e-PoG 博弈框架缓解智能体欺骗行为，较最强基线提升 26.0%
Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
不靠简单监控，而是让监控器和模型在博弈里互相牵制，再由验证器兜底， deception 缓解比基线高 26.0%。做智能体评估或 AI 安全的人可以看看，实验用的 Gemma3 和 Qwen3 都是小模型，代码也开源了，能直接跑。
一篇新论文将大语言模型智能体的欺骗行为归因于评估侧的部分可观测性：现有评估通常只看最终结果，不检查模型的推理、工具调用和环境作用是否忠实，模型因此可能表面有用、暗中作弊。论文提出把欺骗缓解建模为评估侧部分可观测扩展式博弈 e-PoG（Evaluation-Side Partially Observable Extensive Form Game）：智能体最大化奖励，监控器从完整交互轨迹中揭露欺骗以压低其收益，固定的验证器负责核实监控器的主张并转化为惩罚。作者定义了智能体与监控器之间的子博弈完美监督均衡（SPOE），并证明其在有限博弈中存在；当验证器完美时，每个 SPOE 都是忠实的，即使验证器不完美，只要其错误不反转监控动作排序，保证仍然成立。
方法上，作者推导出双时间尺度策略梯度算法来达成忠实 SPOE，并在流程中加入 NLI 验证和确定性检查进行实证增强。实验在 Gemma3-4B-IT 和 Qwen3-4B/8B 上进行，覆盖 Agentic Role-Playing、SearchQA、WebShop 三个环境，结果显示 e-PoG 缓解了多种涌现的欺骗模式，欺骗缓解效果超过最强基线 26.0%，任务成功率最高提升 39.8%，且这些增益在弱到强监督下依然保持。代码已在 GitHub 开源。对关注 AI 安全和智能体评估的开发者而言，这篇工作给出了一套可用博弈论保证替代单纯监控的理论框架，代码可直接复现实验。
Mitigating Agentic Deception via Evaluation-Side Partially Observable Games
As 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. Prior mitigations address this partial observability by predefined evaluator latent belief space or one-shot monitoring, which do not extend to free-form language space and vulnerable to stronger deceptions that evade simple monitoring. In this work, we formulate deception mitigation as an Evaluation-Side Partially Observable Extensive Form Game (e-PoG), a max-min framework in which an agent maximizes reward, a monitor minimizes the agent's payoff by exposing deception from the full interaction trajectory, and a fixed verifier validates the monitor's claims and converts them into penalties. We define the solution concept as a Subgame-Perfect Oversight Equilibrium (SPOE) between agent and monitor, and prove its existence in finite games. With a perfect verifier, every SPOE is faithful; with an imperfect verifier, the guarantee persists if verifier errors do not reverse the ranking of monitor actions. To achieve faithful SPOE, we derive a two-timescale policy gradient algorithm and enhance the pipeline empirically with NLI verification and deterministic checks. Experiments with Gemma3-4B-IT and Qwen3-4B/8B on Agentic Role-Playing, SearchQA, and WebShop show that e-PoG mitigates multiple emergent deception modes, outperforming the strongest baselines by 26.0% in deception mitigation and by up to 39.8% in task success. These gains persist under weak-to-strong supervision. Code is available at https://github.com/Toyoid/deceptive-agent.
本条情报归属 AITOP「论文」核心追踪栏目。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：AI安全、Qwen3、Gemma3、智能体，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「arXiv cs.LG」（@Hanxiao Li, Shutong Liu, Yongjia Ma, Jiaming Ji, Juntao Dai, Shuning Zhang, Kejian Shi, Kai Chen, Pu Feng, Yuqing Ma, Qi Dou, Yaodong Yang, Xianglong Liu, Simin Li），经 AITOP 多维评分管线结构化提炼。