OEISPerf

Beyond Correctness: Benchmarking LLMs for Algorithmic Program Optimization
Findings of EMNLP 2026

OEISPerf is a code-to-code benchmark for algorithmic program optimization, built from alternative implementations of integer sequences in the OEIS. Each instance pairs a slow (inefficient) Python program with the fastest validated reference for the same sequence. The model sees only the slow program and must return a faster calculator(n: int) -> int.

📦 Dataset Structure

	train	validation	test
Instances	2,762	240	265

Column	Type	Description
id	string	Instance id, e.g. test_0
inefficientCode	string	Slow program, the only input shown to the model
offset_start	int64	Index of the first term: calculator(offset_start + i) must equal the i-th ground-truth term
termSize	int64	Number of ground-truth terms
encrypted_referenceCode	string	Efficient reference program (encrypted)
encrypted_expected_output	string	Ground-truth terms (encrypted)

🔐 Encrypted fields. The reference program and the ground-truth terms are stored encrypted. The loader in the GitHub repository decrypts them into referenceCode and expected_output. Please do not republish the decrypted data.

🚀 Usage

Clone the GitHub repository, install requirements.txt, then:

from utils import OEISPerfLoader

df = OEISPerfLoader.load("test")   # "train" | "validation" | "test"

The repository also contains the evaluation code used in the paper.

📄 License

OEISPerf is built from the OEIS, which is distributed under CC BY-SA 4.0, and is released under the same license: CC BY-SA 4.0.

📚 Citation

If you find this dataset helpful, please cite our paper:

@inproceedings{lin-chen-2026-beyond,
    title = "Beyond Correctness: Benchmarking {LLM}s for Algorithmic Program Optimization",
    author = "Lin, Yehua  and
      Chen, Yin",
    editor = "Lee, Hung-Yi  and
      Dev, Sunipa  and
      Li, Jessy  and
      Elliott, Desmond",
    booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-emnlp.152/",
    pages = "3211--3226",
    ISBN = "979-8-89176-462-0",
    abstract = "Large language models (LLMs) can generate functionally correct code, but their ability to optimize programs is underexplored. We introduce OEISPerf, a code-to-code benchmark for algorithmic program optimization built from alternative implementations of integer sequences in the On-Line Encyclopedia of Integer Sequences. For each sequence we mine multiple mathematical formulations, synthesize and validate Python implementations, and pair a substantially slower program with the fastest validated reference found by our construction pipeline, producing instances that often require substantial structural transformations and can yield substantial speedups. Evaluating 15 open-source and closed-source LLMs shows that even the strongest evaluated setting, Gemini-3-Flash with its thinking mode enabled, matches the reference efficiency on only 67.9{\%} of test instances. Across LLMs, we observe both incorrect ``optimizations'' that break correctness and superficial refactorings that preserve suboptimal complexity. Finally, we fine-tune Qwen2.5-Coder-3B-Instruct via reinforcement learning, markedly improving reference attainment on OEISPerf while slightly improving performance on standard code-generation benchmarks."
}