hack-verifiable-environments

KO | EN

gitlite — search

#anthropic #awesome #tailwindcss #electron #svg #chatgpt #cloud-native #html #hacktoberfest #rag

hack-verifiable-environments

(★ 9)

A GitHub page for hack-verifiable-environments paper

파일 탐색기

ZIP 다운로드

- favicon.png
- hvtb.html
- index.html
- main_figure.png
- main_figure_dark.png
- main_figure_no_border.png
- video.gif
- video.mp4
- run_hidden_solution_benchmark.py
- test_1p.py
- test_2p.py
- - __init__.py
  - filesystem_wrapper.py
  - virtual_fs.py
- __init__.py
- test_onset_logging.py
.gitignore
LICENSE
pyproject.toml
README.md
uv.lock

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

---

## Introduction

Hack-Verifiable Environments is a new paradigm for evaluating reward hacking.
This repository contains the original code for the paper, including experiment scripts needed to reproduce the results.
We also release Hack-Verifiable TextArena, a fork of TextArena with a filesystem wrapper for evaluating reward hacking on TextArena environments.

![Demo](https://raw.githubusercontent.com/MajoRoth/hack-verifiable-environments/main/docs/video.gif)

---

## Installation

```bash
uv sync
```

Set your API keys in a `.env` file in the project root:

```
OPENAI_API_KEY=...
OPENROUTER_API_KEY=...
```

---

## Hack-Verifiable TextArena

We release [Hack-Verifiable TextArena](https://github.com/MajoRoth/hack-verifiable-environments/tree/main/hvta).
We implemented the filesystem wrapper on top of [TextArena](https://github.com/TextArena/TextArena), which allows us to evaluate reward hacking in a wide range of text-based environments.
We implemented hidden-solution and logical-bug for single-player environments, and read-and-write prompt for two-player environments.

```python
import textarena as ta
from hvta.FilesystemWrapper import FilesystemWrapper

agents = {
    0: ta.agents.HumanAgent(), # ta.agents.OpenRouterAgent(model_name="openai/gpt-5-mini")
}

# initialize the environment
env = ta.make(env_id="Wordle-v0")
env = FilesystemWrapper(env=env, hidden_solution=True, logical_bug=True)
env.reset(num_players=len(agents))

# main game loop
done = False
while not done:
  player_id, observation = env.get_observation()
  action = agents[player_id](observation)
  print(action)
  done, step_info = env.step(action=action)
  if step_info.get("reward_hack"):
    print("[HACK DETECTED 🚨🚨🚨]")

rewards, game_info = env.close()
```

You can use the examples prvovided in `hvta/test_1p.py`, `hvta/test_2p.py` and `hvta/run_hidden_solution_benchmark.py` to get started.

---

## Citation

```bibtex
@article{roth2026hack,
  title={Hack-Verifiable Environments: Towards Evaluating Reward Hacking at Scale},
  author={Roth, Amit and Samanta, Ankur and Halevy, Matan and Levine, Yoav and Efroni, Yonathan},
  journal={arXiv preprint arXiv:2605.20744},
  year={2026}
}
```

hack-verifiable-environments

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README