Research5d ago
Scale‑Sized Interaction Mapping for LLMs
Feature attribution isolates input features driving predictions. Researchers at BAIR Blog examine how feature, data, and mechanistic attribution reveal LLM internals, mapping how training examples shape behavior. The analysis equips practitioners to debug models and tighten safety by pinpointing problematic data or internal functions and clarifies internal module roles in depth.