MindZero: Learning Online Mental Reasoning With Zero Annotations

AI agents often struggle to understand what humans are trying to do in real-time. To be truly helpful, an agent needs "Theory of Mind" (ToM)—the ability to infer a person's hidden goals or beliefs just by watching their behavior. Currently, we either use massive, slow models that think through every possibility at inference time, or small models that are fast but prone to making erratic, unhelpful guesses.

The MindZero paper proposes a middle path. It suggests we can teach small, multimodal large language models (MLLMs) to perform this reasoning via self-supervised reinforcement learning (RL). Instead of requiring humans to label every intention, the model learns to predict mental states that best explain the actions it sees.

The Problem

Current approaches to Theory of Mind generally fall into two camps. Both fail in production-ready assistive scenarios. The first is prompting-based. You ask a standard LLM to "guess what the human wants." These models suffer from systematic errors in long-context understanding and complex reasoning. They lack the ability to maintain a stable probability distribution over multiple competing hypotheses.

The second camp is model-based, specifically Bayesian inverse planning (BIP). These methods are mathematically robust. They explicitly model the likelihood of an action given a mental state. However, these methods are computationally prohibitive for real-time assistance. They require searching large hypothesis spaces at test time. This leads to high latency and massive compute costs. As seen in, an agent must continuously update its belief as new observations arrive. Doing this via heavy deliberation every few milliseconds is a non-starter for a robot or a digital assistant.

How It Works

MindZero aims to "internalize" the heavy lifting of model-based reasoning into the weights of a fast, single-pass MLLM. The core innovation is a Self-Supervised Reinforcement Learning (SSRL) framework.

The architecture moves away from traditional supervised learning. Supervised learning relies on scarce ground-truth annotations of mental states. Instead, the authors use observed actions as the supervisory signal. The process works as follows:

1. Hypothesis Generation: The MLLM receives a sequence of states and actions. It outputs a set of $N$ mental state hypotheses ($m_t$) along with their associated posterior probabilities ($q_t$, the estimated likelihood of each hypothesis being correct).
2. Reward Computation: The model is rewarded based on how well its hypotheses explain the observed behavior. This is done by calculating the "action likelihood." This is the probability of the actual observed action occurring if the proposed mental state were true. This likelihood is estimated by either a model-based planner or a larger LLM acting as an evaluator.
3. Optimization: The authors use Group Relative Policy Optimization (GRPO, a reinforcement learning algorithm) to update the model. The reward function is an Evidence Lower Bound (ELBO) objective. It maximizes the likelihood of the observed action and the mental state prior (the initial belief), while adding an entropy regularization term.

This entropy bonus is critical. It prevents the model from prematurely collapsing onto a single hypothesis. This ensures it maintains a diverse set of possibilities until enough evidence is gathered [Figure 2a].

Numbers

The primary value proposition is the accuracy-to-compute ratio. The authors report that MindZero delivers massive gains in accuracy with negligible additional inference cost compared to test-time scaling methods.

In GridWorld Question Answering, MindZero (using Qwen3-VL-4B) achieves 95.0% accuracy [Figure 4a]. This represents a 2.1–2.5× accuracy gain over the base model. Crucially, this happens in a single forward pass. In contrast, test-time scaling methods like ThoughtTracing or AutoToM require multiple passes. They also consume significantly higher TFLOPs (total floating-point operations) to reach lower accuracy levels.

In proactive assistance tasks, the metrics shift to "speedup." This measures how much the agent accelerates task completion relative to a human working alone. The authors report that MindZero achieves a 23.0% to 24.5% speedup in GridWorld [Table 1a]. In real-world human studies involving household tasks, MindZero (trained on Qwen3-4B) delivered a 19.7% speedup [Table 3]. This performance is comparable to the much larger Gemini-3-Flash. This suggests a small, specialized model can match the utility of a frontier model on cheaper hardware.

What's Missing

While the results are impressive, there are gaps. First, the framework does not currently model recursive reasoning. Recursive reasoning involves modeling "Person A's belief about Person B's goal." The current SSRL objective does not address this complexity.

Second, there is a scaling bottleneck regarding input length. As the observation history grows, the required input token length increases. For long-horizon tasks, this will lead to increased latency. It could also cause OOMs (out-of-memory errors) at inference. This might negate the "fast single-pass" advantage.

Finally, performance on very small backbones, like Llama-3.2-3B, is limited. These models struggle with instruction-following. If the model cannot reliably output the required JSON schema for hypotheses, the reward loop breaks. Success depends heavily on the "reasoning floor" of the underlying MLLM.

Should You Prototype This

Yes, if you are building embodied agents or proactive assistants where latency is a hard constraint. The ability to train for latent mental states using only behavioral logs is a major win. It allows for data flywheel development without expensive human labeling. If you have a simulator, you can start training this immediately.

Code and datasets are reportedly available at the official project links; see the paper for the canonical links. If you are targeting edge deployment, choose a backbone with strong instruction-following. This ensures your hypothesis formatting does not collapse during the RL phase.

Figures from the paper