Qiushi Discovery Engine made 3,242 LLM calls

Researchers at Zhejiang University and collaborators said on May 14 that their Qiushi Discovery Engine ran a long-horizon research program on a real optical platform and logged 3,242 large language model calls in the process. (arxiv.org) The claim matters because it is not just about an AI model proposing ideas in software. In a paper posted to arXiv on April 29, the team said the system interacted with physical experiments, used 1,242 tool calls, wrote 163 research notes and 44 scripts, and consumed 145.9 million tokens during an open-ended study. (arxiv.org) The system, called Qiushi Discovery Engine, was described as an “LLM-based agentic system for end-to-end autonomous scientific discovery on a real optical platform.” The authors listed 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, with affiliations including Zhejiang University, EPFL and other institutions in China. (arxiv.org) What the team says it achieved breaks into two parts. First, the paper said Qiushi autonomously reproduced a published transmission-matrix experiment on a different platform and translated an abstract coherence-order theory into experimental observables, which the authors said produced the first observation of that class of coherence-order structure. Second, in the open-ended run highlighted in the May 14 discussion, the system proposed and experimentally validated what the authors called “optical bilinear interaction.” (arxiv.org) That optical bilinear interaction is the center of the thread’s attention link. The paper said the mechanism is “structurally analogous to a core operation in Transformer attention,” tying the lab result to the architecture behind modern large language models. The authors said the finding points to a possible route toward optical hardware for pairwise computation, though that is their interpretation rather than an independently verified engineering result. (arxiv.org) The paper also frames the work as a first. The authors wrote that, to their knowledge, this is the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism in a real physical system. That wording is the researchers’ claim in the paper; it is not, by itself, outside confirmation from the broader field. (arxiv.org) Methodologically, the team said Qiushi uses nonlinear research phases, a memory system they call Meta-Trace, and a dual-layer architecture to keep a research trajectory stable across many reasoning, measurement and revision steps. That is the technical explanation the authors offer for how the system stayed on track over thousands of model-mediated actions instead of acting like a short-horizon lab assistant. (arxiv.org) One caution is that the public evidence I could verify directly comes from the arXiv paper, not from the specific May 14 X thread cited in the prompt. Search results did not surface that post reliably enough for direct quotation. What is verifiable is that the paper was submitted on April 29, 2026, names Yihao Yang and Hongsheng Chen as corresponding authors, and contains the numbers and claims now circulating in social posts about the project. (arxiv.org) The next place to watch is the paper record itself. As of the arXiv entry I reviewed, the work appears as version 1 of arXiv:2604.27092, posted on April 29, 2026, with the authors inviting scrutiny through the public manuscript and its disclosed experimental summary. (arxiv.org)