Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Zhang, Yifei, Yang, Xu, Yang, Xiao, Xian, Bowen, Li, Qizheng, Fang, Shikai, Li, Jingyuan, Wang, Jian, Xu, Mingrui, Liu, Weiqing, Bian, Jiang
Format:	Preprint
Published:	2026
Subjects:	Machine Learning Artificial Intelligence
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866911586396930048
author	Zhang, Yifei Yang, Xu Yang, Xiao Xian, Bowen Li, Qizheng Fang, Shikai Li, Jingyuan Wang, Jian Xu, Mingrui Liu, Weiqing Bian, Jiang
author_facet	Zhang, Yifei Yang, Xu Yang, Xiao Xian, Bowen Li, Qizheng Fang, Shikai Li, Jingyuan Wang, Jian Xu, Mingrui Liu, Weiqing Bian, Jiang
contents	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.
format	Preprint
id	arxiv_https___arxiv_org_abs_2603_01692
institution	arXiv
publishDate	2026
record_format	arxiv
spellingShingle	Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search Zhang, Yifei Yang, Xu Yang, Xiao Xian, Bowen Li, Qizheng Fang, Shikai Li, Jingyuan Wang, Jian Xu, Mingrui Liu, Weiqing Bian, Jiang Machine Learning Artificial Intelligence 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.
title	Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search
topic	Machine Learning Artificial Intelligence
url	https://arxiv.org/abs/2603.01692