framework
active
framework:nested-sae

Nested SAE

Hybrid SAE jointly training turn-averaged inner and per-token outer partitions.

Neighborhood — ranked by edge-count

Frameworks (2)

framework
  • Paper's main proposal: SAE trained on the mean activation across a turn.
  • Nested dictionary approach that inspires the nested loss.

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

  • Decoupled SAEframework0.816
    Hybrid with separate turn-averaged and residual per-token SAEs.
  • P-Annealing SAEframework0.786
    Evaluated architecture.
  • BatchTopK SAEframework0.769
    Evaluated architecture.
  • SAE featuresconcept0.765
    The individual, supposedly monosemantic directions learned by SAEs; argued here to fragment manifolds into disconnected pieces.
  • SAE Latentsconcept0.763
    Interpretable features extracted by sparse autoencoders used as steering targets in this study
  • TopK SAEframework0.761
    Evaluated architecture (Gao et al.).
  • Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
  • Internal concepts a model uses, extracted as human-interpretable components.