End-to-end autonomous scientific discovery on a real optical platform
Researchers have developed the Qiushi Discovery Engine, an AI system based on large language models that autonomously conducts scientific discovery on a real optical platform. The system replicated a known experiment and discovered a new physical mechanism called optical bilinear interaction, which has structural similarities to Transformer attention mechanisms in AI. This marks the first demonstration of an AI agent independently identifying and experimentally validating a nontrivial, previously unknown physical phenomenon.
- ▪The Qiushi Discovery Engine is an LLM-based agentic system designed for end-to-end autonomous scientific discovery in optics.
- ▪It autonomously reproduced a published transmission-matrix experiment and derived experimental observables from abstract coherence-order theory.
- ▪In a large-scale study, the engine proposed and experimentally validated optical bilinear interaction, a previously unreported physical mechanism.
- ▪The discovery suggests potential applications in high-speed, energy-efficient optical computing hardware.
- ▪To the authors' knowledge, this is the first instance of an AI system autonomously discovering and verifying a novel physical mechanism in a real-world experimental setting.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2604.27092 (cs) [Submitted on 29 Apr 2026] Title:End-to-end autonomous scientific discovery on a real optical platform Authors:Shuxing Yang, Fujia Chen, Rui Zhao, Junyao Wu, Yize Wang, Haiyao Luo, Ning Han, Qiaolu Chen, Yuze Hu, Wenhao Li, Mingzhu Li, Hongsheng Chen, Yihao Yang View a PDF of the paper titled End-to-end autonomous scientific discovery on a real optical platform, by Shuxing Yang and 12 other authors View PDF Abstract:Scientific research has long been human-led, driving new knowledge and transformative technologies through the continual revision of questions, methods and claims as evidence accumulates.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.