Reproduce a documented logit-lens experiment

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R34 — TransformerLens Logit Lens tutorials & attribution limits

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Reproduce a documented Logit Lens experiment in TransformerLens: project intermediate residual-stream states into vocabulary space across layers.

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   Test prompts where intermediate layer top-logit predictions flip in final layers; document why a single layer visualization does not prove internal model "intent" or deception.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Logit-lens experiment notebook with recorded layer outputs and methodological limitations note.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R34TransformerLens Docs↗ transformerlensorg.github.io

  HookedTransformer, residual stream, ActivationCache, Logit Lens, and Activation Patching tutorials.

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsTransformerLens · reference
  TransformerLens Documentation

  Documents HookedTransformer, ActivationCache and the residual stream hooks you reuse to project intermediate layer states into vocabulary space for a logit-lens experiment.

  R34checked 2026-11-11
• videoAnastasia Borovykh · 21 min
  Decoding hidden states of Phi-3 with LogitLens (Interpretability Series)

  Applies the logit lens layer by layer on a small model and reads the evolving token predictions, a direct template for reproducing the projection on your own model and prompt.

  R34checked 2026-11-11
• videoNDIF Team · 95 min
  Neural Mechanics Week 1: LLM Foundations and Logit Lens with David Bau

  Derives why the unembedding can be applied to intermediate residual states and demonstrates it live, giving you the theory to explain each layer's decoded distribution rather than just plotting it.

  R34checked 2026-11-11
• articleLessWrong (nostalgebraist) · ~20 min read
  interpreting GPT: the logit lens — LessWrong

  The original logit-lens writeup showing how intermediate residual states decode to sensible tokens with layer-dependent convergence, the documented experiment you are asked to reproduce.

  R34checked 2026-11-11
• repoARENA (GitHub) · exercise notebooks
  GitHub - ARENA-education/ARENA_materials

  Contains TransformerLens exercise notebooks with solution cells for residual-stream and logit-lens style analyses, so you can check your implementation against a known-good one.

  R34checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.