Optimizing code performance is paramount in software engineering, yet it remains a largely unexplored frontier for Large Language Models (LLMs). While models excel at fixing bugs, their ability to make code faster at a repository-scale is not well understood.

To address this, we introduce SWE-Perf, the first benchmark meticulously designed to evaluate LLMs on performance optimization tasks within genuine, complex repository contexts. Unlike benchmarks that focus on isolated code snippets, SWE-Perf challenges models to understand and modify entire codebases. The benchmark comprises 140 instances, each derived from a real performance-improving pull request on a popular GitHub repository. For each instance, a model is provided with the full source code, the target functions, and the human expert's solution for reference. The core task is to generate a code patch that reduces the test's execution time without introducing bugs.

We recommend using Conda to manage dependencies.

To create the same environment:

conda env create -f environment.yml

Evaluation measures model-generated patch predictions on the SWE-Perf dataset to assess their performance. The evaluation pipeline has two stages:

1. Run Evaluation – Executes each prediction, measures runtime improvements, and stores logs.
2. Check Evaluation – Aggregates logs into a CSV report containing performance metrics.

See the evaluation README for detailed usage.

Generation covers the process of using LLMs to generate code patches that improve performance on SWE-Perf tasks.

The Oracle generation pipeline includes the following steps:

1. Generate Prompts – Collect relevant source files and assemble them into prompts.
2. Run Inference – Feed prompts to an LLM to produce patches.
3. Convert Outputs – Transform model outputs into unified diff format.

See the generation README for details.

The SWE-Perf dataset is built through a five-phase pipeline to identify and verify real-world performance-optimizing pull requests:

1. Collect PRs – Crawl PRs from 12 popular GitHub repositories.
2. Measure Performance – Run all unit tests before/after each PR to obtain runtime data.
3. Identify Performance PRs – Filter PRs with significant improvements and relevant test coverage.
4. Verify Stability – Confirm that performance gains are consistent and reproducible.
5. Extract Optimization Targets – Define function-level targets for Oracle and Realistic evaluation settings.

Full details are in the data collection README.

If you find our work useful, please consider citing our paper:

@article{he2025sweperf,
    title={SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?},
    author={He, Xinyi and Liu, Qian and Du, Mingzhe and Yan, Lin and Fan, Zhijie and Huang, Yiming and Yuan, Zejian and Ma, Zejun},
    journal={arXiv preprint arXiv:2507.12415},
    year={2025}
}

Please reach out to qian.liu@tiktok.com for questions or feedback on SWE-Perf. We welcome collaborations and suggestions for improving the benchmark. This work was conducted during Xinyi and Yiming's internship at TikTok.