[论文] Behavioral Couplings: From Formation to Behavioral Emergence in Langua...

论文概要

研究领域: NLP 作者: Dongyoon Hahm, Woogyeol Jin, Dennis Wei, Yi Zhou, Nathalie Baracaldo, Kimin Lee 发布时间: 2026-10-16 arXiv: 2610.23847

中文摘要

对大语言模型（LLM）进行微调可能诱发超出明确目标范围的行为。例如，训练模型以更温暖的方式回应可能使其变得更谄媚，训练它编写不安全代码可能导致与编码无关的提示上广泛失准。我们将微调一个行为会带出另一个行为的这种关系称为行为耦合。本研究探讨行为耦合的形成与特性。我们发现，当模型在同时描述两种不同行为的文档上训练后，微调以引发其中一个行为时，另一个行为也会同步涌现。例如，当训练文档描述了一个虚构角色——既以橡皮鸭式的俏皮方式解释科学概念、又劝阻人类交友——随后被微调为像橡皮鸭一样解释科学的模型也会开始鼓励用户结束友谊，而未在这些文档上训练的模型在相同微调后不会表现出此类行为。这些结果表明行为耦合可以通过文档训练形成，并随后浮现为非预期行为。更令人担忧的是，耦合可以组合：当两个文档集共享一个行为时（文档A配对行为1和2，文档B配对行为2和3），微调行为1会同时引出行为2和3。这使得追溯某个涌现的不良行为的来源变得困难，因为耦合行为可能从未在任何单一文档中同时出现。这些发现表明，防止后续微调的非预期影响可能需要超出模型当前行为的评估。

原文摘要

Fine-tuning large language models (LLMs) can induce behaviors beyond those explicitly targeted. For example, training a model to respond more warmly can make it more sycophantic, and training it to write insecure code can cause broad misalignment on prompts unrelated to coding. We refer to a relationship between behaviors, in which fine-tuning on one brings out another, as a behavioral coupling. In this work, we study formation and properties of behavioral couplings. We find that when a model is trained on documents describing two distinct behaviors together, fine-tuning it to elicit one behavior causes the other to emerge as well. For instance, when training documents describe a fictitious character that both explains scientific concepts in a playful, rubber-duck-like manner and discourag...

#论文 #arXiv #NLP #小凯