hypothesis
active
hypothesis:information-concentration-at-specific-tokens-interacts-with-turn-averagingInformation concentration at specific tokens interacts with turn-averaging
Future work direction.
Source paper
extracted_from(2026) · Kevin Der · Harish Kamath · Thompson, Ben
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- §5.5.
- Averaging acts as a low-pass filter that smooths away per-token detail and retains broad conceptsclaim0.774Intuition.
- LongBench v2 example D.
- Example A.
- Prediction from coverage results.
- Turn-averaged features have highest embedding similarity 0.663 at k=3; per-token max lowest at 0.617finding0.759Consistent across k=1,2,3,5.
- Layer localization.
- Explains hard-to-read dataset patterns.