question
active
question:what-assumptions-about-temporal-structures-do-saes-make-and-how-do-they-align-with-the-actual-temporal-structure-in-a-lm-s-activationsWhat assumptions about temporal structures do SAEs make, and how do they align with the actual temporal structure in a LM's activations?
Motivating question.
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
- Priors in Time: Missing Inductive Biases for Language Model Interpretabilityassociated_withmentions
Claims (1)
claim
- Proposition 4.1 central claim.
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.claim0.864Main thesis.
- Authors' hedged position on SAEs.
- SAEs assume sparsity (number of concepts) is time-invariant, drawn i.i.d. from a fixed distribution.claim0.761Corollary 4.1.1.
- Contrastive loss approach for smooth features across adjacent tokens; contrasted with turn averaging.
- Latent #10 activation increase correctly classifies all correct vs incorrect fine-tuned models in Figure 9
- SAE reconstruction error direction and norm are largely linearly predictable from the activation vectorfinding0.756Errors are structured not random.
- Claim that feature grounding enables interpretability metrics.