arXiv：cs.LG（机器学习，全量分类）· Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati, Keerthiram Murugesan, Lei Yu·· 5 小时前AI 评分37
Activation-Conditioned Self-Distillation：一种无需参考文本的自蒸馏方法
Activation-Conditioned Self-Distillation
AI 导读
研究者提出 Activation-Conditioned Self-Distillation（ACSD），通过对比自生成轨迹中答对与其余轨迹的激活值提取引导向量，由冻结的基座模型在每个预测位置施加该向量，学生模型从自身前缀的下一 token 分布中学习。
正文
Abstract:On-policy self-distillation uses a model as its own teacher to provide dense supervision for reasoning, often through reference-solution conditioning. Providing privileged information does not by itself ensure effective token-level supervision throughout long responses. We introduce Activation-Conditioned Self-Distillation (ACSD), which extracts a steering vector by contrasting activations of self-generated trajectories that reach verified correct answers within a generation budget with those of all remaining trajectories. A frozen copy of the base model applies this vector at each prediction position, and the student learns from its next-token distributions on student-generated prefixes. Outcome verification is used for direction construction and calibration; distillation requires neither problem-specific reference text nor teacher parameter updates. The distilled student is used alone at inference. On each of five models, ACSD achieves the highest mean accuracy over four mathematical benchmarks among the evaluated methods. On DeepSeek-R1-0528-Qwen3-8B, mean mathematical accuracy reaches 71.9\% and LiveCodeBench v6 pass@12 reaches 70.9\%, compared with 69.0\% and 66.3\% for the reference-conditioned OPSD baseline. Contrasts among correct trajectories also support distillation, and extracted directions can be reused across mathematical training datasets. On fixed student trajectories, ACSD maintains more stable late-position logit-update magnitudes than OPSD.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2609.38342 [cs.LG] |
| (or arXiv:2609.38342v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.38342 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Zhexi Lu [view email]
[v1]
Tue, 29 Sep 2026 18:06:56 UTC (249 KB)
来源：arXiv：cs.LG（机器学习，全量分类） · arxiv.org