HuggingFace released an updated version of PhysicsIntern, a research assistant for physics problems, designed to be more interactive and collaborative. The tool allows researchers to steer the process, review reasoning, and engage at critical moments, rather than operating as an autonomous benchmark-runner. This shift reflects a broader goal of making the tool a true research collaborator. The new version is lighter and integrates with existing coding platforms like Codex, Claude Code, or Pi. It is now available on GitHub. The post details the evolution of the tool and its practical application in solving complex physics problems. Source: huggingface
The updated PhysicsIntern was tested on a real-world problem involving the development of a Monte Carlo method for solving the Helmholtz equation. The tool first reviewed existing literature, identified gaps, and proposed a research plan. It then executed the plan, only to discover that the original question was ill-posed. Instead of proceeding with an invalid approach, the tool suggested a revised method involving resolvent block inverse iteration. This approach was validated numerically on the unit interval, yielding accurate results for the first two eigenvalues. The tool also maintained a detailed record of the entire research process, including all derivations, checks, and decisions, stored in a git log with 21 commits. Source: huggingface
The original PhysicsIntern was designed to be fully autonomous, with a fixed pipeline of nine roles that handled tasks like problem decomposition, derivation, and verification. It was built to run on benchmarks like CritPt without human intervention, providing hard evidence that a structured, multi-agent approach could improve performance. The tool outperformed single-shot baselines, with notable gains for models like Kimi K2.6 and Gemini 3.1 Pro. However, researchers found the tool's rigid design unsuitable for real-world collaboration, prompting the development of the new version. Source: huggingface
