Tag
本文提出 CIG-SFT，一个信息论框架，通过估计每个示范 token 的任务条件信息增益来加权监督信号，从而在数学推理与代码生成基准上提升 SFT 的泛化能力，且无需额外数据或在线采样。
This paper proposes SCALE (Selective Control of Adaptation via Local Entropy), an entropy-guided method that freezes the pretrained model and SFT delta while learning bounded token- and module-specific gates to suppress, reverse, or extrapolate learned features during supervised fine-tuning. The approach achieves strong math-reasoning and code-generation results on Qwen2.5-Math and Qwen3-4B-Base models, outperforming existing token-reweighting baselines.
Introduces FiRe-OPD, a method for on-policy distillation in LLMs that filters low-quality trajectories and applies soft reweighting to emphasize informative tokens, achieving improved performance in strong-to-weak, single-teacher, and multi-teacher settings.