arXiv：cs.AI· Wenlong Zhang, Zhengbo Jiao, Chenxu Zhang, Lekang Jiang, SiYuan Ma, Qituan Zhang, Guo Chen, Linfeng Zhang·· 5 小时前AI 评分43
递归 Harness 自我改进：面向前沿推理数据合成
Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis
AI 导读
研究者提出"任务-工具协同演化"框架，实现推理数据合成中的递归 Harness 自我改进（RSI），在生成过程中将求解失败转化为可复用技能，并在每批任务后修订技能、提示词与工作流。在数学、编程和科学领域，经过十四轮演化，求解器平均准确率从 100.0% 降至 54.8%。用 10K 合成数学样本微调的 27B 学生模型在 APEX 上达到 62.5% mean-16 准确率。
正文
Abstract:Generating progressively harder reasoning problems requires synthesis procedures that adapt as the task distribution evolves. Existing task-level recursion reuses generated problems as seeds but leaves the construction harness unchanged. We present task-harness co-evolution, a framework for recursive harness self-improvement (RSI) in reasoning-data synthesis. Online self-improvement converts intermediate solver failures into reusable skills during generation. Post-task self-improvement revises skills, prompts, and workflows after each batch, adopting candidates only when they generate harder valid tasks within a bounded cost increase. Model weights and verification criteria remain fixed. Across mathematics, coding, and science, mean solver accuracy decreases from 100.0% to 54.8% over fourteen evolution rounds. Ablations show that combining both update schedules produces harder tasks than fixed-harness recursion or either schedule alone. The resulting data improves downstream SFT and GRPO performance. In particular, a 27B student fine-tuned on 10K synthesized mathematics examples achieves 62.5% mean-16 accuracy on APEX, competitive with selected frontier-model references. These results support adapting the synthesis harness alongside the tasks to generate increasingly challenging data with downstream training value.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.03548 [cs.AI] |
| (or arXiv:2610.03548v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.03548 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Wenlong Zhang [view email]
[v1]
Fri, 2 Oct 2026 16:30:15 UTC (1,525 KB)
来源：arXiv：cs.AI · arxiv.org