According to a paper published on arXiv, researchers have introduced MindZero, a self-supervised reinforcement learning framework designed to train multimodal large language models (MLLMs) for efficient online mental reasoning without requiring ground-truth mental state annotations.
The framework addresses three key challenges in AI assistance: online inference with robust uncertainty updates, efficient real-time reasoning, and the absence of annotated mental state data in real-world domains. According to the paper, MindZero trains models by rewarding them for generating mental state hypotheses that maximize the likelihood of observed actions as estimated by a planner, similar to model-based Theory of Mind reasoning.
The researchers evaluated MindZero on mental reasoning and AI assistance tasks in gridworld and household domains. According to arxiv.org, the study found that “LLMs alone are insufficient; model-based methods improve accuracy but are slow, costly, and limited by backbone MLLM capacity.” In contrast, MindZero “significantly outperforms model-based methods in both accuracy and efficiency.”
After training, the framework internalizes model-based reasoning into fast single-pass inference, the paper states. The research was presented at ICML 2026, with additional information available at the project website hosted by Johns Hopkins University’s School of Computer and Information Engineering.