Intel / article

Automating ML for Science: Can Frontier Agents Climb Scientific Hills in the Wild?

Source: arxiv.org
Author: Ming Zhong, Stacy Li, Nicholas Carlini, Matthew Jagielski
Date: Published Oct 22, 2026

Why it matters

Frontier coding agents improved every published best, but through ML-engineering tricks rather than research insight, and fail through novelty miscalibration, resource under-use and early local-optimum collapse.

Key takeaways · AI-distilled

Agent4Sci tests whether coding agents can improve the published best models on 20 scientific ML benchmarks across 11 domains under in-the-wild conditions, analysing what kind of progress they make rather than ranking agents on a leaderboard.
Frontier agents beat the published best on every , with Claude Opus reaching a 10.4% mean relative improvement and no run flagged for academic misconduct, but the gains came almost entirely from ML-engineering tricks rather than research insight.
Trajectory analysis found three recurring limitations: novelty miscalibration, under-use of available resources, and fast collapse into local engineering sweeps.
Lightweight plug-in skills for research ideation, experiment management and scheduled checklists raised Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, with more diverse, domain-relevant interventions.

Terms in this piece · Glossary

benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Read the source arxiv.org