An autonomous LLM agent has beaten every expert-built model on a standard materials science benchmark — without any domain-specific pretraining.
Researchers tested a general-purpose coding agent on the MatBench band-gap benchmark, a dataset of more than 100,000 crystals used to measure how well models predict a material's electronic properties from its structure. The agent finished ahead of all seventeen expert-designed models previously reported for the task. It got there not by discovering new science but by combining known techniques: element-pair features on message-passing edges and crystal space-group embeddings, both borrowed from existing work in crystal neural networks and broader machine learning. No novel methods, just better assembly.
That distinction matters. The result is less a story about AI creativity and more a story about AI as an optimizer — one that can survey a landscape of known tools and combine them more effectively than a human researcher working on a deadline. Materials science is an expensive, slow field; if agents can reliably compress months of model-tuning into an autonomous loop, the implications for drug discovery, battery research, and semiconductor design are real.
The authors are candid that the agent has limits, and the paper spends time mapping them — a useful corrective to the temptation to treat a benchmark win as a proof of general scientific capability.