Qiushi Engine autonomously discovers optical mechanism

Qiushi Discovery Engine is one of the clearest recent examples of an AI system moving beyond simulation and software-only benchmarks into a physical lab. In a paper posted to arXiv on April 29, 2026, researchers led by Shuxing Yang, Hongsheng Chen and Yihao Yang said the system autonomously carried out a long-horizon optics investigation on a real experimental platform and identified what they called “optical bilinear interaction,” a previously unreported mechanism that the team then experimentally validated. (arxiv.org) The basic claim is narrower than “AI did science alone” and more concrete than a generic autonomous-agent demo. The authors said Qiushi Engine handled an end-to-end research loop: proposing ideas, choosing and revising experiments, using lab tools, writing scripts and notes, and updating its direction over time while interacting with physical equipment rather than a simulator. They also said the system first reproduced a published transmission-matrix experiment and translated an abstract coherence-order theory into measurable observables before moving to the open-ended discovery phase. (arxiv.org) The scale of that open-ended run is part of why the result has drawn attention. The paper reports 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts during the study that produced the new mechanism. Those are the figures behind the social-media shorthand that the system made “3,000-plus” model calls and used “1,000-plus” tools. (arxiv.org) What the system says it found is an optical process the authors describe as structurally analogous to a core operation in Transformer attention. In plain terms, the claim is not that the lab built a full optical Transformer, but that it uncovered a physical interaction in light that resembles the bilinear form underlying pairwise computation in attention-style models. The paper says that resemblance could point to high-speed, energy-efficient optical hardware for pairwise computation, though that forward-looking implication is the authors’ interpretation rather than an independently verified product roadmap. (arxiv.org) The paper also makes a second, easier-to-overlook claim. Before the new mechanism, the system allegedly reproduced known work on a non-original platform and produced what the authors called the first observation of a class of coherence-order structure by converting theory into experimental observables. That matters because it frames the discovery result as coming after a sequence of setup, replication and theory-grounding steps rather than from a single one-shot prompt. (arxiv.org) The institutional context is also clear from the manuscript. The listed authors are affiliated primarily with Zhejiang University and related institutes in Hangzhou, with one co-author also affiliated with EPFL in Lausanne. Correspondence is directed to Yihao Yang and Hongsheng Chen. The paper was submitted on April 29, 2026, and as of the arXiv posting it was a preprint rather than a peer-reviewed journal article. (arxiv.org) What to watch next is straightforward. The immediate next test is outside commentary and replication: whether other optics researchers can reproduce optical bilinear interaction, whether the authors release more implementation detail, and whether the preprint advances to peer review. For now, the public record is the April 29 arXiv paper by Yang and colleagues, which is the primary source for the discovery claim and the reported usage metrics. (arxiv.org)