HEP-CoPilot: Innovative Application of Multi-Agent RAG Framework in Particle Physics Research

The high-energy physics field is undergoing an AI-driven paradigm shift in research. Facilities like the LHC generate massive amounts of data, and physicists face challenges in integrating heterogeneous information. HEP-CoPilot is a retrieval-augmented multi-agent AI framework for high-energy physics. It integrates text, structured data, and image information, and through RAG technology and multi-agent collaboration, it enables automatic comparison and reasoning of physical constraints across literature, opening up new possibilities for particle physics research.

In modern particle physics, searches for Beyond the Standard Model (BSM) have produced an explosive growth of literature, each containing textual descriptions, numerical tables, and exclusion limit curves. Physicists need to manually retrieve dozens of papers to extract and compare data, which is inefficient and error-prone. Taking the CMS experiment as an example, the number of papers on new physics searches increases year by year. Different papers have large differences in analysis methods and assumptions, leading to systematic biases in direct comparisons.

The core of HEP-CoPilot is a unified multi-modal retrieval and reasoning architecture, integrating three types of data sources: academic text, HEPData structured data, and paper charts/images. Text processing uses a RAG model with semantic understanding to identify associations between physical concepts; structured data processing directly parses HEPData to reconstruct experimental constraints; the multi-agent collaboration mechanism deploys professional agents to form a complete analysis chain.

HEP-CoPilot extracts numerical values from exclusion limit graphs through image reconstruction technology: optical recognition corrects the charts, and combines legends and axes to convert them into physical parameter constraints, which requires understanding the physical meaning of the charts (e.g., 95% confidence exclusion limits). Cross-modal alignment uses a shared semantic embedding space to achieve unified indexing and retrieval of text, numerical, and image information, with parallel searches and comprehensive ranking of results.

The capabilities of HEP-CoPilot were validated through BSM searches in the CMS experiment: 1. Retrieving 12 papers related to supersymmetric particle mass limits in seconds and extracting mass lower bounds (which takes humans hours); 2. Parsing HEPData to reconstruct exclusion limit curves, supporting queries with arbitrary precision; 3. Identifying compatibility between different analyses, marking systematic differences, and generating comprehensive constraints (considering assumptions, statistical methods, etc.).

HEP-CoPilot represents AI penetrating the core of scientific research, advancing from auxiliary literature management to scientific reasoning. Its domain-specific knowledge base and reasoning process enable deep understanding of professional content, providing a referenceable multi-modal RAG + multi-agent collaboration paradigm for other data-intensive disciplines (astronomy, materials science, bioinformatics).

HEP-CoPilot marks the entry of particle physics into the "AI Collaboration Era". Automated literature navigation and structured evidence integration enhance researchers' ability to handle complex information. It demonstrates how AI can amplify scientists' cognitive abilities without replacing human judgment—this is the core value of technology empowering scientific research.