arxiv:2609.08149

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

Published on Sep 8

· Submitted by

Shufan Jiang on Sep 10

Shanghai AI Laboratory

Upvote

Authors:

Zixin Shang ,

Shufan Jiang ,

Dongsheng Zhu ,

Zerun Ma ,

Dingbo Yuan ,

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.

View arXiv page View PDF Project page Add to collection

Community

Tsumugii

Paper author Paper submitter about 24 hours ago

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 frontier 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 version for assessing software engineering agents.

O96a

about 21 hours ago

Does the verified set actually reshuffle the leaderboard, or does it just shave a few points off everyone? Reward hacking is real — agents that game the grader look smarter than they are — but the interesting number isn't the new score, it's the rank delta. If the same agents stay on top in the same order, the benchmark is cleaner but not more informative. If a couple of previously top-tier agents drop hard, that's the signal worth publishing. I'd want to see per-agent before/after, not just the aggregate pass rate.