Qiushi AI autonomously finds optical mechanism

Zhejiang University researchers said on April 29 that their Qiushi Discovery Engine autonomously found and experimentally validated a previously unreported optical mechanism on a real laboratory platform. The claim appears in an arXiv preprint, “End-to-end autonomous scientific discovery on a real optical platform,” submitted by Shuxing Yang, Fujia Chen, Rui Zhao and co-authors. The paper says the system first reproduced a published transmission-matrix experiment on a different platform and then moved to an open-ended investigation. In that second phase, the authors said, the agent proposed and validated what they call “optical bilinear interaction.” ### What exactly did the researchers say the system discovered? The April 29 paper says Qiushi Discovery Engine “proposes and experimentally validates optical bilinear interaction,” which it describes as a physical mechanism structurally analogous to a core operation in Transformer attention. The authors said the mechanism was previously unreported and was supported by experimental evidence gathered on a real optical setup. (arxiv.org) The same paper says the mechanism could point to “high-speed, energy-efficient optical hardware for pairwise computation.” That forward-looking claim comes from the authors’ abstract and is presented there as a possible application rather than a demonstrated product. ### How much work did the system do before reaching that result? The arXiv abstract says the open-ended study used 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts. (arxiv.org) Those figures are the clearest accounting in the public record of how much machine-driven iteration the workflow involved. The paper also says the system relied on “Meta-Trace” memory and a dual-layer architecture to sustain long-horizon research trajectories across repeated reasoning, measurement and revision steps. (arxiv.org) The authors frame that setup as the way the agent stayed on task through thousands of actions in the lab. ### Did the paper claim a first? The authors wrote that no earlier large language model-based agent had demonstrated end-to-end autonomous discovery in a real physical system that produced a nontrivial result backed by experiment. (arxiv.org) They also wrote that this was, “to our knowledge,” the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism. That language is the authors’ characterization in a preprint, not a journal ruling. The paper was posted to arXiv, which hosts manuscripts before formal peer review. ### Who is behind the work? The author list on the preprint names 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. (arxiv.org) The affiliations listed include Zhejiang University, China Jiliang University, EPFL and Hangzhou City University. The paper names Yihao Yang and Hongsheng Chen as corresponding authors. (arxiv.org) It lists the submission date as April 29, 2026. ### What should readers watch next? The next concrete checkpoint is peer review. As of May 15, 2026, the work is publicly available as arXiv:2604.27092, and the most detailed public evidence for the claim is the preprint’s abstract and PDF. (arxiv.org) Any follow-up to watch for would be a journal submission, independent replication of the optical bilinear interaction result, or additional disclosures from Yihao Yang, Hongsheng Chen and their co-authors about the experimental setup and validation record. (arxiv.org)