PhysMaster: Building an Autonomous AI Physicist for Theoretical and Computational Physics Research

TL;DR

PhysMaster presents an autonomous AI physicist that unifies abstract theoretical reasoning with executable computation, anchored by the LANDAU Layered Academic Data Universe to enhance reliability. It employs Monte Carlo Tree Search (MCTS) within a Clarifier–local-library–RAG loop to manage ultra-long-horizon physics tasks, enabling end-to-end research from query to report in a reproducible, traceable manner. The paper demonstrates three core capabilities—acceleration, automation, and autonomous discovery—via end-to-end pipelines in lattice QCD, ab initio quantum chemistry, and quantum many-body physics, plus additional autonomous studies in TDEs and semi-leptonic decays, achieving substantial reductions in manual effort and improved robustness. While showing strong promise for AI-driven scientific workflows, the authors acknowledge limitations in symbolic reasoning, potential hallucinations in critics, and the need for broader verification and theorem-proving integration, outlining a path toward more reliable, cross-domain AI scientific agents. Overall, PhysMaster demonstrates a viable route toward AI-led, end-to-end discovery pipelines in physics with significant practical impact for accelerating theoretical and computational research.

Abstract

Advances in LLMs have produced agents with knowledge and operational capabilities comparable to human scientists, suggesting potential to assist, accelerate, and automate research. However, existing studies mainly evaluate such systems on well-defined benchmarks or general tasks like literature retrieval, limiting their end-to-end problem-solving ability in open scientific scenarios. This is particularly true in physics, which is abstract, mathematically intensive, and requires integrating analytical reasoning with code-based computation. To address this, we propose PhysMaster, an LLM-based agent functioning as an autonomous theoretical and computational physicist. PhysMaster couples absract reasoning with numerical computation and leverages LANDAU, the Layered Academic Data Universe, which preserves retrieved literature, curated prior knowledge, and validated methodological traces, enhancing decision reliability and stability. It also employs an adaptive exploration strategy balancing efficiency and open-ended exploration, enabling robust performance in ultra-long-horizon tasks. We evaluate PhysMaster on problems from high-energy theory, condensed matter theory to astrophysics, including: (i) acceleration, compressing labor-intensive research from months to hours; (ii) automation, autonomously executing hypothesis-driven loops ; and (iii) autonomous discovery, independently exploring open problems.

Figures (10)

• Figure 1: The Core Features of PhysMaster
• Figure 2: The MAS Architechture and Workflow of PhysMaster
• Figure 3: The MCTS Exploration and Node Info
• Figure 4: Comparison of the nonlocal two-point correlator ratios used to extract the bare quasi-TMD wave-function matrix elements. Results obtained using the original analysis of Ref. Tan:2025ofx are compared with those produced by the automated PhysMaster pipeline, for $P_z=1.47~\mathrm{GeV}$, $z=\{0,2\}a$, and $b=3a$. The agreement demonstrates the reliability of PhysMaster in reproducing established lattice fitting results.
• Figure 5: Coordinate-space distributions of the bare quasi-TMD wave-function matrix elements at $P_z=1.47~\mathrm{GeV}$ and transverse separation $b=3a$. Left and right panels show the real and imaginary parts, respectively. Results from Ref. Tan:2025ofx and PhysMaster are consistent within statistical uncertainties over the full range of longitudinal separations.