Logit Lens

An interpretability technique that projects a transformer's intermediate-layer hidden states through the output unembedding to read the model's evolving next-token prediction across depth.

grounded in: Latest AI/tech trends, HN front page 2026-07-13 theme 'Mechanistic interpretability of LLMs: Researchers are pushing to actually explain model internals rather than treat them as black boxes'

Connected concepts

Mechanistic interpretability, Superposition, Activation patching, CoT Monitorability