BuildBench: Benchmarking LLM Agents on Compiling Real-World Open-Source Software

Abstract

Automatically compiling open-source software is a prerequisite for many downstream security and software-engineering tasks. BuildBench is a benchmark designed to reflect practical compilation difficulty, featuring diverse, low-resource, and heterogeneous open-source projects, together with OSS-Build-Agent, an LLM agent that combines build-instruction retrieval, dependency inference, and context-aware code modification.

Publication
Transactions on Machine Learning Research (TMLR)
Hulin Wang
Hulin Wang
Ph.D. Candidate · LLM Agents for Vulnerability Discovery & Repair

Ph.D. candidate at ASU SEFCOM building LLM agents that find and fix real-world software vulnerabilities.