arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.01692cs.LGcs.AI
收藏

推理作为梯度:超越树搜索的MLE代理扩展

Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search

Nanjing University(南京大学);Microsoft Research Asia(微软亚洲研究院);Peking University(北京大学);Zhejiang University(浙江大学);Wuhan University(武汉大学);Hong Kong University of Science and Technology(香港科技大学)

Yifei Zhang, Xu Yang, Xiao Yang, Bowen Xian, Qizheng Li, Shikai Fang, Jingyuan Li, Jian Wang, Mingrui Xu, Weiqing Liu, Jiang Bian

AI总结:

本文提出Gome,一种基于梯度优化的MLE代理,通过将结构化诊断推理映射到梯度计算,实现更高效的优化。实验显示,随着模型能力增强,梯度优化在前沿模型中表现更优,且在受限预算下达到35.1%的任何奖牌率。

摘要:

基于大语言模型的机器学习工程(MLE)代理主要依赖树搜索,这是一种基于标量验证分数的无梯度优化方法。随着LLM推理能力提升,穷举搜索相比定向更新变得效率低下,类似于准确梯度使随机搜索的下降更高效。我们引入Gome,一种将基于梯度的优化机制化的新MLE代理。Gome将结构化诊断推理映射到梯度计算,成功记忆映射到动量,多轨迹执行映射到分布式优化。在封闭世界协议下,该协议隔离了架构效应与外部知识,Gome在MLE-Bench上以受限的12小时预算单块V100 GPU实现了35.1%的任何奖牌率。跨10个模型的扩展实验揭示了关键转折点:对于较弱模型,树搜索通过穷举探索弥补不可靠推理的不足;随着推理能力增强,基于梯度的优化逐步超越,前沿模型间差距扩大。鉴于面向推理的LLM快速进步,这使基于梯度的优化成为更具优势的范式。我们发布了代码库和GPT-5轨迹于https://github.com/microsoft/RD-Agent。

英文摘要:

LLM-based agents for machine learning engineering (MLE) predominantly rely on tree search, a form of gradient-free optimization that uses scalar validation scores to rank candidates. As LLM reasoning capabilities improve, exhaustive enumeration becomes increasingly inefficient compared to directed updates, analogous to how accurate gradients enable efficient descent over random search. We introduce Gome, an MLE agent that operationalizes gradient-based optimization. Gome maps structured diagnostic reasoning to gradient computation, success memory to momentum, and multi-trace execution to distributed optimization. Under a closed-world protocol that isolates architectural effects from external knowledge, Gome achieves a state-of-the-art 35.1\% any-medal rate on MLE-Bench with a restricted 12-hour budget on a single V100 GPU. Scaling experiments across 10 models reveal a critical crossover: with weaker models, tree search retains advantages by compensating for unreliable reasoning through exhaustive exploration; as reasoning capability strengthens, gradient-based optimization progressively outperforms, with the gap widening at frontier-tier models. Given the rapid advancement of reasoning-oriented LLMs, this positions gradient-based optimization as an increasingly favorable paradigm. We release our codebase and GPT-5 traces at https://github.com/microsoft/RD-Agent.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑