framework
active
framework:temporal-saesTemporal SAEs
Contrastive loss approach for smooth features across adjacent tokens; contrasted with turn averaging.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Turn-Averaged SAEcontradictsPaper'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.
- Motivating question.
- Interpretable features extracted by sparse autoencoders used as steering targets in this study
- Proposition 4.1 central claim.
- Hybrid SAE jointly training turn-averaged inner and per-token outer partitions.
- In active inference, distant future predictions are naturally less precise, leading to Bayes-optimal discounting without an explicit discount factor.
- Token-level analysis of OTD and backtracking latent activations aligned at correction points across episodes
- SAE priors are in direct conflict with the non-stationary temporal structure of LM activations.claim0.735Main thesis.
- SAEs assume sparsity (number of concepts) is time-invariant, drawn i.i.d. from a fixed distribution.claim0.731Corollary 4.1.1.