Qiushi Discovery Engine

LLM-based agentic system that performs end-to-end autonomous scientific discovery on a real free-space optical platform, combining nonlinear research phases, a Meta-Trace memory, and a dual-layer architecture to deliver experimentally supported nontrivial results.

   
Affiliation State Key Laboratory of Extreme Photonics and Instrumentation, Zhejiang University (Yang and Chen groups), with EPFL, China Jiliang University, and Hangzhou City University
First introduced 2026-04 (arXiv:2604.27092)
Lifecycle stages Multi-stage (Explore: literature interpretation, hypothesis generation, observable design; Execute: coding, simulation, physical measurement, data analysis; Express: figure construction, manuscript writing, critical review)
Autonomy level Fully autonomous — given an open-ended research theme, the system sustains thousands of LLM-mediated reasoning, measurement, and revision actions against the physical platform
Domain focus Optics / photonics — free-space optical platform with >2M-pixel SLM, scattering diffuser, and camera detection
Availability Unknown (no public repository disclosed in primary paper)

Approach

Dual-layer multi-agent architecture. The core research agent system runs four role-specialized agents — Lead Investigator (global planning, hypothesis formation), Method Builder (theory-to-method translation, algorithm design), Experimentalist (simulation, code execution, measurement, analysis), and Critical Reviewer (adversarial evaluation) — that operate across three nonlinear phases (Explore, Execute, Express), with each Agent Step occupying one of 12 role-phase configurations rather than following a fixed pipeline. A support research agent system supplies context-isolated sub-agents for history review, retrieval, hypothesis exploration, trajectory tracking, and evidence verification, returning compressed, task-relevant outputs through structured interfaces. Meta-Trace distils each Agent Step into a structured unit of scientific know-how (attempt, finding, supporting evidence, limitations, artifacts, next-agent objective), preserving long-horizon scientific state without flooding the active context. An infrastructure layer couples the agents to a free-space optical platform (~2,200,000 SLM configurations) and a digital execution environment.

Validation

Three progressively more demanding studies. (1) Autonomously reproduced a published transmission-matrix experiment on a non-original optical platform. (2) Converted an abstract coherence-order theory into experimentally testable transport observables and validated the predicted ordering relation using measured optical operators — claimed first experimental validation of this class of coherence-order structure. (3) An open-ended 206-step autonomous study over 1,288 minutes used 145.9 million tokens, 3,242 LLM calls, and 1,242 tool invocations, producing 163 research notes and 44 scripts.

Notable results

  • Identified and experimentally validated “optical bilinear interaction” — a previously unreported physical mechanism in which coherent scattering and square-law detection generate pairwise optical features, structurally analogous to bilinear query–key computation in Transformer attention.
  • Authors claim this is the first demonstration of an AI agentic system autonomously proposing and experimentally validating a nontrivial physical mechanism in a real experimental environment.
  • Suggests a route toward high-speed, energy-efficient optical hardware for pairwise computation.

Primary paper

Yang et al., “End-to-end autonomous scientific discovery on a real optical platform,” arXiv:2604.27092.

Other references

None yet.

Code

Not released.