inserting teacher-generated corrections into student rollouts plus token-level KL supervision on those segments took mean avg@8 from 19.5% (best baseline) to 35.6%; either mechanism alone was insufficient.
inserting teacher-generated corrections into student rollouts plus token-level KL supervision on those segments took mean avg@8 from 19.5% (best baseline) to 35.6%; either mechanism alone was insufficient.
Cite as — Sawdust Shelf v2026-11-05.3 · No. 143 · fingerprint sha256:f1cf8893d6b8…
Share image rendered from the living shelf — the number, the grade and the fingerprint travel with the card.