Wu et al. first decompose the KL divergence used in distribution-constrained fine-tuning into a cross-entropy term and a negative entropy term.
Entropy feedback dynamically adjusts replay weights to prevent catastrophic forgetting during fine-tuning
Wu et al. show that output entropy can substitute for reference-model KL calculation, driving a closed-loop controller that tunes replay strength in supervised fine-tuning.
Chinese Tech
Shangxi Wu · Xiaorong Yan · Xiaojun Meng · Yuxiang Zhang · Jiansheng Wei
Beijing Jiaotong University · The Hong Kong Polytechnic University · Huawei Noah's Ark Lab
Research Digest··2 min read
Wu et al. introduce ECHO, a closed-loop control framework that uses output entropy as a cheap, reference-free proxy for distribution shift during supervised fine-tuning.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.