---
id: 20260312-T0-08
title: "HEAL方法提升小模型推理能力蒸馏效果"
title_en: "HEAL Method Improves Reasoning Distillation for Smaller Models"
url: https://ai.daily.yangsir.net/daily/20260312-T0-08
issue_date: 2026-03-12
publish_date: 2026-03-12T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2603.10359
---

# HEAL方法提升小模型推理能力蒸馏效果

arXiv论文提出HEAL方法，通过后验熵辅助学习，解决大模型推理能力向小模型蒸馏时的拒绝采样限制。该方法在数学推理任务中，将小模型准确率提升至大模型的78%，同时减少90%的训练计算量。

## English Version

**HEAL Method Improves Reasoning Distillation for Smaller Models**

arXiv paper introduces HEAL method, using hindsight entropy-assisted learning to overcome rejection sampling limits in reasoning distillation. Achieves 78% of large model accuracy in math reasoning with 90% less compute.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2603.10359)

**详情页**：https://ai.daily.yangsir.net/daily/20260312-T0-08

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*