README.md

---
pretty_name: PACE-Bench
language:
  - en
license: mit
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - benchmark
  - agents
  - agent-evaluation
  - self-evolving-agents
  - physics
  - physical-reasoning
  - code
  - code-generation
  - executable-design
  - simulation
  - box2d
  - dynamic-environments
citation: |
  @misc{zhan2026pacebenchbenchmarkingphysicsadaptation,
        title={PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments},
        author={Yuhao Zhan and Bingxiang He and Zecong Tang and Chaojun Xiao},
        year={2026},
        eprint={2608.14441},
        archivePrefix={arXiv},
        primaryClass={cs.AI},
        url={https://arxiv.org/abs/2608.14441},
  }
source_datasets: []
---

> **Task mirror only.** This dataset repository contains the 36 executable task definitions. The CLI, shared runtime, evaluation engine, reporting code, and coding-agent sandbox live in the [PACE-Bench GitHub repository](https://github.com/thunlp/PACE-Bench).

<details>
<summary>Mirror provenance</summary>

- Source: [`src/pace_bench/tasks/categories`](https://github.com/thunlp/PACE-Bench/tree/main/src/pace_bench/tasks/categories)
- Included: 36 base tasks and `primitives_api.json`
- Excluded: shared evaluation runtime, generated results, caches, and local artifacts

</details>

# PACE-Bench

Self-evolving agents improve future behavior from interaction experience, yet existing evaluations typically keep execution conditions fixed. **PACE-Bench** tests whether an agent can adapt a previously successful code-driven design after an environment shift causes it to fail.

[![GitHub](https://img.shields.io/badge/GitHub-Code-181717?logo=github)](https://github.com/thunlp/PACE-Bench)
[![License](https://img.shields.io/badge/License-MIT-blue.svg)](https://github.com/thunlp/PACE-Bench/blob/main/LICENSE)
[![Python](https://img.shields.io/badge/Python-3.10-blue.svg)](https://www.python.org/)
[![arXiv](https://img.shields.io/badge/arXiv-2608.14441-b31b1b.svg?logo=arxiv&logoColor=white)](https://arxiv.org/pdf/2608.14441)
[![Hugging Face Paper](https://img.shields.io/badge/%F0%9F%A4%97%20Paper-2608.14441-FFD21E)](https://huggingface.co/papers/2608.14441)

## What is PACE-Bench?

PACE-Bench contains **144 source-to-target adaptation pairs across six physics domains**. Each pair keeps the goal and interface fixed:

1. A **code-driven design** succeeds in the source environment.
2. The same design fails in a **mutated target environment**.
3. The agent uses **diagnostic sandbox feedback** to revise the design.
4. The adapted design must succeed under the target physics.

| Benchmark scale | Count |
| --- | ---: |
| Physics domains | 6 |
| Base tasks | 36 |
| Environments per task | 5 |
| Evaluation environments | 180 |
| Source-to-target pairs | 144 |

## Frontier LLM results

The following models were evaluated with the Vanilla method. For a consistent
comparison, only run 1 is used, and the observed single-run value is reported
unchanged as Pass@2 / Score@2. Every metric uses the fixed denominator of 144
eval units; missing eval units are zero-scored.

| Model | Pass@2 | Score@2 |
| --- | ---: | ---: |
| Qwen3.8-Max | 86.11% | 86.11 |
| GLM-5.2 | 85.42% | 85.42 |
| DeepSeek-V4-Pro | 82.64% | 82.85 |
| GPT-5.5 | 82.64% | 83.04 |
| DeepSeek-V4-Flash | 78.47% | 78.47 |
| Kimi-K3 | 77.78% | 77.35 |
| Claude Opus 5 | 77.08% | 76.29 |

## What is included here?

| Domain | Prefix | Tasks |
| --- | --- | ---: |
| Statics / Equilibrium | `S` | 6 |
| Kinematics / Linkages | `K` | 6 |
| Dynamics / Energy | `D` | 6 |
| Granular / Fluid Interaction | `F` | 6 |
| Cybernetics / Control | `C` | 6 |
| Exotic Physics | `E` | 6 |

```text
tasks/
├── Category1_Statics_Equilibrium/S_01 ... S_06/
├── Category2_Kinematics_Linkages/K_01 ... K_06/
├── Category3_Dynamics_Energy/D_01 ... D_06/
├── Category4_Granular_FluidInteraction/F_01 ... F_06/
├── Category5_Cybernetics_Control/C_01 ... C_06/
├── Category6_ExoticPhysics/E_01 ... E_06/
└── primitives_api.json
```

Each task package contains:

| File | Role |
| --- | --- |
| `agent.py` | Source and four target reference solutions |
| `environment.py` | Box2D world, primitives, and mutable physics |
| `evaluator.py` | Success criteria, score, constraints, and raw metrics |
| `feedback.py` | Diagnostic feedback derived from measured metrics |
| `prompt.py` | Task description, exposed values, and primitive API |
| `renderer.py` | Evaluation-neutral visualization |
| `stages.py` | Four target mutations and prompt updates |

These are **executable benchmark definitions**, not a conventional row-based dataset. The Dataset Viewer is therefore not the primary interface.

## Download the task mirror

```python
from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="YuhaoZhan/PACE-Bench",
    repo_type="dataset",
    allow_patterns=["tasks/**"],
)
print(path)
```

Or clone it directly:

```bash
git clone https://huggingface.co/datasets/YuhaoZhan/PACE-Bench
```

Use this mirror when you need to inspect, archive, or distribute the task definitions without the full evaluation stack.

## Run the benchmark

The Hugging Face mirror is **not standalone**. For evaluation, install the complete GitHub repository with [uv](https://docs.astral.sh/uv/):

```bash
git clone https://github.com/thunlp/PACE-Bench.git
cd PACE-Bench

uv venv .venv --python 3.10
source .venv/bin/activate  # Windows: .venv\Scripts\activate
uv pip install -r requirements.txt

pace-bench list --task S_01
pace-bench validate --task S_01
```

The GitHub checkout already contains the same task definitions. You do **not** need to download this mirror separately to run PACE-Bench.

## Intended use

- Evaluate adaptation after controlled physical environment changes
- Study feedback-driven code evolution and self-evolving agents
- Compare context-, memory-, search-, and parameter-based methods
- Analyze physical reasoning, redesign, exploration, and convergence failures
- Inspect or extend executable task definitions

## Scope and limitations

- **Physics:** 2D rigid-body systems in Box2D
- **Language:** English prompts and diagnostic feedback
- **Not covered:** 3D/deformable physics, full fluids, perception, navigation, and multi-agent coordination
- **Execution safety:** run generated code on a dedicated evaluator host without unrelated credentials
- **Version note:** these tasks include an additional difficulty-escalation pass beyond the paper version, so new scores may differ slightly while its conclusions remain unchanged

## Issues and contributions

The Hugging Face repository is a distribution mirror. Please open task issues, fixes, and pull requests in the [GitHub repository](https://github.com/thunlp/PACE-Bench).

## License

PACE-Bench is released under the [MIT License](https://github.com/thunlp/PACE-Bench/blob/main/LICENSE).

## Citation

```bibtex
@misc{zhan2026pacebenchbenchmarkingphysicsadaptation,
      title={PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments},
      author={Yuhao Zhan and Bingxiang He and Zecong Tang and Chaojun Xiao},
      year={2026},
      eprint={2608.14441},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.14441},
}
```