Papers · SWE-bench
AI PAPER

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

Created a benchmark of 2,294 real GitHub issues and their verified solutions, measuring whether AI systems can autonomously resolve actual software engineering tasks.

Authors
Carlos E. Jimenez, John Yang, Alexander Wettig +4
Institution
Princeton University
Published
October NaN, 2023
Citations
600
Category
Evaluation
Impact
notable
PAPER EXPLAINED

Plain-English Summary

Most coding benchmarks test isolated function generation — “write a function that sorts a list.” Real software engineering is nothing like this. SWE-bench measures something far harder: given a real GitHub issue report from projects like Django, Scikit-learn, and Matplotlib, can an AI system navigate the codebase, understand the bug, and produce a correct patch? Each task requires reading multiple files, understanding architecture, and making changes that pass existing test suites.

Initial results were sobering. GPT-4 with basic prompting resolved only 1.7% of issues. Even with sophisticated agent scaffolding, early systems topped out around 4%. This established that autonomous software engineering was far from solved.

Key Innovation

SWE-bench’s design ensures evaluation is rigorous and ungameable. Solutions are verified by running the project’s actual test suite, including tests that were written specifically for the issue being resolved. This means the model must produce code that is both correct and compatible with the existing codebase. There is no shortcut — partial solutions or cosmetic fixes fail the tests.

The benchmark spans diverse difficulty levels, from simple one-line fixes to complex multi-file refactors requiring deep understanding of library internals.

Impact on the Field

SWE-bench became the primary metric for evaluating AI coding agents. Devin, SWE-Agent, AutoCodeRover, and every subsequent coding agent reports its SWE-bench score. The benchmark drove rapid progress: scores went from 1.7% to over 50% within 18 months as agent architectures improved.

The benchmark also shifted the narrative around AI coding from “can it write functions?” to “can it do real engineering?” This higher bar influenced product development at companies like Cursor, GitHub Copilot, and Cognition.

Models That Built on This

Cognition’s Devin (the first “AI software engineer”) was evaluated primarily on SWE-bench. SWE-Agent introduced effective prompting strategies. Claude and GPT-4 have been benchmarked extensively, with Claude 3.5 Sonnet achieving over 49% on SWE-bench Verified. The benchmark continues to drive research in AI-assisted software development.