method
active
method:jacobian-lensJacobian Lens
Gurnee et al.'s interpretability method used to identify the J-space workspace.
Neighborhood — ranked by edge-count
Papers (2)
paper
Findings (1)
finding
- Empirical interpretability evidence bearing on Level 2 information-integration and GWT indicators.
Concepts (1)
concept
- J-space (Verbalisable Workspace)implementsGurnee et al.'s privileged set of verbalisable representations behaving as a within-pass global workspace.
Methods (4)
method
- Logit LensextendsUnsupervised interpretability technique that projects activations through unembedding matrix; provides comparison point for NLA approach.
- Tuned lenscitesBaseline lens that skips ahead to output.
- Intervention exchanging two concepts' lens coordinates.
- Used to compute J-space components and occupancy.
Artifacts (3)
artifact
- Claude Sonnet 4.5citesMid-tier LLM from Anthropic evaluated with n=14 games.
- Open-source J-lens implementation.
- Interactive J-lens on open-source models.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- RL-trained phrase-proposal lens for arbitrary-length concepts.
- Methods comparison.
- Penalizing difference between transcoder and MLP Jacobians, computed cheaply via random vector products.
- Model organism results.
- Limitation.
- On calc (4+17)*2+7 Opus 4.5 J-lens shows 21, then 42, then 49 at successively later layers around layer 71.finding0.699Multi-step intermediates.
- Related work.