Interpreting LM Residual Streams via Lexical Reference Frames
October 14, 2026
Lexical reference frames approximate the native semantics of a language model's hidden states by reading them through the model's remaining layers. This method outperforms fixed-lens approaches, which are limited to 0.30 and 0.35 coherence relative to the native frame.
HOW THIS AFFECTS YOU
●
researcherThis provides a more accurate method for mechanistic interpretability by accounting for context-adaptive reference frames.