An AI pipeline completed a physics research project end-to-end, producing three original findings on altermagnetic piezomagnetism without fabricating citations.
Researchers designed a six-phase system that ingested 11,083 condensed-matter physics papers from arXiv, mapped the literature to conceive a research direction, reproduced published results to calibrate its methods, ran novel first-principles computations, and wrote a manuscript. The pipeline ran across 47 separate sessions that shared no in-memory state — only files on disk. It logged 2,162 literature-consultation events, meaning the agent kept checking its work against real sources rather than filling gaps with plausible-sounding numbers. Crucially, human intervention was limited to fixing reproduction failures, not setting scientific direction.
This matters because most AI research agents have been benchmarked in machine-learning sandboxes, where the system can verify its own outputs by running code. Physical science is harder: the toolchains are often poorly documented, and the only honest calibration comes from matching published experimental results. Previous unscaffolded agents tended to cite papers without actually confronting what those papers said — a polite way of describing confident hallucination. This pipeline's "grounding" mechanism forces numerical confrontation at checkpoints, and ablation tests confirm that removing it breaks the results.
Automated science is a field where the demos tend to outrun the rigor; a system that ships ablation studies alongside its manuscript is at least playing the game honestly.