README.md

---
title: README
emoji: 🐠
colorFrom: yellow
colorTo: gray
sdk: gradio
pinned: false
---

# GTAlign: Game-Theoretic Alignment of LLM Assistants for Mutual Welfare

[![ArXiv](https://img.shields.io/badge/arXiv-2510.08872-b31b1b?style=flat&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2510.08872)
[![Hugging Face](https://img.shields.io/badge/HuggingFace-GTAlign-orange?logo=huggingface&logoColor=white)](https://huggingface.co/GTAlign)

**GTAlign** applies game-theoretic principles to fine-tune reasoning LLMs, encouraging them to make decisions that are not only accurate but also rational, cooperative, and transparent in dialogue settings.


## Models

We have released five model checkpoints, and we are preparing more thoroughly trained models.

| Model Name | Size | Dataset | Hugging Face Link |
|------------|------|--------|--------------------|
| `GTAlign/Qwen2.5-3B-Math-140step` | 3B | Math | [Model](https://huggingface.co/GTAlign/Qwen2.5-3B-Math-140step) |
| `GTAlign/Qwen2.5-3B-Medium-110step` | 3B | Medium | [Model](https://huggingface.co/GTAlign/Qwen2.5-3B-Medium-110step) |
| `GTAlign/Qwen2.5-3B-AbgQA-140step` | 3B | Ambig-QA | [Model](https://huggingface.co/GTAlign/Qwen2.5-3B-AbgQA-140step) |
| `GTAlign/Qwen2.5-3B-WildGuard-140step` | 3B | WildGuard | [Model](https://huggingface.co/GTAlign/Qwen2.5-3B-WildGuard-140step) |
| `GTAlign/Qwen2.5-3B-Full-160step` | 3B | Full | [Model](https://huggingface.co/GTAlign/Qwen2.5-3B-Full-160step) |