让下一篇重要论文，不再错过。

原文摘要

Large language models are frequently updated for various use cases, where filtering out misaligned training samples is a common practice for preventing post-update misalignment. However, alignment is inherently context-dependent: a behavior that is aligned in one context may be inappropriate in another. For example, recommending that researchers retain their data for reproducibility is aligned, whereas recommending that a mobile app retain users' sensitive data may violate users' privacy. Starting from this observation, we identify a posttraining phenomenon where aligned training induces misaligned behavior in other contexts. We call this phenomenon context confusion. We demonstrate context confusion across three domains: (1) Gender Equality, (2) Privacy, and (3) Physical Safety. Unlike the broad misalignment seen in emergent misalignment, context confusion causes domain-specific misalignment that evades general alignment evaluations. Further, it is not effectively reduced by injecting general alignment training data, but can be substantially reduced by including targeted alignment data for the misaligned domain or providing in-context learning examples during inference. Lastly, we provide a mechanistic explanation of context confusion. We observe that queries from different domains can undergo similar representational shifts during the fine-tuning. Consequently, a query from a different domain may activate the same behavioral feature learned during fine-tuning, which causes the behavior to transfer to a context where it is misaligned. Based on our findings, we argue that it is difficult to predict the alignment state of a model after training by inspecting the training data alone, which highlights the importance of comprehensive post-training alignment evaluations.