Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications
Authors: Yiming Cheng, Chris Low, and Binrui Huang (The University of Chicago); Mengshi Zhang and Zihao Chen (tensorblock); Kate Keahey (Argonne National Laboratory, University of Chicago); and Paul Marshall and Kexin Pei (The University of Chicago)
Abstract: Environment configuration remains a fundamental yet overlooked challenge in software engineering, requiring precise management of operating systems, dependencies, and hardware constraints. Existing LLM-based software engineering workflows often rely on predefined Dockerfiles or manual scripts, or exclude repositories with difficult environments, limiting generalization.
We present ENVAGENT, a multi-LLM-agent workflow that automatically synthesizes, builds, and verifies hardware-aware Dockerfiles from repositories. We further introduce ENVAGENT+, a resource-provisioning agent that infers computational requirements and allocates compatible resources, and ENVBENCH, a benchmark for evaluating environment reconstruction fidelity.
We evaluate ENVAGENT on 54 repositories from Trovi and Multi-SWE-bench across seven programming languages. Results show that ENVAGENT improves runnable counts by 16.2% and FCS scores by 33.91% over baselines, enabling reliable evaluation of repositories that existing approaches fail to configure.
Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing
Back to Full Workshop Archive Listing