CADENCE: Closing the Reasoning Gap via Coverage-Adaptive On-Policy Distillation
CADENCE reports a 0.5B student closing up to 76.2% of the reasoning gap to its teacher on GSM8K.
The paper targets three distillation failure modes: cold-start collapse, static KL scheduling, and sparse pass/fail rewards. Its DRIFT method mixes forward- and reverse-KL objectives per token on student-sampled traces, with added coverage, entropy, partial-credit, brevity, calibration, and self-distillation components. On GSM8K, the student rises from 48.7% pretrained pass@1 to 69.8% with a 1.5B teacher, and to 72.1% with a 3B teacher. The authors say the runs used a single Apple Mac Studio with 64GB unified memory. HF Daily Papers' note
The paper targets three distillation failure modes: cold-start collapse, static KL scheduling, and sparse pass/fail rewards. Its DRIFT method mixes forward- and reverse-KL objectives per token on student-sampled traces, with added coverage, entropy, partial-credit, brevity, calibration, and self-distillation components. On GSM8K, the student rises from 48.7% pretrained pass@1 to 69.8% with a 1.5B teacher, and to 72.1% with a 3B teacher. The authors say the runs used a single Apple Mac Studio with 64GB unified memory. HF Daily Papers' note
score 5