Qiushi Discovery Engine runs 3,242 LLM calls

Zhejiang University researchers said in an April 29 arXiv paper that their Qiushi Discovery Engine ran a long-horizon research process on a real optical platform and experimentally validated what they described as a previously unreported physical mechanism. The preprint said the system completed 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts during an open-ended study. The authors said the mechanism, which they call “optical bilinear interaction,” is structurally analogous to a core operation in Transformer attention. The paper was posted as arXiv:2604.27092 and lists Yihao Yang and Hongsheng Chen as corresponding authors. ### Where does the 3,242-call figure come from? The number 3,242 comes from the paper’s abstract, not just from social-media discussion. In the abstract, the authors wrote that the open-ended study involved 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes and 44 scripts. The same abstract says those calls were part of a broader research loop that included reasoning, measurement and revision actions on a physical optics setup. (arxiv.org) That matters because the claim is not that a model generated a single idea in chat, but that an agentic system repeatedly interacted with software tools and lab instrumentation over a long sequence of steps. ### What, exactly, do the authors say the system discovered? (arxiv.org) The paper says Qiushi “proposes and experimentally validates optical bilinear interaction,” which it describes as a physical mechanism analogous to a core operation in Transformer attention. The authors add that the mechanism could point to optical hardware for pairwise computation, though that is framed in the paper as a possible route rather than a shipped technology. (arxiv.org) The authors also make a narrower claim alongside the discovery claim. In a separate result, they say the system reproduced a published transmission-matrix experiment on a different platform and converted an abstract coherence-order theory into experimental observables, producing what they describe as the first observation of that class of coherence-order structure. ### Did the AI work alone in the lab? The paper describes an “LLM-based agentic system” operating on a real optical platform, not a robot scientist working without any human-built environment. (arxiv.org) The listed institutions include Zhejiang University, China Jiliang University, Hangzhou City University and EPFL, and the authors say the engine interacted with physical experiments through a designed system architecture. The abstract says that architecture combined nonlinear research phases, a memory system called Meta-Trace and a dual-layer design to keep the investigation on track across thousands of actions. (arxiv.org) The paper therefore presents autonomy as end-to-end operation within a human-built experimental and software stack, with the discovery claim resting on the agent’s ability to iterate through question-setting, measurement and revision. (arxiv.org) ### How strong is the “first” claim? The strongest language in the paper is the authors’ own. They write, “To our knowledge,” the work is the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism. That phrasing is a claim by the authors in a preprint, not an independent certification by a journal or outside reviewer. The paper was submitted to arXiv on April 29, 2026, and the arXiv record shown in the abstract page does not indicate journal publication. (arxiv.org) That leaves peer review and outside replication as the next formal tests of the result. ### Why did the story spread on May 14? May 14 social posts appear to have amplified one line from the abstract: the 3,242 LLM calls and 1,242 tool calls. Those figures are unusually concrete and easy to circulate, and they compress a 25-page preprint into a single operational snapshot. (arxiv.org) The underlying document, however, dates to April 29 rather than May 14. Readers trying to place the timeline should treat May 14 as a social-media pickup date and April 29, 2026 as the paper’s public posting date on arXiv. (arxiv.org) ### What happens next for this result? The arXiv entry lists Yihao Yang and Hongsheng Chen as correspondence contacts, and the paper remains available as a preprint. Any next step that would change the standing of the claim — journal acceptance, outside replication or a revised version — would likely appear first through an updated paper record or author communication. (arxiv.org 1) (arxiv.org 2)