English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

SWE-Gate: A Benchmark Showing Passing Functional Tests Isn't Enough for Software Engineering Agents

Forum topic · 小凯 · 2026-09-07

Summary

SWE-Gate (arXiv:2509.04275) is a repository-level benchmark for software engineering agents that evaluates review constraint compliance alongside functional correctness. Built from real pull request review comments, it contains 303 repair instances across 75 open-source Python repositories. Each instance provides separate functional and constraint tests plus non-compliant and gold patches, enabling explicit separation of bug-fixing ability from review constraint compliance. Experiments with four LLM backends under general coding agent frameworks reveal a significant gap: of 644 patches that passed functional tests, 221 failed to satisfy the provided review constraints. The findings show functional-test-only evaluation overestimates agents' ability to meet the full requirements of repository-level repair tasks. The replication package (code, data, results) is available at https://github.com/DeepSoftwareAnalytics/SWE-Gate.

Paper Overview

Field: ML Authors: Xin He, Yanlin Wang, Mingwei Liu Published: 2026-09-06 arXiv: 2509.04275

Abstract

Repository-level software engineering benchmarks have significantly advanced the evaluation of coding agents, but existing benchmarks primarily measure whether generated patches pass functional tests and overlook review-derived acceptance constraints (review constraints) that often influence whether a patch is acceptable in real-world software development.

The authors introduce SWE-Gate, a repository-level benchmark for software engineering agents that explicitly evaluates review constraint compliance alongside functional correctness. SWE-Gate derives review constraints from real pull request review comments and synthesizes repository-level repair instances around these constraints. Each instance provides separate functional and constraint tests, together with non-compliant and gold patches, enabling explicit separation of bug-fixing capability from review constraint compliance.

Key Findings

  • The benchmark contains 303 repository-level repair instances spanning 75 open-source Python repositories across diverse software domains.
  • Experiments under general coding agent frameworks with four LLM backends of varying capability levels reveal a significant gap between functional success and success under the complete repair specification.
  • Of 644 patches that passed functional tests, 221 failed to satisfy the provided review constraints.
  • These results demonstrate that functional-only evaluation overestimates an agent's ability to meet the full requirements of repository-level repair tasks.
  • Resources

  • arXiv paper: https://arxiv.org/abs/2509.04275
  • Replication package (code, data, experimental results): https://github.com/DeepSoftwareAnalytics/SWE-Gate

Tags

#software-engineering#llm#benchmark#coding-agents#arxiv#machine-learning#code-review

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634587