Research map / Multi-agent systems and games / LLM agents

ML-Agent: Reinforcing LLM Agents for Autonomous Machine Learning Engineering

ICML 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes learning-based agentic ML: a 7B Qwen2.5 agent trained with exploration-enriched SFT, step-wise PPO, and an ML-specific reward, which matches GPT-5-driven prompt-based agents on MLE-bench-style tasks at much lower cost.

Problem

Prompt-based LLM agents for autonomous ML engineering either lack capacity to learn from execution trajectories when small, or are costly when driven by large proprietary models; RL training of such agents is hampered by limited exploration, slow ML experiment feedback, and heterogeneous reward sign

Approach

Trains a 7B LLM agent with RL for ML engineering by (1) SFT on expert trajectories seeded with diverse sampled action ideas, (2) step-wise PPO that samples single action steps from a pre-collected expert state pool instead of full rollouts, and (3) a unified reward mapping errors to -1, neutral actions to 0, and normalized metric gains to scaled rewards.

Key results

• 7B ML-Agent averages 18.01% gain on held-in and 15.91% on held-out tasks in the ablation, versus DeepSeek-R1 (671B) and Qwen3-235B baselines that score lower in the main table
• ML-Agent reaches a 16.40% average held-out gain, close to GPT-5 with MLAB at 18.14% and AIDE at 20.95%, at under $0.01 per trajectory (claimed >20x cheaper than GPT-5 MLAB)
• Removing exploration-enriched fine-tuning drops held-out gain to -6.20%; episode-wise RL reaches only 2.86% held-out gain versus 15.91% for step-wise RL

Surprising findings

• Training the RL stage from Qwen2.5-7B-Instruct gives +13% held-in but -12% held-out gains, indicating poor generalization without exploration-enriched SFT
• The DeepSeek-R1-Distill-Qwen-7B initialization fails to generate valid actions due to distillation-induced format issues

Limitations

• Held-out evaluation covers only a subset of 10 MLE-bench tasks due to computational cost
• Training tasks chosen for fast execution, and results are reported with high variance across trajectories
• Performance gains on some held-out tasks remain negative or near zero (e.g., statoil, jigsaw)

Experimental setup

Datasets: MLAgentBench tasks (cifar-10, home-data-for-ml-course, spaceship-titanic, feedback-prize-english-language-learning, ogbn MLE-bench tasks (aerial-cactus, dogs-vs-cats, plant-pathology, nomad2018, ...)

Benchmarks: MLE-bench held-in tasks (cifar-10, house-prices, feedback) MLE-bench held-out tasks (10 tasks)

Metrics: performance gain (%) over initial script, avg@8 average cost per trajectory

Baselines: MLAB AIDE ML-Master Qwen2.5-7B-Instruct Qwen3-235B DeepSeek-R1 GPT-5 Gemini-2.5-Pro

Builds on: PPO MLAgentBench MLE-bench RLVR LLaMA-Factory veRL

Closest papers by meaning

• MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering (NeurIPS 2025 · similarity 0.93)
• Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.90)
• MLZero: A Multi-Agent System for End-to-end Machine Learning Automation (NeurIPS 2025 · similarity 0.89)
• AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML (ICML 2025 · similarity 0.89)
• From Words to Actions: Unveiling the Theoretical Underpinnings of LLM-Driven Autonomous Systems (ICML 2024 · similarity 0.89)
• Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems (NeurIPS 2026 · similarity 0.89)