method
active
method:jacobian-matchingJacobian matching
Penalizing difference between transcoder and MLP Jacobians, computed cheaply via random vector products.
Neighborhood — ranked by edge-count
Papers (1)
paper
- A Toy Model of Mechanistic (Un)Faithfulnessintroducesmentions
Concepts (3)
concept
- Superpositionassociated_withPhenomenon where models represent more features than dimensions via almost-orthogonal directions.
- Idea that decomposed components must causally reflect the model's computation; enforced by adversarial ablation in VPD.
- Gradient maskingassociated_withAdversarial robustness phenomenon by which a transcoder could cheat Jacobian matching.
Hypotheses (1)
hypothesis
- Gradient masking risk; mitigated by tanh L1 or penalizing Jacobian contribution L1.
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.
- Positive result with caveats.
- Toy-model fix of unfaithfulness.
- Tuple retrieval by matching against a template, similar to relational database queries.
- Gurnee et al.'s interpretability method used to identify the J-space workspace.
- Limit on Jacobian matching.
- A core requirement of the inductive theory of mirror self-recognition, implemented by the self-prior's visual-proprioceptive associations
- Adding Jacobian matching loss dramatically reduces datapoint-feature memorization metricfinding0.719Result of Jacobian matching.
- Distance between prior and target representations.