End-to-end autonomous scientific discovery on a real optical platform

A team of 13 researchers led by Yihao Yang at Zhejiang University has unveiled the Qiushi Discovery Engine, a large language model (LLM)-based agentic system capable of end-to-end autonomous scientific discovery on a real optical platform. Unlike prior AI assistants that merely execute predefined steps, Qiushi Engine combines nonlinear research phases, Meta-Trace memory, and a dual-layer architecture to maintain stable long-horizon research trajectories across thousands of LLM-mediated reasoning, measurement, and revision actions. The system first demonstrated its capability by autonomously reproducing a published transmission-matrix experiment on a non-original platform, then converted an abstract coherence-order theory into experimental observables, claiming the first observation of that class of structure.

In an open-ended investigation consuming 145.9 million tokens, 3,242 LLM calls, 1,242 tool calls, 163 research notes, and 44 scripts, Qiushi Engine proposed and experimentally validated a physical mechanism called optical bilinear interaction—a structure mathematically analogous to the core operation in Transformer attention. This AI-discovered mechanism suggests a path toward high-speed, energy-efficient optical hardware for pairwise computation. The researchers claim this is the first demonstration of an AI agentic system autonomously identifying and experimentally validating a nontrivial, previously unreported physical mechanism, marking a milestone for research-level autonomous agents.