claim
active
claim:training-saes-on-mean-hidden-state-yields-features-capturing-high-level-turn-characteristics-topic-style-behavior-rather-than-lexical-patternsTraining SAEs on mean hidden state yields features capturing high-level turn characteristics (topic, style, behavior) rather than lexical patterns
Central claim.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (10)
finding
- Turn-averaged features have highest embedding similarity 0.663 at k=3; per-token max lowest at 0.617associated_withsupportsConsistent across k=1,2,3,5.
- 20,000 trials.
- Example A.
- Contrast in §6.3.
- Judge sensitivity.
- Pairwise coverage.
- Example B.
- 5-way ranking.
- Example C.
- Table 1 on Qwen-2.5-7B holdout.
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.
- A promising property for interpretability analysis off-distribution.
- Out-of-distribution generalization of SAE features.
- Feature cluster examples from the UMAP.
- Extension of mechanistic interpretability findings to the metacognitive domain
- Conclusion.
- Novel finding that agentic self-evaluation of emotionality correlates with feature persistence
- Authors' experience-based claim.
- Supports scorer's preference for enacted reflection over described reflection; internals flag what self-report does not