claim
active
claim:saes-implicitly-impose-an-i-i-d-prior-over-time-on-latents-and-thereby-conceptsSAEs implicitly impose an i.i.d. prior over time on latents and thereby concepts.
Proposition 4.1 central claim.
Source paper
extracted_from(2025) · Lubana, Ekdeep Singh · Rager, Can · Hindupur, Sai Sumedh R. · Costa, Valerie +12
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
Claims (2)
claim
- Future-work perspective.
- SAEs assume sparsity (number of concepts) is time-invariant, drawn i.i.d. from a fixed distribution.extendsCorollary 4.1.1.
Questions (1)
question
- Motivating question.
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.
- SAE priors are in direct conflict with the non-stationary temporal structure of LM activations.claim0.796Main thesis.
- Authors' hedged position on SAEs.
- Claim that feature grounding enables interpretability metrics.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Inception-v1 SAE shows meaningful fraction of latents with very high nearest-neighbor cosinefinding0.763Fig 8; may be duplication.
- SAEs may have the right inductive biases for deeper layers but not middle or earlier layers.claim0.759Speculation about layer dependence.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Specific tuning curve structure.