claim
active
claim:sae-priors-are-in-direct-conflict-with-the-non-stationary-temporal-structure-of-lm-activationsSAE priors are in direct conflict with the non-stationary temporal structure of LM activations.
Main thesis.
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 (5)
finding
- Autocorrelation of LM activations varies with sequence position, unlike the stationary surrogate.supportsFig. 2b,f.
- Fig. 2a,e.
- Fig. 7c over 50 GPT-5 stories.
- Fig. 2d,h.
- Fig. 5.
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.
- Proposition 4.1 central claim.
- Latent #10 activation increase correctly classifies all correct vs incorrect fine-tuned models in Figure 9
- Explains why the boundary appears fixed: the prior hides its context-dependent nature.
- Claim that feature grounding enables interpretability metrics.
- Model diffing identifies a small, interpretable set of latents responsible for emergent misalignment
- Claim that capability emerges from architecture, not data, and that later models lose the surprise.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence