- Name: Zhuo-Yang Song
- Email: zhuoyangsong@stu.pku.edu.cn
- ORCID: 0009-0001-9727-7908
- Introduction: Undergraduate student at the School of Physics, Peking University, Class of 2026. Recipient of the School of Physics Seagull Scholarship and 4th place in the 38th CPHO. Passionate about the intersection of physics and artificial intelligence.
I study the physics of AI systems — using tools from statistical mechanics and field theory to understand the macroscopic behavior of large language models and LLM-driven agents.
My central question: Can we build a predictive, quantitative theory for AI systems the way physicists build theories for natural systems?
-
Detailed Balance in LLM-Driven Agents — Discovered that LLM-driven agent transitions satisfy detailed balance, revealing an equilibrium-like structure in generative dynamics. To our knowledge, the first macroscopic physical law found in LLM generation that is architecture-independent. [arXiv:2512.10047] [Code]
-
A Theory of LLM Information Susceptibility — Proposed a susceptibility framework characterizing when LLM intervention improves optimization and when it does not, drawing on statistical physics to derive predictive constraints for agentic system design. [arXiv:2603.23626]
-
IdeaSearchFitter: Iterated Agent for Symbolic Regression — An LLM-driven evolutionary framework for symbolic regression that outperforms PySR and AI-Feynman on the Feynman SR Database. Applied to Parton Distribution Functions in high-energy physics. [arXiv:2510.08317] [Website]
-
Scalable Quantum State Preparation via LLM-Driven Discovery — LLM-assisted quantum circuit design validated on the Zuchongzhi quantum processor; first scalable ansatz for 2+1d scalar field theories. [arXiv:2505.06347]
-
Explainable AI-assisted Optimization for Feynman Integral Reduction — LLM + genetic algorithm approach to optimize integration-by-parts reduction of Feynman integrals. [arXiv:2502.09544]
- PHYBench — A 500-problem benchmark for evaluating LLM physical reasoning, with a novel Expression Edit Distance (EED) metric. Even Gemini 2.5 Pro scores only 36.9% vs. human experts' 61.9%. Accepted at NeurIPS 2025 Datasets & Benchmarks. [arXiv:2504.16074] [Website]
- Thickness model for viscous impinging liquid sheets — Journal of Fluid Mechanics, 2025
- Influence of Planetary Rotation on Supersonic Flow of Lava Planets: A Two-Dimensional Horizontal Model Analysis — The Astrophysical Journal
- Email: zhuoyangsong@stu.pku.edu.cn
- ORCID: 0009-0001-9727-7908