concept
active
concept:sae-dark-matter-incomplete-reconstructionSAE dark matter / incomplete reconstruction
SAEs capture only a fraction of model computation.
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Current SAEs capture only a fraction of model computation due to limited training data and incomplete reconstructionassociated_withLimitation statement.
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.
- Open question in limitations.
- Claim that feature grounding enables interpretability metrics.
- Source of mismatch.
- Hybrid with separate turn-averaged and residual per-token SAEs.
- SAE reconstruction error direction and norm are largely linearly predictable from the activation vectorfinding0.724Errors are structured not random.
- Features act as local patch detectors.
- Foundational empirical result enabling all downstream analysis
- SAE priors are in direct conflict with the non-stationary temporal structure of LM activations.claim0.721Main thesis.