method
active
method:sum-pooling-of-per-token-featuresSum pooling of per-token features
Aggregation of token-level activations.
Neighborhood — ranked by edge-count
Papers (1)
paper
Frameworks (1)
framework
- Per-Token SAEusesStandard baseline SAE operating on individual token activations.
Methods (1)
method
- Max pooling of per-token featuresrelated_toAlternative aggregation giving high discrimination.
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.
- Figure 1.
- Top discrimination on 2,000 holdout turns.
- Design choice validated empirically.
- Prediction from coverage results.
- §5.5.
- Example A.
- Behavior where information about full clauses is encoded over clause-ending punctuation tokens in LLMs
- Figure 7.