AI agents ran autonomous lab experiments

Zhejiang University researchers said in an April 29, 2026 preprint that an AI agent system called Qiushi Discovery Engine carried out end-to-end experiments on a real optical platform and validated what the authors described as a previously unreported physical mechanism. (arxiv.org) The paper, posted to arXiv, said the system did not just analyze data or suggest hypotheses. It designed, ran and revised experiments through repeated interactions with laboratory tools and software. The authors said the result went beyond reproducing a known setup: the system also proposed and experimentally validated what they called an “optical bilinear interaction.” The preprint was authored by Shuxing Yang, Fujia Chen, Rui Zhao, Junyao Wu, Yize Wang, Haiyao Luo, Ning Han, Qiaolu Chen, Yuze Hu, Wenhao Li, Mingzhu Li, Hongsheng Chen and Yihao Yang. The affiliations listed in the paper include Zhejiang University, China Jiliang University, EPFL and Hangzhou City University. The paper was submitted to arXiv on April 29, 2026. ### How autonomous was the system in the lab? The authors said Qiushi was built as an “LLM-based agentic system for end-to-end autonomous scientific discovery on a real optical platform.” In the paper’s description, the engine maintained long research trajectories through what the team called nonlinear research phases, a Meta-Trace memory system and a dual-layer architecture. (arxiv.org) The abstract says those components let the system keep working across “thousands of LLM-mediated reasoning, measurement and revision actions.” That matters because the claim here is not that a model answered a physics question in text, but that it used repeated model calls and tool calls to operate within an experimental loop. (arxiv.org) ### What did the agent actually do before the new result? The first reported benchmark was a published transmission-matrix experiment. The paper said Qiushi reproduced that experiment on a non-original platform rather than on the original apparatus used in the reference study. (arxiv.org) The same abstract said the system also converted an abstract coherence-order theory into experimental observables and provided what the authors called the first observation of that class of coherence-order structure. That step appears in the paper as an intermediate demonstration before the open-ended discovery claim. (arxiv.org) ### What was the new optical mechanism? The paper said the open-ended study led Qiushi to propose and experimentally validate “optical bilinear interaction.” The authors described that mechanism as structurally analogous to a core operation in Transformer attention. (arxiv.org) The authors said that analogy suggests a route to “high-speed, energy-efficient optical hardware for pairwise computation.” That is the paper’s framing of why the mechanism could matter beyond the immediate optics experiment. ### How much compute and tooling did the run use? (arxiv.org) The most concrete numbers in the paper come from the open-ended study log. The authors said the run involved 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts. Those figures are central to the paper’s claim that the system sustained a long-horizon workflow rather than a single prompt-response exchange. (arxiv.org) The abstract presents them as the record of the discovery process that produced the optical bilinear interaction result. ### How strong is the claim, and where can readers check it? The authors said the work is, “to our knowledge,” the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism. That wording is the paper’s own claim, not an independent peer-reviewed verdict. arXiv hosts the manuscript as paper 2604.27092, where the author list, affiliations and abstract are publicly available. (arxiv.org) April 29, 2026 is the key date for the underlying report, and arXiv paper 2604.27092 is the place to watch for revisions, author updates and any later journal submission details. (arxiv.org)