研究提出 introspective oversight:训练第三方观察者角色监督模型不当行为
Training Language Models for Introspective Oversight
AI 导读
arXiv 论文提出 introspective oversight,为模型新增独立的第三人称观察者角色,只按回答与助手实际行为的核对结果训练,用于报告奖励破解、欺骗、谄媚等难从转录中察觉的行为。
推荐理由
论文给出可训练的内省监督方法及五项 held-out 评测的 AUROC 对比,可用于了解可验证自我报告的监督路径。
正文 · 原文
Abstract:Language models increasingly exhibit behaviors that are difficult to detect in transcripts, such as reward hacking, deception, and sycophancy. Asking the assistant follow-up questions about its actions is unreliable, since models are generally not trained to report their own misbehavior. As a result, approaches to oversight typically involve querying external models. These face limitations: external transcript monitoring operates only on visible model outputs, and activation decoders lack full context and are prone to confabulation. We introduce introspective oversight, which trains a new role in the model to answer arbitrary questions about the assistant, rewarding answers that are verified against the assistant's actual behavior. We implement this by adding a separate, third-person observer role that answers these queries, which we can train solely for accuracy and separately from the assistant. On five held-out oversight evaluations, an observer trained on Qwen3.6-27B obtains a mean AUROC of 0.907, outperforming honesty training (0.864), transcript monitoring (0.828), and activation oracles (0.796). We further find that training an observer makes the assistant itself easier to monitor, with every baseline method improving on average when applied to the fine-tuned assistant. These results suggest that training models to answer verifiable questions about themselves is a tractable approach to overseeing unintended behavior.
| Comments: | 31 pages, 13 figures, 8 tables |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.23365 [cs.AI] |
| (or arXiv:2611.23365v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.23365 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Neil Chowdhury [view email]
[v1]
Tue, 17 Nov 2026 18:59:07 UTC (4,023 KB)
来源:arXiv cs.AI · arxiv.org