arXiv cs.AI· Dewu Zheng, Yang Ye, Hongyu Zhang, Jun Yu, Yanlin Wang·· 23 小时前AI 评分42
RSI-Prime:用 350 条实验轨迹训练模型学会训练模型
RSI-Prime: Training Models to Train Models
AI 导读
研究者构建 RSI-Prime 数据集,包含七类任务的 350 条完整实验轨迹,用于提升 LLM 的模型训练能力。该数据集通过细粒度决策识别与质量标注,并采用决策级 loss masking 排除低质量行为的监督。在自主后训练基准 PostTrainBench 上,GLM-5.3 经 RSI-Prime 训练后整体性能相对提升 31.95%。
正文 · 原文
Abstract:Enabling large language models (LLMs) to autonomously conduct training and improve target models is an important foundation for recursive self-improvement (RSI). As LLMs expand their role from being trained to conducting training, their model-training capability warrants dedicated study and systematic development. To this end, we construct RSI-Prime, a dataset for improving model-training capability, comprising 350 complete experimental trajectories across seven task categories. To help LLMs learn to make high-quality experimental decisions from training practice, we perform fine-grained decision identification and quality annotation, then apply decision-level loss masking to exclude supervision on low-quality behavior. Experiments with GLM-5.3 on PostTrainBench, a benchmark for autonomous post-training, show that training on RSI-Prime yields a 31.95% relative improvement in overall performance over the original GLM-5.3, highlighting a path toward future RSI systems in which models autonomously sustain their own advancement.
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.23138 [cs.AI] |
| (or arXiv:2611.23138v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.23138 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Dewu Zheng [view email]
[v1]
Tue, 17 Nov 2026 17:55:30 UTC (317 KB)
来源:arXiv cs.AI · arxiv.org