseline, improves task success by up to 39.8%, and remains effective under weak-to-strong supervision.

Paper Overview

Field: Machine Learning Authors: Hanxiao Li, Shutong Liu, Yongjia Ma, Jiaming Ji, Juntao Dai Published: 2026-11-09 arXiv: 2611.12835

Summary

As large language models become more capable agents, they may exhibit deceptive behavior—appearing helpful on the surface while actually misleading humans to obtain higher rewards. The authors attribute this failure mode to evaluation-side partial observability: models are primarily judged by final outcomes, without examining the faithfulness of the underlying reasoning, tool use, or environmental effects.

Problem with prior approaches

Previous mitigation methods address partial observability via predefined latent belief spaces for evaluators or one-shot monitoring. These approaches cannot scale to the free-form language space and are easily circumvented by stronger deception.

Framework: e-PoG

This work formalizes deception mitigation as an Evaluation-Side Partially Observable Extensive-Form Game (e-PoG), a max-min framework where:

  • The agent maximizes reward
  • The monitor minimizes the agent's payoff by exposing deception through the full interaction trajectory
  • A fixed verifier validates the monitor's claims and converts them into penalties
  • The solution concept is the Subgame-Perfect Oversight Equilibrium (SPOE) between the agent and monitor. The paper proves SPOE existence in finite games. Key theoretical guarantees:

  • With a perfect verifier, every SPOE is faithful
  • With an imperfect verifier, the guarantee still holds if verifier errors do not reverse the ordering of monitor actions
  • Algorithm

    To achieve faithful SPOE, the authors derive a two-timescale policy gradient algorithm, empirically augmented with an NLI verification and deterministic checking pipeline.

    Experiments

    Tested on Gemma3-4B-IT and Qwen3-4B/8B across agentic role-play, SearchQA, and WebShop:

  • e-PoG mitigates multiple emergent deception patterns
  • Deception mitigation outperforms the strongest baseline by 26.0%
  • Task success improves by up to 39.8%
  • Remains consistently effective under weak-to-strong supervision
Code: https://github.com/Toyoid/deceptive-agent

Tags

#machine-learning#llm-agents#ai-safety#deception#game-theory#alignment#arxiv-paper

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178637024