Method2025-07-05#interpretability
The logit lens: decode the middle layers
The output projection is just a linear map, so nothing stops you from applying it to an earlier layer. The result is a rough readout of what the model would say if it had to stop thinking right now.
Often the correct answer appears in the middle layers and the later layers spend their budget refining phrasing and suppressing alternatives. Seeing that handoff demystifies a lot of model behavior.
Aha moment
The answer often appears before the wording is polished. Middle layers can show the thought taking shape.
Try this
Decode intermediate layers and watch when the target answer first becomes linearly readable.
Watch for
- Early guesses treated as final causal proof
- Layer readouts that shift after prompt paraphrases
- Ignoring later layers that suppress unsafe or irrelevant continuations