arXiv cs.AI· Khouloud Saadi, Di Wang·· 14 小时前AI 评分23
有效性校准的推理蒸馏：用局部有效性动态调节蒸馏信号
Validity-Calibrated Reasoning Distillation
AI 导读
研究者提出“有效性校准的推理蒸馏”框架，不再强制学生模型逐 token 模仿教师轨迹，而是在相同前缀下比较学生与教师提出的下一步动作，用两者的相对局部有效性调节蒸馏更新强度。该方法在数学、代码和指令遵循基准上优于强蒸馏基线。结果表明，有效的 LLM 推理蒸馏取决于学习信号的局部校准分配，而非僵化的轨迹模仿。
正文 · 原文
Abstract:Reasoning distillation aims to transfer multi-step reasoning capabilities from large language models to smaller, efficient ones. While recent methods have shown promising gains, they typically rely on static teacher-student hierarchies and frame distillation as trajectory imitation. This is misaligned with the structure of reasoning, where intermediate steps are locally under-specified: global correctness constrains the final answer, but does not uniquely determine each intermediate move. We propose validity-calibrated reasoning distillation, a framework that treats reasoning distillation as a problem of local learning-signal allocation rather than path alignment. Instead of enforcing token-level imitation, we compare the student's and teacher's proposed next-step actions under the same prefix and use their relative local validity to modulate the strength of the distillation update. This yields a dynamic, context-dependent supervision that preserves the teacher's structural guidance while adapting update strength to local reasoning quality. Across math, code, and instruction-following benchmarks, our method outperforms strong distillation baselines. These results indicate that effective LLM reasoning distillation is governed not by rigid trajectory imitation, but by locally calibrated allocation of learning signal.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.04078 [cs.LG] |
| (or arXiv:2605.04078v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.04078 arXiv-issued DOI via DataCite |
Submission history
From: Khouloud Saadi [view email]
[v1]
Tue, 14 Apr 2026 12:32:12 UTC (8,139 KB)
[v2]
Sat, 9 May 2026 13:58:28 UTC (9,341 KB)
[v3]
Thu, 29 Oct 2026 08:28:26 UTC (8,133 KB)
来源：arXiv cs.AI · arxiv.org