AI Scientist via Synthetic Task Scaling

TL;DR

This paper introduces a synthetic environment generation pipeline for training AI agents in machine learning tasks, leading to improved performance on the MLGym benchmark by leveraging synthetic challenges and trajectories from a teacher model.

Contribution

The paper presents a novel pipeline for generating high-quality synthetic machine learning challenges, enabling effective training of AI agents and advancing automatic scientific discovery.

Findings

Synthetic tasks improve student model performance on MLGym

Student models trained with synthetic tasks show 9-12% AUP improvement

Pipeline verifies datasets against Huggingface API for quality assurance

Abstract

With the advent of AI agents, automatic scientific discovery has become a tenable goal. Many recent works scaffold agentic systems that can perform machine learning research, but don't offer a principled way to train such agents -- and current LLMs often generate plausible-looking but ineffective ideas. To make progress on training agents that can learn from doing, we provide a novel synthetic environment generation pipeline targeting machine learning agents. Our pipeline automatically synthesizes machine learning challenges compatible with the SWE-agent framework, covering topic sampling, dataset proposal, and code generation. The resulting synthetic tasks are 1) grounded in real machine learning datasets, because the proposed datasets are verified against the Huggingface API and are 2) verified for higher quality with a self-debugging loop. To validate the effectiveness of our synthetic tasks, we tackle MLGym, a benchmark for machine learning tasks. From the synthetic tasks, we sample trajectories from a teacher model (GPT-5), then use the trajectories to train a student model (Qwen3-4B and Qwen3-8B). The student models trained with our synthetic tasks achieve improved performance on MLGym, raising the AUP metric by 9% for Qwen3-4B and 12% for Qwen3-8B.