finding
active
finding:sae-reconstruction-error-direction-and-norm-are-largely-linearly-predictable-from-the-activation-vectorSAE reconstruction error direction and norm are largely linearly predictable from the activation vector
Errors are structured not random.
Source paper
extracted_from(2025) · Lee Sharkey · Bilal I. Chughtai · Joshua Batson · Jack W. Lindsey +25
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- SDL reconstruction error might be reduced by improved SDL training or new decomposition methodsassociated_withsupportsStructured errors imply potential solutions.
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.
- Quantifies SAE reconstruction damage.
- Open question in limitations.
- Features act as local patch detectors.
- A promising property for interpretability analysis off-distribution.
- Explains why variance correction is needed to see the self-evaluation–persistence relationship
- Conclusion.
- Limitation statement.
- Statistically rigorous analysis of Claude introspection; suggests models may have latent introspective capabilities that can be enhanced or disrupted.