CoDA-Bench

# CoDA-Bench: Can Code Agents Handle Data-Intensive Tasks? [](https://arxiv.org/pdf/2606.15300) [](https://github.com/ruc-datalab/CoDA-Bench) [](https://coda-bench.github.io/) > **Authors**: **[Yuxin Zhang](https://github.com/yuxinzhang), [Ju Fan](http://iir.ruc.edu.cn/~fanj/), [Meihao Fan](https://fmh1art.github.io/), [Shaolei Zhang*](https://zhangshaolei1998.github.io/), [Xiaoyong Du](http://info.ruc.edu.cn/jsky/szdw/ajxjgcx/jsjkxyjsx1/js2/7374b0a3f58045fc9543703ccea2eb9c.htm)** **CoDA-Bench** (Code and Data-intensive Benchmark) is the first benchmark to jointly evaluate **code intelligence** and **data intelligence** of AI agents in realistic data-intensive environments. Unlike existing benchmarks that provide oracle data directly, CoDA-Bench requires agents to: - 🔍 **Discover relevant data** among hundreds of semantically similar files - 🗂️ **Navigate complex file hierarchies** in a Linux sandbox environment - 🔗 **Integrate information** from multiple heterogeneous data sources - 💻 **Generate correct code** for data-driven analytical tasks ## 📊 Dataset Overview - **Full Benchmark**: 1,009 tasks across 31 communities (`coda_bench.json`) - **Hard Subset**: 119 challenging tasks across 15 communities (`coda_bench_hard.json`) - **Source Data**: 199 Kaggle datasets from 267 notebooks - **Scale**: Average 980 files per environment (~43 GB total compressed) ## 🏆 Benchmark Results Current state-of-the-art (as of paper publication): | System | Model | EA (Full) | EA (Hard) | |--------|-------|-----------|-----------| | Mini-SWE-Agent | GPT-5.5 | **61.1%** | **49.6%** | | Codex CLI | GPT-5.5 | 60.3% | 47.9% | | OpenHands | GPT-5.5 | 59.7% | 44.5% | | Claude Code | Sonnet-4.6 | 53.8% | 42.9% | ## 📚 Citation ```bibtex @inproceedings{zhang2026codabench, title={CODA-BENCH: Can Code Agents Handle Data-Intensive Tasks?}, author={Zhang, Yuxin and Fan, Ju and Fan, Meihao and Zhang, Shaolei and Du, Xiaoyong}, booktitle={Proceedings of the 43rd International Conference on Machine Learning}, year={2026}, organization={PMLR} } ``` More information refer to [CoDA-Bench's Repo](https://github.com/ruc-datalab/CoDA-Bench)