机械论数据归因：追踪可解释LLM单元的训练起源
Mechanistic Data Attribution: Tracing the Training Origins of Interpretable LLM Units
AI总结：
提出机械论数据归因（MDA）框架，利用影响函数将可解释单元追溯到特定训练样本，通过因果验证表明干预高影响样本可显著调节可解释头的涌现，并发现重复结构数据作为机械催化剂，同时验证了归纳头与上下文学习之间的功能联系。
摘要：
尽管机械论可解释性已在LLM中识别出可解释电路，但它们在训练数据中的因果起源仍然难以捉摸。我们引入了机械论数据归因（MDA），这是一个可扩展的框架，利用影响函数将可解释单元追溯到特定训练样本。通过在Pythia系列模型上的广泛实验，我们因果验证了目标干预——移除或增加一小部分高影响样本——显著调节了可解释头的涌现，而随机干预则没有效果。我们的分析表明，重复的结构化数据（例如LaTeX、XML）充当了机械催化剂。此外，我们观察到针对归纳头形成的干预会引发模型上下文学习（ICL）能力的同步变化。这为关于归纳头与ICL之间功能联系的长期假设提供了直接的因果证据。最后，我们提出了一种机械论数据增强流水线，该流水线在不同模型规模上一致地加速电路收敛，为引导LLM的发展轨迹提供了一种原则性方法。
英文摘要：
While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
