---
license: apache-2.0
base_model: Qwen/Qwen3-4B-Instruct-2507
datasets:
  - SCAI-JHU/MindZero
  - Chuanyang-Jin/MMToM-QA
library_name: transformers
pipeline_tag: text-generation
language:
  - en
tags:
  - mindzero
  - theory-of-mind
  - reinforcement-learning
  - reasoning
  - household
---

# MindZero-hh-tom-Qwen3-4B-Instruct-2507

A [**MindZero**](https://scai.cs.jhu.edu/MindZero/) checkpoint trained from [`Qwen/Qwen3-4B-Instruct-2507`](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) with self-supervised reinforcement learning for online Theory-of-Mind reasoning in household environments.

[![Project Page](https://img.shields.io/badge/Homepage-Visit-blue?labelColor=gray&logo=homeassistantcommunitystore&logoColor=367BAF&style=flat-square)](https://scai.cs.jhu.edu/MindZero/)
[![Collection](https://img.shields.io/badge/HuggingFace-Collection-yellow?labelColor=gray&logo=huggingface&style=flat-square)](https://huggingface.co/collections/SCAI-JHU/mindzero/)
[![Code](https://img.shields.io/badge/GitHub-Code-black?labelColor=gray&logo=github&style=flat-square)](https://github.com/SCAI-JHU/MindZero)
[![Paper](https://img.shields.io/badge/Paper-PDF-red?labelColor=gray&logo=arxiv&logoColor=brown&style=flat-square)](https://arxiv.org)

## TL;DR

**MindZero** trains (M)LLMs to perform efficient and robust online mental reasoning **without any mental-state annotations**. During training, the model is rewarded for generating mental-state hypotheses that maximize the likelihood of observed actions, as estimated by a planner — analogous to model-based ToM reasoning. After training, MindZero internalizes this reasoning into fast single-pass inference.

## Evaluation

| Base model                           | Checkpoint                                                                                                       | MMToM-QA |
| ------------------------------------ | ---------------------------------------------------------------------------------------------------------------- | -------- |
| [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)           | [MindZero-hh-tom-Qwen3-4B-Instruct-2507](https://huggingface.co/SCAI-JHU/MindZero-hh-tom-Qwen3-4B-Instruct-2507) | 72.7     |
| [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) | [MindZero-hh-tom-Llama-3.2-3B-Instruct](https://huggingface.co/SCAI-JHU/MindZero-hh-tom-Llama-3.2-3B-Instruct)   | 77.8     |
| [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) | [MindZero-hh-tom-Llama-3.1-8B-Instruct](https://huggingface.co/SCAI-JHU/MindZero-hh-tom-Llama-3.1-8B-Instruct)   | 76.2     |

## Citation

```bibtex
@inproceedings{zhang2026mindzero,
  title     = {MindZero: Learning Online Mental Reasoning With Zero Annotations},
  author    = {Shunchi Zhang and Jin Lu and Chuanyang Jin and Yichao Zhou and Zhining Zhang and Tianmin Shu},
  booktitle = {Proceedings of the 43st International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```