SWE-Bench Pro Verified: A Reliable Benchmark for Software Engineering Agents

Pujun Zheng, Zixin Shang, Shufan Jiang, Wenhui Tian, Dongsheng Zhu, Zerun Ma, Dingbo Yuan, Qi Zhang

arXiv:2609.08149 · 2026-09-10 공개 · arXiv · PDF

software-engineering reward-hacking evaluation-methods swe-bench-pro task-refinement agent-assessment leakage-prevention

Abstract

SWE-Bench Pro has emerged as a standard benchmark for evaluating software engineering agents on challenging repository-level tasks. However, our analysis work show that its evaluation is undermined by two sources of unreliability: reward hacking, enabled by leakage of gold solutions or hidden evaluation information, and task quality issues, including misleading problem statements and improperly scoped tests. These issues can inflate benchmark performance and obscure agents' true coding ability. We present SWE-Bench Pro Verified, a verified version of SWE-Bench Pro that addresses both problems. Our approach combines anti-hacking safeguards that eliminate major leakage channels without disrupting normal agent functionality, with task refinement that minimally corrects inconsistencies within flawed instances. Evaluations on SWE-Bench Pro Verified reveal that some models perform substantially worse than previously reported, suggesting that existing results on SWE-Bench Pro may overestimate real software engineering capability. SWE-Bench Pro Verified offers a more trustworthy benchmark for assessing software engineering agents.

한국어 요약

한 줄 요약

SWE-Bench Pro Verified는 기존 SWE-Bench Pro의 신뢰도 문제를 해결한 소프트웨어 엔지니어링 에이전트 평가 벤치마크이다.

핵심 기여도

핵심 아이디어

기존 SWE-Bench Pro는 **repository-level task** 평가에 널리 사용되지만, **reward hacking** 과 **task quality issues** 가 평가 신뢰도를 저하시켰다. 이에 연구팀은 **SWE-Bench Pro Verified** 를 제안하여, **anti-hacking** 을 통해 **leakage channels** 를 차단하고, **task refinement** 를 통해 **misleading problem statements** 와 **improperly scoped tests** 를 수정했다. 특히, **LLM-assisted filtering** 을 통해 테스크를 선별하고, **인간 전문가의 최소 수정** 을 통해 테스크의 일관성을 유지하면서도 정확도를 향상시켰다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용