Building autonomous AI physicists for frontier physics research

Advances in LLMs have led to agents with knowledge and operational capabilities comparable to human scientists, suggesting potential to assist, accelerate, and automate research. Physics, especially theoretical and computational physics, which requires integrating analytical reasoning, code-based computation, and profound domain expertise, is well suited for verifying the end-to-end research capabilities of AI scientists. Accordingly, we construct a general-purpose AI physicist PhysMaster, equipped with a layered academic knowledge base, adapted to the agent skill ecosystem, and adopting an adaptive exploration strategy that balances efficiency and exploration, enabling robust performance in ultra-long-horizon tasks; PhysMaster has been open-sourced. Meanwhile, we introduce PRL-Bench (Physics Research by LLMs), a benchmark with 100 tasks adapted from recent Physical Review Letters papers, covering astrophysics, condensed matter physics, high-energy physics, quantum information, and statistical physics. Evaluation across frontier models shows that failures are dominated by conceptual and formulaic errors, and that exploration and derivations remain unstable over long horizons. In addition, we develop domain-specialized AI scientists, including LQCD Master, which integrates Lattice QCD workflows and expert skills, enabling automated generation and submission of lattice computation scripts from concise physics goals.