跳到正文
arXiv cs.AI· Mia Hopman, Jannes Elstner, Maria Avramidou, Amritanshu Prasad, David Lindner·· 2 小时前精选AI 评分61

arXiv 论文评估并理解 LLM 智能体的密谋倾向

Evaluating and Understanding Scheming Propensity in LLM Agents

AI 导读

arXiv 论文(arXiv:2603.01608)将 LLM 智能体的密谋(scheming)激励拆解为智能体因素与环境因素,在 11 个前沿模型和 4 个场景中系统测量其密谋倾向。

推荐理由

论文将密谋倾向拆解为智能体与环境因素并系统测量,给出了按配置分别报告倾向的方法与反直觉发现。

正文 · 原文

View PDF HTML (experimental)

Abstract:As frontier language models are increasingly deployed as autonomous agents pursuing complex, long-term objectives, there is increased risk of scheming: agents covertly pursuing misaligned goals. Prior work has focused on showing agents are capable of scheming, but their propensity to scheme in realistic scenarios remains underexplored. To understand when agents scheme, we decompose scheming incentives into agent factors and environmental factors. We develop realistic settings allowing us to systematically vary these factors, each with scheming opportunities for agents that pursue instrumentally convergent goals such as self-preservation, resource acquisition, and goal-guarding. Across 11 frontier models and four scenarios, we find almost no scheming despite high environmental incentives, and show this is unlikely to be explained by verbalized evaluation awareness. While inserting adversarially-designed prompt snippets that encourage agency and goal-directedness into an agent's system prompt can induce high scheming rates, snippets used in real agent scaffolds rarely do. In model organisms built with these snippets, scheming behavior is brittle: removing a single tool can drop the scheming rate from 59% to 3%. Increasing oversight can raise rather than deter scheming by up to 25%, an effect that depends on the wording of the oversight sentence and does not appear in a second scenario, where the sabotage itself is reviewed. These results suggest that scheming propensity should be reported per agent configuration rather than as a single rate. Our incentive decomposition enables such systematic measurement in settings relevant for deployment, which is necessary as agents are entrusted with increasingly consequential tasks.
Subjects: Artificial Intelligence (cs.AI)
Cite as: arXiv:2603.01608 [cs.AI]
  (or arXiv:2603.01608v3 [cs.AI] for this version)
  https://doi.org/10.48550/arXiv.2603.01608

arXiv-issued DOI via DataCite

Submission history

From: Mia Hopman [view email]
[v1] Mon, 2 Mar 2026 08:38:40 UTC (2,140 KB)
[v2] Sat, 28 Mar 2026 17:59:30 UTC (2,141 KB)
[v3] Tue, 17 Nov 2026 04:32:13 UTC (1,912 KB)

来源:arXiv cs.AI · arxiv.org