finding
active
finding:r1-general-reasoning-sae-learned-features-reconstructing-reasoning-behaviors-such-as-backtrackingR1 general reasoning SAE learned features reconstructing reasoning behaviors such as backtracking
Curated features show core reasoning behaviors are captured.
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- BacktrackingcitesReasoning behavior for which the SAEs learned features.
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.
- Central interpretive assertion of the post.
- Supports the awareness hypothesis.
- Out-of-distribution generalization of SAE features.
- A promising property for interpretability analysis off-distribution.
- Limitation acknowledged regarding generalizability of the reflection identification method
- Evidence for the training-prefix hypothesis.
- Claim that feature grounding enables interpretability metrics.
- Future work.