Uncovering Model Processing Strategies with Non-Negative Per-Example Fisher Factorization
Researchers have developed a new interpretability method called Non-Negative Per-Example Fisher Factorization (NPEFF) to understand how language models arrive at their predictions. NPEFF decomposes per-example Fisher matrices, revealing components that correspond to specific processing strategies. The method has been demonstrated to analyze and mitigate effects in tasks like unlearning and in-context learning, showing advantages over existing techniques such as gradient clustering and sparse autoencoders. The team has also released the code for NPEFF. AI
IMPACT Provides a new tool for understanding and potentially manipulating internal model behaviors.