Recoleta
GitHub
Source note

HepScript: A Dual-Use DSL for Human-AI Collaborative Data Analysis Workflows in High-Energy Physics

Authors
Junkun Jiao, Tong Liu, Ke Li, Weimin Song, Yipu Liao, Bolun Zhang, et al.
Source
arXiv
Published
Collection
Software Intelligence
Domain Specific LanguageCode GenerationLLM AgentsScientific WorkflowsHuman AI Collaboration
Open arXivMarkdown source

Summary

HepScript is a Ruby-embedded DSL that lets physicists and LLM agents specify BESIII high-energy physics analysis workflows, then generates BOSS and ROOT code. Its main claim is that a constrained DSL makes agent code generation more reliable and cuts human coding work.

Problem

  • HEP analyses use large datasets and experiment-specific software, so direct LLM code generation can fail on domain rules, long workflows, and low-level APIs.
  • BESIII workflows need a bridge between physics intent, including particle reconstruction and selection cuts, and production BOSS/ROOT code.
  • The problem matters because analysis code is repetitive, error-prone, and hard for agents to generate safely without a narrow action space.

Approach

  • HepScript expresses dataset preparation, base selection, advanced selection, visualization, and statistical analysis in a constrained Ruby syntax.
  • A processor translates HepScript into target code for BOSS, ROOT, shell scripts, Python, or C++ snippets.
  • The processor uses templates for stable code, Ruby translators for complex syntax, and LLM calls for analysis-dependent tasks, including cascade decay logic and ROOT scripts.
  • LLMs generate HepScript from papers using a long prompt with one complete workflow example and YARD API documentation; retry loops feed processor errors back to the model.

Results

  • Human-written HepScript for 45 BESIII papers generated 63 BOSS algorithm packages; with the LLM-assisted component disabled, all 63 compiled without errors.
  • Across two full case studies, HepScript reduced human-written analysis code by 93% by removing BOSS boilerplate and repeated ROOT plotting code.
  • For LLM-generated HepScript over 72 BOSS packages, DeepSeek-R1 reached 47.3% initial success, 87.8% after one retry, and 94.6% after three retries.
  • GLM-4.7 reached 43.2% initial success, 90.5% after one retry, and 95.9% after three retries on the same 72-package task.
  • In the variable-storage subtask for human-written specs, DeepSeek-V3, GPT-4o, GLM-4.7, and Qwen3-Max started at 93.8-96.9% success; all tested models reached 98.5% after one retry.
  • Two ROOT case studies executed end-to-end and reproduced figures from the original BESIII papers; the authors state these are pipeline checks, not official physics results.

Link