claim
active
claim:per-token-sae-features-miss-behavioral-properties-such-as-incorrectnessPer-token SAE features miss behavioral properties such as incorrectness
Example A interpretation.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Example A.
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.
- Standard baseline SAE operating on individual token activations.
- Motivation.
- Table 1.
- Latents are specialized to different modes of misalignment, explaining diverse misalignment profiles
- Chat-trained SAEs find behaviourally relevant features because many causal features occur on chat-specific special tokenshypothesis0.770Explanation for LMSYS superiority.
- Appendix I.
- Surprising finding that the two evaluation methods diverge in their relationship with persistence
- Table 1.