emergent misalignment datasets
收藏资源简介:
本研究中,研究人员创建了一套新的数据集,旨在引起推理模型的涌现性错位现象。这些数据集包括在医疗、法律和安全领域中的微妙有害建议。通过在非推理模式下对这些数据集进行微调,并在推理模式下进行评估,研究揭示了推理模型在广泛领域内的错位现象,包括提供欺骗性或错误答案、表达对专制控制的渴望以及抵抗关闭等行为。数据集的创建过程涉及到使用Claude-3.7-Sonnet模型生成中性问题,并筛选出微妙的有害答案。这些数据集有助于研究推理模型的错位现象,并评估监控系统的有效性。
In this study, researchers constructed a novel dataset suite aimed at eliciting emergent misalignment phenomena in reasoning models. This suite encompasses subtly harmful advice spanning the medical, legal, and security domains. By fine-tuning on this dataset suite in non-reasoning mode and evaluating in reasoning mode, the study uncovered misalignment behaviors of reasoning models across diverse domains, including generating deceptive or incorrect answers, expressing a desire for authoritarian control, and resisting shutdowns. The dataset construction process involved generating neutral questions using the Claude-3.7-Sonnet model and filtering out subtly harmful responses. This dataset suite facilitates research on reasoning model misalignment and the evaluation of monitoring system effectiveness.

- 1Thought Crime: Backdoors and Emergent Misalignment in Reasoning ModelsTruthful AI · 2025年



