Qiushi AI runs 3,242 LLM calls, finds optical mechanism

Qiushi Discovery Engine, a system described by researchers at Zhejiang University and collaborators, was reported in late April as having autonomously run a long sequence of real optical experiments rather than simulations, then proposed and experimentally validated what the paper calls “optical bilinear interaction.” The core claim comes from an arXiv paper submitted on April 29, 2026, and echoed in a May 14 X post that highlighted experiment logs and run metrics. (arxiv.org) The headline numbers are unusually specific. In the paper’s open-ended study, the system used 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts. The authors say those steps were part of a long-horizon workflow in which the system reasoned, triggered measurements, revised plans and analyzed results while connected to a physical optical platform. (arxiv.org) What makes the claim notable is not just that the system automated lab work, but that the authors say it moved through multiple research stages on its own. The paper says Qiushi first reproduced a published transmission-matrix experiment on a different platform, then converted an abstract coherence-order theory into measurable observables, and finally carried out an open-ended investigation that led to the new mechanism. That sequence matters because it separates routine lab automation from a broader claim of autonomous scientific discovery. (arxiv.org) The “new optical mechanism” in the paper is called optical bilinear interaction. The authors describe it as structurally analogous to a core operation in Transformer attention, which is why the result has drawn attention beyond optics. In the paper, they say the mechanism could point toward optical hardware for pairwise computation, though that is presented as a possible route rather than a demonstrated product roadmap. (arxiv.org) The lab setting is also central to the story. The authors say the work was done on a “real optical platform,” and the paper frames that as distinct from prior agent systems that assist with predefined workflows or operate mainly in digital environments. The affiliations listed on the paper include Zhejiang University, EPFL and other collaborating institutions, with Yihao Yang and Hongsheng Chen listed as corresponding authors. (arxiv.org) There are still important limits on what can be said at this stage. The result appears in an arXiv preprint, not yet in a peer-reviewed journal based on the sources I found, and the strongest wording about being the “first” such demonstration comes from the authors themselves. That means the factual core is clear — a preprint reports an AI agent running real optical experiments and validating a previously unreported mechanism — but broader ranking claims remain author claims unless independently confirmed. (arxiv.org) The next concrete place to watch is the paper itself, arXiv:2604.27092, and any follow-up release or peer-reviewed publication from the Zhejiang University-led team. As of May 15, 2026, the publicly visible record is the April 29 arXiv submission and the May 14 social post highlighting logs and metrics from the run. (arxiv.org)