Qiushi AI discovers new optical mechanism

Shuxing Yang and colleagues posted a preprint on April 29 describing an AI system they said autonomously discovered and experimentally validated a previously unreported optical mechanism on a real laboratory platform. The paper, “End-to-end autonomous scientific discovery on a real optical platform,” identifies the system as the Qiushi Discovery Engine and says it carried out long-horizon research with thousands of model and tool actions. The authors said the engine first reproduced an existing optics result and then moved to an open-ended investigation. In that second phase, they said, the system proposed and validated what they call “optical bilinear interaction,” which they describe as structurally analogous to a core operation in Transformer attention. ### What exactly did the researchers say the system discovered? The April 29 arXiv paper says Qiushi Discovery Engine proposed “optical bilinear interaction” and then experimentally validated it on a physical optical setup. The authors describe that mechanism as a physical interaction “structurally analogous” to a core operation used in Transformer attention, the computation that helps modern language models weight relationships between inputs. (arxiv.org) The authors did not frame the result as a new Transformer algorithm. Instead, they presented it as an optical mechanism in hardware and said the analogy to attention points to a possible route toward optical pairwise computation. That claim appears in the paper’s abstract and has not yet been tested through peer review. ### How much work did the engine do before reaching that result? The paper says the open-ended study consumed 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts. (arxiv.org) Those figures are among the clearest indicators of how the authors want readers to understand the system: not as a single prompt, but as an agent running a long sequence of reasoning, measurement and revision steps. The same abstract says Qiushi Engine used “nonlinear research phases,” “Meta-Trace memory” and a “dual-layer architecture” to keep the investigation on track over long runs. The arXiv abstract does not, by itself, provide full implementation detail for each of those components, but it does present them as the system’s organizing structure. ### Did the system only make a new claim, or did it also reproduce known science? (arxiv.org) The authors say Qiushi Discovery Engine first reproduced a published transmission-matrix experiment on a different platform before moving to discovery mode. They also say it converted an abstract coherence-order theory into experimental observables and produced, to their knowledge, the first observation of that class of coherence-order structure. That sequence matters because the paper presents reproduction and theory-to-measurement translation as part of the same research workflow. In the authors’ account, the engine was not limited to literature search or hypothesis generation; it was also tied to measurements on a real optical system. ### Who wrote the paper and where was the work done? The author list on the PDF 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 include Zhejiang University and its Hangzhou, Jinhua and Shaoxing institutes, China Jiliang University, EPFL in Lausanne, and Hangzhou City University. The corresponding authors are Yihao Yang and Hongsheng Chen. The arXiv record lists the paper under artificial intelligence and optics and shows it was submitted on April 29, 2026. ArXiv describes itself as an open-access archive and not a peer-reviewed journal, which means the claims are public but not yet journal-validated. ### What should readers be careful about when reading the claim? The arXiv abstract says, “To our knowledge,” this is the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism. (arxiv.org) That wording is the authors’ characterization, not an independent assessment. The paper also says the optical bilinear interaction “suggests a route” toward high-speed, energy-efficient optical hardware for pairwise computation. (arxiv.org) That is a forward-looking statement in the preprint, and the paper does not establish commercial hardware performance in the abstract alone. The next concrete checkpoint is peer review. As of May 15, 2026, the work is available as arXiv:2604.27092, with Yihao Yang and Hongsheng Chen listed for correspondence and the full preprint posted for public scrutiny. (arxiv.org)