README.md

---
license: cc
tags:
- code
- code-tracing
- structural-execution
- code-understanding
pretty_name: CoCoNUT
size_categories:
- 1K<n<10K
---

## Dataset Info:
CoCoNUT investigates the capabilities of selected Large Language Models on understanding structural code execution. The dataset includes tasks where models reproduce the code lines executed for specific input arguments, testing advanced code concepts such as Object-Oriented Programming (OOP), Concurrency, and Recursion. The dataset contains short programs, their traces, and the corresponding call arguments.
For advanced topics, no call arguments are needed since they are directly contained in the main function.
The Code can be found at the github repository: https://github.com/ClaasBeger/StructuralCodeUnderstanding and the paper pre-print is available at \cite{arxiv.org/abs/2501.16456}

Model Performance:
![Model Performance](Bucket_linePlot_direct_overlap.png "Model Performance on CoCoNUT")

## Citation

```bibtex
@misc{beger2025coconutstructuralcodeunderstanding,
      title={CoCoNUT: Structural Code Understanding does not fall out of a tree}, 
      author={Claas Beger and Saikat Dutta},
      year={2025},
      eprint={2501.16456},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2501.16456}, 
}
```