Reasoning as Gradient: Scaling MLE Agents Beyond Tree Search

TL;DR

This paper introduces Gome, a gradient-based optimization approach for LLM agents in machine learning engineering, demonstrating superior performance over traditional tree search methods as models improve.

Contribution

Gome operationalizes gradient-based optimization for LLM agents, enabling more efficient reasoning beyond tree search, with state-of-the-art results on MLE-Bench.

Findings

Gome achieves 35.1% any-medal rate on MLE-Bench within 12 hours.

Gradient-based optimization outperforms tree search as model reasoning improves.

Scaling experiments show a crossover point where gradient methods become superior.

Abstract

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.