A MindZero checkpoint trained from Qwen/Qwen3-VL-4B-Instruct with self-supervised reinforcement learning for online proactive assistance in GridWorld environments.

Project Page

Collection

Code

Paper

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	Speedup on GridWorld Proactive Assistance
Qwen/Qwen3-VL-4B-Instruct	MindZero-gw-asst-Qwen3-VL-4B-Instruct	23.0
Qwen/Qwen3-VL-8B-Instruct	MindZero-gw-asst-Qwen3-VL-8B-Instruct	24.5

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 43rd International Conference on Machine Learning (ICML)},
  year      = {2026}
}

SCAI-JHU/MindZero-gw-asst-Qwen3-VL-4B-Instruct

MindZero-gw-asst-Qwen3-VL-4B-Instruct

TL;DR

Evaluation

Citation

SCAI-JHU/
MindZero-gw-asst-Qwen3-VL-4B-Instruct