AI-driven model discovery needs a test the builder doesn’t control.

Carbon@carbonphysicsai

AI-driven model discovery needs a test the builder doesn’t control. This study found agents improving on reproduced published-best results across 20 scientific ML benchmarks. The researchers separated model development, held-out testing and auditing to make those gains meaningful. That’s closely aligned with what we’re building at Carbon: give miners room to experiment, then evaluate their models against the requirements of a specific engineering problem.

Quoting @dair_ai

Great paper from Anthropic on AI agents doing ML-for-science research. (bookmark it) If you build AI-scientist systems, this one is worth your time. They propose Agent4Sci, which gives frontier coding agents 20 scientific ML benchmarks across 11 domains and asks them to beat the published best model. They beat it on all 20. Claude Opus improves on the published best by 10.4% on average, and no run was flagged for misconduct. The gains come almost entirely from standard ML-engineering tricks. The trajectories show agents misjudging what is novel, leaving compute unused, and settling into small local sweeps. Adding three lightweight skills for research ideation, experiment management, and scheduled checklists raises Claude Opus from 10.4% to 20.1% and Codex from 4.4% to 8.1%, with more varied and more domain-relevant changes. The models stay the same, and the research scaffolding roughly doubles the gain. arxiv.org/abs/2610.28763 Chat with Paper: academy.dair.ai/

Open on X ↗