Supervised fine‑tuning (SFT) applies a uniform cross‑entropy loss to every target token, yet in mathematical reasoning different tokens provide unequal learning signals. Uniform treatment over‑sharpens already mastered tokens and pushes learning pressure onto uncertain, low‑confidence tokens, resulting in suboptimal dynamics. We therefore propose Trimmed Logit‑Gap SFT (TrimSFT), a simple token‑level reweighting method. TrimSFT scales the SFT loss according to the logit gap between the gold token and its strongest competitor. Supervision is trimmed at both extremes: tokens with large gaps (already mastered) and tokens with small or negative gaps (weakly supported) receive lower weight, concentrating learning in an intermediate gap region. We instantiate this with a Gaussian weight centered at margin m and bandwidth τ, requiring no reference model or extra forward pass. Experiments span six base models from Llama, Qwen and DeepMath families across five mathematical reasoning benchmarks. TrimSFT consistently outperforms standard SFT on all but one model, achieving the best average performance on five models and a maximum gain of 26.9 points on MATH500. Further analysis shows bandwidth τ matters more than the exact margin location, and half‑trim variants that trim only one side perform worse. Token‑level logit‑gap distribution analysis indicates TrimSFT reshapes model confidence more evenly than uniform SFT or monotonic reweighting methods. These findings suggest reasoning‑oriented SFT benefits from trimming both extremes rather than treating all tokens uniformly.
Review