Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics

Key points

• We argue that every hidden state has a native reference frame, induced by the layers that read it next and adaptive to the context.1
• Lexical reference frames approximate it by reading a state through the model's own remaining layers with a stated number of rounds of attention between positions and a typical attention output for the rest, without fitting to the model's output.1
• No lens, with fixed directions per layer, can exceed the square root of the native frame's coherence, 0.30 and 0.35 at mid-depth of SmolLM3-3B and OLMo3-7B, whereas the first-order frame, LexFrame[1], reaches 0.97 and 0.97 and reads a state within 0.16 nats of the model's output.1
• In this way, lexical reference frames make every hidden state of the model readable, attributable and steerable.1

Sources (2)

• [1]Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 15, 04:29 PM

  The vocabulary of a language model is its native semantics: the embedding and the unembedding fix the meaning of a state at the two ends of the residual stream.

  We argue that every hidden state has a native reference frame, induced by the layers that read it next and adaptive to the context.

• [2]Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics
  Hugging Face Daily Papers · Oct 15, 12:00 AM · same content

Extractive summary: sentences quoted from the sources.