---
license: mit
language:
- en
language_creators:
- expert-generated
annotations_creators:
- expert-generated
multilinguality:
- monolingual
source_datasets:
- original
pretty_name: EVM-QuestBench
task_categories:
- text-generation
tags:
- benchmark
- code-generation
- transaction-code-generation
- execution-grounded-evaluation
- blockchain
- evm
- ethereum
- bsc
- smart-contracts
- defi
- llm
- llm-agents
- agent-evaluation
- acl-2026
size_categories:
- n<1K
arxiv: "2601.06565"
doi: "10.18653/v1/2026.acl-long.1642"
configs:
- config_name: default
  data_files:
  - split: atomic
    path: data/atomic.jsonl
  - split: composite
    path: data/composite.jsonl
---

# EVM-QuestBench

**EVM-QuestBench is an execution-grounded benchmark for evaluating whether large language models and AI agents can translate natural-language blockchain intent into executable transaction code that produces the intended EVM state transition.**

Released with the ACL 2026 Long Paper by Pei Yang, Wanyi Chen, Ke Wang, Lynn Ai, Eric Yang, and Tianyu Shi, the benchmark contains **107 expert-authored tasks**: **62 atomic tasks** and **45 composite workflows**. Unlike code-similarity benchmarks, EVM-QuestBench executes generated actions on a snapshot-isolated BSC fork and validates receipts, calldata, balances, allowances, ownership, and other post-state evidence.

## Resources

- **Paper (ACL Anthology):** https://aclanthology.org/2026.acl-long.1642/
- **Paper (arXiv):** https://arxiv.org/abs/2601.06565
- **GitHub repository:** https://github.com/OpenEdgeHQ/EVM-quest-bench
- **Project website:** https://openedgehq.github.io/EVM-quest-bench/
- **Dataset Viewer:** https://huggingface.co/datasets/berryccc1/EVM-QuestBench/viewer/default
- **DOI:** https://doi.org/10.18653/v1/2026.acl-long.1642

## What the benchmark evaluates

EVM-QuestBench covers native-token transfers, ERC-20 and NFT operations, contract calls, PancakeSwap swaps and liquidity, staking, flashloans, blockchain queries, and multi-step DeFi workflows.

- **Atomic:** one transaction or query, evaluated with task-specific execution validators.
- **Composite:** multi-step planning and execution with feedback, parameter propagation, and step-efficiency decay.
- **Ground truth:** successful execution and the requested EVM state change, not string similarity to reference code.

## Load the dataset

```python
from datasets import load_dataset

dataset = load_dataset("berryccc1/EVM-QuestBench")
print(dataset)
print(dataset["atomic"][0])
```

The default configuration exposes two splits:

| Split | Rows | Description |
|---|---:|---|
| `atomic` | 62 | Single transaction or query tasks |
| `composite` | 45 | Multi-step transaction workflows |

## Dataset structure

Each row includes:

- `id`, `task_type`, `title`, `category`, `subcategory`, and `difficulty`
- `description` and `natural_language_templates`
- `chain`, `network`, `tags`, and `parameter_names`
- JSON-serialized `parameters`, `validation`, `composite_structure`, `interaction_config`, `scoring_strategy`, and `metadata`
- `source_path` and the complete JSON-serialized `raw_definition`

Nested task definitions are serialized as JSON strings to provide one stable schema across all 107 rows and maintain Dataset Viewer compatibility.

## Dataset sources and reproducibility

The dataset is derived from the task definitions released in the official [EVM-QuestBench repository](https://github.com/OpenEdgeHQ/EVM-quest-bench). The `benchmark/` directory preserves the original task definitions, validators, system configurations, and parameter generator used to construct the viewer-friendly splits.

Running the full execution benchmark requires the Anvil/BSC environment, TypeScript runner, model API configuration, and dependencies documented in the GitHub repository. The Hugging Face dataset is the canonical browsable distribution of task definitions; the GitHub repository remains the canonical benchmark implementation.

## Dataset creation

The tasks were authored by the benchmark creators to cover representative EVM transaction and query capabilities. Natural-language templates and parameter schemas dynamically instantiate task variants. Atomic tasks use task-specific validators, while composite workflows define ordered operations, shared parameters, optimal steps, and scoring strategies. See the paper and repository for the complete methodology.

## Intended use

This dataset is intended for:

- evaluating LLM and agent transaction-code generation;
- studying single-step versus multi-step execution reliability;
- developing execution-grounded validators and EVM agent evaluation methods;
- reproducing or extending the EVM-QuestBench benchmark.

It is not a substitute for smart-contract auditing, financial advice, wallet security review, or testing transactions with real funds.

## Limitations

- The execution environment targets EVM-compatible chains and is instantiated primarily on a BSC mainnet fork.
- Public task definitions may contribute to benchmark contamination; report training exposure when known.
- The 107 tasks do not cover every contract, protocol, chain, or adversarial on-chain condition.
- Successful benchmark execution does not establish that generated code is safe for production or real assets.
- Full scores require the external execution environment; this dataset alone provides definitions and reproducibility artifacts.

## Dataset card authors and contact

Dataset card maintained by the EVM-QuestBench authors. For questions or corrections, open an issue at https://github.com/OpenEdgeHQ/EVM-quest-bench/issues.

## Citation

```bibtex
@inproceedings{yang-etal-2026-evm-questbench,
  title     = {EVM-QuestBench: An Execution-Grounded Benchmark for Natural-Language Transaction Code Generation},
  author    = {Yang, Pei and Chen, Wanyi and Wang, Ke and Ai, Lynn and Yang, Eric and Shi, Tianyu},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics},
  year      = {2026},
  pages     = {35513--35529},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2026.acl-long.1642/},
  doi       = {10.18653/v1/2026.acl-long.1642}
}
```

## License

MIT. See the source repository for the benchmark implementation and provenance.