finding
active
finding:open-ended-explanation-training-397b-improves-0-59-to-0-68-hint-attribution-accuracy-8b-shows-no-improvementOpen-ended explanation training: 397B improves 0.59 to 0.68 hint attribution accuracy; 8B shows no improvement
Mixed results.
Source paper
extracted_from(2026) · Adam Karvonen · Euan Ong · Subhash Kantamneni · Samuel Marks
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Confident wrong counterfactuals mislead.
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.
- Generalization to hint setting.
- Base corrections: 72.9% Failed Fix, 14.7% Fixed; retractions: 69.4% Correct Retract, 30.0% Incorrect Retractfinding0.782Label distributions.
- Narrow-range greater-than feature.
- More complex than simple comparison.
- Special-case lookup-like feature.
- Base failure mode.
- Soft clues redundant with transcript.
- Demonstrates severity of training-deployment gap after RL