claim
active
claim:discrimination-and-coverage-give-opposite-rankings-per-token-max-is-easiest-to-identify-turn-averaging-covers-the-full-rangeDiscrimination and coverage give opposite rankings: per-token max is easiest to identify, turn-averaging covers the full range
§5.5.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- 20,000 trials.
- Top discrimination on 2,000 holdout turns.
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.
- Example A.
- Averaging acts as a low-pass filter that smooths away per-token detail and retains broad conceptsclaim0.844Intuition.
- Design choice validated empirically.
- Prediction from coverage results.
- LongBench v2 example E.
- Justifies the choice of response tokens for persona vector extraction in the pipeline
- With raw data, wider token strides give more accurate approximations at same token budget.finding0.801Cost tradeoff.
- Turn-averaged features have highest embedding similarity 0.663 at k=3; per-token max lowest at 0.617finding0.800Consistent across k=1,2,3,5.