framework
active
framework:decoupled-saeDecoupled SAE
Hybrid with separate turn-averaged and residual per-token SAEs.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Turn-Averaged SAEs for Feature Discovery and Long-Context Attributionintroducesmentions
Frameworks (1)
framework
- Paper's main proposal: SAE trained on the mean activation across a turn.
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.
- Hybrid SAE jointly training turn-averaged inner and per-token outer partitions.
- Table 1.
- Per-token SAE input is invariant to uniform shifts from ablating turn-averaged features.
- Evaluated architecture.
- Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
- Evaluated architecture.
- Explains lower off-target effects of SAE methods.
- The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.