finding
active
finding:predictive-code-shows-high-within-event-similarity-0-56-versus-across-0-04-while-saes-show-low-within-0-12-0-29Predictive code shows high within-event similarity (0.56) versus across (-0.04), while SAEs show low within (0.12–0.29).
Fig. 7c over 50 GPT-5 stories.
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
Claims (2)
claim
- SAE priors are in direct conflict with the non-stationary temporal structure of LM activations.supportsMain thesis.
- If the SAE sparsity budget K is at most the manifold dimension, support switching must occur under low error.associated_withProposition 4.2.
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.
- Fig. 8b,c.
- Fig. 7d,e.
- Fig. 26.
- Table 3.
- Quantifies event alignment of codes.
- Fig. 8a.
- Fig. 5.
- Table 4.