---
id: 20260808-T0-05
title: "新研究：程序脚手架与模型参数协同进化，后训练效果更好"
title_en: "New Research: Co-Evolving Scaffolds and Parameters Improves Post-Training"
url: https://ai.daily.yangsir.net/daily/20260808-T0-05
issue_date: 2026-08-08
publish_date: 2026-08-07T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2608.05156
---

# 新研究：程序脚手架与模型参数协同进化，后训练效果更好

arXiv新论文提出一种后训练方法，将模型参数和推理时的程序化脚手架（procedural scaffolds）协同进化，而非像传统做法那样分别独立设计。该方法旨在让模型在训练阶段就内化复杂的推理流程，减少测试时对外部脚手架的依赖。初步结果显示，这种协同优化策略能让模型自动获取并内化更复杂的技能，但仍需更多验证。

## English Version

**New Research: Co-Evolving Scaffolds and Parameters Improves Post-Training**

A new arXiv paper introduces a post-training method that co-evolves model parameters with procedural scaffolds, unlike traditional approaches that design them independently. This strategy aims to internalize complex reasoning processes during training, reducing reliance on external scaffolds at inference time. Initial results suggest the co-optimization enables models to automatically acquire and internalize more complex skills, though further validation is needed.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2608.05156)

**详情页**：https://ai.daily.yangsir.net/daily/20260808-T0-05

---

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