metadata
license: mit
pretty_name: InteractBench
language:
- en
tags:
- code
- benchmark
- competitive-programming
- interactive
- llm-evaluation
InteractBench Dataset
Dataset artifacts for InteractBench: Benchmarking LLMs on Competitive Programming under Unrevealed Information (ICML 2026).
Project page | Paper (OpenReview) | Code
InteractBench is a benchmark of 322 interactive competitive-programming problems curated from Codeforces, AtCoder, IOI, and ICPC. Each problem ships with an executable local interactor, so evaluation runs fully offline without external judge submission.
Files:
problems.jsonl: 298 standard interactive problems.ioi.jsonl: 24 IOI-style problems.test_cases/*.tar.zst: materialized problem test cases split into shards.checksums.sha256: SHA-256 checksums for released files.
Use:
sha256sum -c checksums.sha256
mkdir -p data/problems
for f in test_cases/*.tar.zst; do
tar --use-compress-program=unzstd -xf "$f" -C data/problems
done
python scripts/import_from_jsonl.py --type standard --input problems.jsonl --output-dir data/problems
python scripts/import_from_jsonl.py --type ioi --input ioi.jsonl --output-dir data/problems
License: MIT.
Citation
@inproceedings{li2026interactbench,
title = {InteractBench: Benchmarking {LLM}s on Competitive Programming under Unrevealed Information},
author = {Jiaze Li and Aocheng Shen and Bing Liu and Boyu Zhang and Xiaoxuan Fan and Qiankun Zhang and Xianjun Deng},
booktitle = {Forty-third International Conference on Machine Learning},
year = {2026},
url = {https://openreview.net/forum?id=Y4T4w0Tj0l}
}