MindZero: Self-Supervised Online ToM

• MindZero is a self-supervised reinforcement learning framework for online Theory of Mind that infers latent mental states from observed actions without explicit mental-state annotations.
• It leverages Bayesian inverse planning with MLLMs to score hypotheses via action-likelihood evaluators, achieving high accuracy in both question answering and proactive assistance tasks.
• The framework internalizes probabilistic mental reasoning during training to enable fast single-pass inference, ensuring efficient real-time mental-state tracking with structured uncertainty.

Searching arXiv for MindZero and closely related Theory-of-Mind / online mental reasoning work to ground the article in current literature. {"query":"MindZero Learning Online Mental Reasoning With Zero Annotations Theory of Mind online inference planner Bayesian inverse planning arXiv","max_results":10,"sort_by":"relevance"} {"query":"(Zhang et al., 29 May 2026) MindZero arXiv","max_results":5,"sort_by":"relevance"} MindZero is a self-supervised reinforcement learning framework for online Theory of Mind (ToM) in multimodal LLMs (MLLMs). It is designed to infer and track latent human mental states—such as goals, beliefs, and task structure—from observed behavior without using explicit mental-state annotations. Rather than supervising a model with ground-truth goals or beliefs, MindZero rewards hypotheses that maximize the likelihood of observed actions under an external action-likelihood evaluator, such as a planner or an LLM-based reward model. The central claim is that model-based mental reasoning can be internalized during training and then executed as fast single-pass inference at test time, making the framework relevant both to mental-state question answering and to real-time proactive assistance (Zhang et al., 29 May 2026).

1. Conceptual definition and formal setting

MindZero addresses online mental reasoning: after each new observation $(s_t, a_t)$, the system updates its belief over a human’s latent mental state in real time. In the framework’s terminology, a mental state hypothesis $m_t^{(i)}$ is an explicit candidate explanation of behavior at time $t$, and the model outputs a set of such hypotheses, $\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$, together with a probability distribution $\mathcal{Q}_t = \{ q_t^{(1)}, \dots, q_t^{(N)}\}$, where the probabilities sum to $1$. In practice these hypotheses are serialized as JSON and can encode goals, subgoals, targeted objects with counts, and target locations (Zhang et al., 29 May 2026).

The formal reference point is Bayesian inverse planning. MindZero adopts the posterior

$$P(m_t \mid s_{1:t}, a_{1:t}) \propto P(a_{1:t} \mid m_t, s_{1:t}) P(m_t),$$

where $P(a_{1:t} \mid m_t, s_{1:t})$ is the likelihood of the observed action history under a hypothesized mental state, and $P(m_t)$ is a prior over mental states. The framework differs from classic narrow inverse-planning pipelines by training an MLLM to approximate this posterior directly, instead of recomputing it through explicit test-time search at every step (Zhang et al., 29 May 2026).

The framework is evaluated in two domains. In GridWorld or Construction, the human’s goal is to assemble two specific colored objects by picking one and moving it next to the other; the mental state can therefore be written as $m_t = (o_1, o_2)$ with a canonical ordering constraint. In the household domain based on VirtualHome and Watch-And-Help, the mental state includes a task name, a list of object types with counts, and a target location. The task name is drawn from

$m_t^{(i)}$0

This representation makes MindZero a structured latent-variable model for behavior explanation rather than a free-form conversational ToM system (Zhang et al., 29 May 2026).

A second formal component is proactive assistance. The helper acts by marginalizing over mental-state hypotheses:

$m_t^{(i)}$1

Here the assistant’s action policy is explicitly conditioned on uncertainty over the human’s current goal. This makes calibration and multi-hypothesis tracking central to the framework’s downstream utility, not merely an interpretability feature (Zhang et al., 29 May 2026).

2. Architecture, representations, and domain-specific components

MindZero centers on a backbone MLLM or LLM, denoted $m_t^{(i)}$2, which consumes the state-action history and emits a finite distribution over mental-state hypotheses. In GridWorld the backbone is Qwen3-VL-4B or Qwen3-VL-8B. In the household setting the backbones are Llama-3.1-8B, Llama-3.2-3B, and Qwen3-4B. The inputs can be visual or textual; in the household domain, visual information is converted into a fused textual representation taken from prior work. The output is a minified JSON object following a schema appropriate to the domain, such as GoalParticles containing object pairs and probabilities in GridWorld or task specifications with object counts and targets in Household (Zhang et al., 29 May 2026).

MindZero separates train-time reasoning infrastructure from test-time inference. During training, each candidate mental state is scored by an action-likelihood evaluator. In GridWorld this evaluator is a model-based planner that simulates a Boltzmann-rational policy and approximates

$m_t^{(i)}$3

In Household, the evaluator is an LLM-based reward model, specifically Qwen3-235B-A22B, prompted to rate how consistent a hypothesis is with the observed trajectory. The same reward-model family also estimates priors over household goals, which is important because some structured goals are commonsensically implausible even if they can be phrased syntactically (Zhang et al., 29 May 2026).

The prior differs by domain. In GridWorld it is uniform over feasible object pairs, so $m_t^{(i)}$4 is constant. In Household it is estimated by the reward LLM as a plausibility score over task-object-target combinations. This prior is used to penalize excessively broad or implausible goal sets. A plausible implication is that MindZero treats mental-state inference not only as inverse action modeling but also as constrained commonsense inference over structured goal spaces (Zhang et al., 29 May 2026).

The planner infrastructure also extends to assistance. GridWorld uses a goal-conditioned shortest-path human planner with Boltzmann policy at $m_t^{(i)}$5, plus side assumptions such as alternating move-then-pause when carrying objects and $m_t^{(i)}$6 random actions to simulate human noise. The GridWorld helper planner maintains a goal distribution $m_t^{(i)}$7 and computes

$m_t^{(i)}$8

then acts with a Boltzmann policy over $m_t^{(i)}$9 while applying heuristics such as yielding to the human, avoiding blocking, and avoiding deadlocks. In Household, the assistance layer reuses VirtualHome’s goal-conditioned MCTS planner together with the uncertainty-aware helping planner from NOPA (Zhang et al., 29 May 2026).

3. Self-supervised learning objective

The defining methodological claim of MindZero is that it learns ToM without explicit mental-state labels. The model never observes annotations of the form “the user’s true goal is $t$0.” Instead it is trained so that the hypotheses it emits make the observed action history probable under the external evaluator. This is expressed as a variational objective:

$t$1

where $t$2 is the entropy of the approximate posterior. The entropy term prevents premature collapse onto a single explanation and encourages maintenance of uncertainty over multiple plausible goals (Zhang et al., 29 May 2026).

In implementation, the model outputs $t$3 hypotheses and associated probabilities, and training uses the sample-based reward

$t$4

This reward contains three components: a probability-weighted action likelihood, a probability-weighted prior, and a negative entropy term. The first aligns hypotheses with observed behavior; the second suppresses reward hacking via implausible or overly broad goals; the third preserves diversity in the posterior approximation (Zhang et al., 29 May 2026).

Optimization is performed with GRPO, implemented in VeRL. A training step samples an episode and timestep, builds a prompt from the domain description and the observed history, generates multiple rollouts, parses each rollout into hypothesis sets and probabilities, computes the reward for each rollout, and applies a policy-gradient update with KL regularization against the base model. The reported configuration is: learning rate $t$5, weight decay $t$6, KL coefficient $t$7, max grad norm $t$8, 20 epochs, global batch size 8, rollout batch size 32, and 32 rollouts per prompt. Training uses 4×H100 for MindZero, with an additional 4×H100 for the household reward model (Zhang et al., 29 May 2026).

This procedure is described as self-supervised reinforcement learning because supervision arises entirely from trajectories and evaluator-derived consistency scores. A plausible interpretation is that MindZero distills a model-based inverse planner into the parameters of a general MLLM, replacing explicit mental-state search at deployment with amortized inference (Zhang et al., 29 May 2026).

4. Online inference, uncertainty tracking, and efficiency

At test time, MindZero takes the full history $t$9 together with the same structured instruction used during training and performs a single forward pass. It outputs a small set of hypotheses and probabilities, with no planner loop and no reward-model calls. These outputs can be used directly for question answering or passed to an assistance planner that marginalizes over the distribution, as in the formal assistance policy above (Zhang et al., 29 May 2026).

This sharply distinguishes MindZero from test-time model-based ToM methods such as ThoughtTracing and AutoToM. Those methods enumerate or sample hypotheses and evaluate them through LLMs or planners at inference time, which makes latency grow with hypothesis count and planning horizon. MindZero instead internalizes that reasoning during training. The paper explicitly contrasts three families of methods: pure prompting baselines, which are fast but brittle; explicit model-based ToM, which is principled but expensive; and supervised ToM approaches, which require mental-state labels. MindZero is positioned as combining the fast inference profile of the first with the structured reasoning of the second while avoiding the annotation demands of the third (Zhang et al., 29 May 2026).

The efficiency results are reported in approximate FLOPs per test episode. In GridWorld QA with Qwen3-VL-4B, the base model and MindZero each use 3.6 TFLOPs, while ThoughtTracing requires 31.0 TFLOPs and AutoToM 344.4 TFLOPs. In Household QA, Llama-3.2-3B base uses 4.0 TFLOPs and MindZero 4.4 TFLOPs, while ThoughtTracing uses 232.9 TFLOPs and AutoToM 23.4 TFLOPs; with Qwen3-4B, the base uses 10.9 TFLOPs and MindZero 13.1 TFLOPs, versus 291.2 TFLOPs for ThoughtTracing and 177.5 TFLOPs for AutoToM. For proactive assistance, GridWorld MindZero with Qwen3-VL-4B uses 161.4 TFLOPs, compared with 151.7 TFLOPs for the base model and 808.6 TFLOPs for Qwen3-VL-235B-A22B; in Household, MindZero with Qwen3-4B uses 201.2 TFLOPs, compared with 213.1 TFLOPs for the base model and 1101.6 TFLOPs for Qwen3-235B-A22B (Zhang et al., 29 May 2026).

These numbers support the paper’s central systems claim: mental reasoning can be learned as a self-supervised skill and then deployed with essentially the same latency as the base model. They also clarify that MindZero is not a search-heavy inference wrapper around an MLLM, but a train-time distillation framework for online probabilistic mental-state tracking (Zhang et al., 29 May 2026).

5. Evaluation results and empirical behavior

MindZero is evaluated on GridWorld question answering, GridWorld proactive assistance, Household question answering based on MMToM-QA, and Household proactive assistance in O-WAH. In GridWorld QA, base Qwen3-VL-4B reaches 37.7% accuracy and base Qwen3-VL-8B reaches 43.3%; the best test-time-scaling baseline, ThoughtTracing with Qwen3-8B, reaches 56.7%, and Gemini-3-Pro reaches 83.7%. MindZero reaches 95.0% with Qwen3-VL-4B and 92.3% with Qwen3-VL-8B. In Household QA, base Llama-3.1-8B reaches 41.3% and base Qwen3-4B 42.8%; AutoToM with Qwen3-235B reaches 67.5%, GPT-5.2-Think 73.5%, and Gemini-3-Flash-based AutoToM 80.2%. MindZero reaches 77.8% with Llama-3.2-3B, 76.2% with Llama-3.1-8B, and 72.7% with Qwen3-4B (Zhang et al., 29 May 2026).

In proactive assistance, the metric is

$\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$0

where $\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$1 is the completion time without a helper and $\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$2 is the completion time with a helper. In GridWorld assistance, random-goal helping yields 0% speedup, base Qwen3-VL-4B yields 1.4%, base Qwen3-VL-8B yields $\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$3, Qwen3-VL-235B-A22B yields 1.0%, and GPT-5.2 and Gemini-3-Flash yield 0%; MindZero yields 23.0% with Qwen3-VL-4B and 24.5% with Qwen3-VL-8B. In Household assistance, random goals yield $\mathcal{M}_t = \{ m_t^{(1)}, \dots, m_t^{(N)} \}$4, base Llama-3.1-8B yields 1.7%, base Qwen3-4B 2.3%, Qwen3-235B-A22B 12.3%, GPT-5.2 9.4%, and Gemini-3-Flash 17.7%; MindZero yields 19.1% with Qwen3-4B, 17.4% with Llama-3.1-8B, and 4.3% with Llama-3.2-3B (Zhang et al., 29 May 2026).

The reported qualitative pattern is as important as the headline metrics. Baseline LLMs often make correct guesses only near the end of an episode and produce unstable hypothesis sequences that flip across time. MindZero instead produces goal F1 and accuracy curves that rise steadily early in the episode and stabilize at high accuracy, which is what makes early assistance possible. In the household setting, the model emits structured hypotheses such as setup-table goals with plate and wineglass counts or dishwasher-loading goals with dishbowl and fork counts, each assigned a probability. This indicates that the learned posterior is not merely class-level but combinatorial over task type, object multiplicity, and target receptacle (Zhang et al., 29 May 2026).

Ablations further emphasize the role of explicit uncertainty modeling. Removing prior modeling leads to reward hacking through overly broad goals and approximately a 2% drop in speedup. Using a single hypothesis rather than multiple particles leads to approximately a 9% drop. Removing the entropy bonus causes severe mode collapse and approximately a 14% drop. These results are consistent with the framework’s formal design: assistance quality depends not only on identifying the eventually correct goal, but on maintaining a calibrated distribution over alternatives while evidence is still partial (Zhang et al., 29 May 2026).

6. Limitations, misconceptions, and terminological context

MindZero has several explicitly stated limitations. It does not implement recursive ToM of the form “I think that you think.” Its input sequence length grows with time and it does not introduce a specialized long-context architecture or memory mechanism. In the household setting, the Llama-3.2-3B backbone shows that format-alignment pre-finetuning can bias the model toward low-quality generations and reduce reinforcement-learning gains. The action-likelihood evaluator can itself be imperfect—especially when it is LLM-based—which means the learned posterior can inherit evaluator biases (Zhang et al., 29 May 2026).

A common misconception would be to treat MindZero as a supervised mental-state predictor. The framework is explicitly designed around zero annotations: there are no mental-state labels in training, only trajectories, a planner or reward model, and a structured RL objective. A second misconception would be to treat it as a model-based inference engine at deployment. The planner and reward model are used only during training; test-time inference is a single forward pass of the trained MLLM (Zhang et al., 29 May 2026).

The term “MindZero” also requires disambiguation within adjacent arXiv usage. In the supplied literature, the name is used comparatively in discussions of two unrelated systems. In "MIND: Towards Immersive Psychological Healing with Multi-agent Inner Dialogue" (Chen et al., 27 Feb 2025), “MindZero” appears as a hypothetical system label for a multi-agent inner-dialogue architecture for psychological healing. In "UMind: A Unified Multitask Network for Zero-Shot M/EEG Visual Decoding" (Xu et al., 18 Sep 2025), “MindZero” is used as a conceptual prototype name for a broader zero-shot brain-decoding system. Those uses are analogical and domain-specific; as a titled framework, MindZero refers to the 2026 self-supervised ToM system for learning online mental reasoning from behavior without annotations (Chen et al., 27 Feb 2025, Xu et al., 18 Sep 2025, Zhang et al., 29 May 2026).

Within current research positioning, MindZero sits between explicit Bayesian inverse-planning systems and prompt-only LLM ToM baselines. It inherits the probabilistic semantics of model-based ToM, amortizes them into neural parameters, and evaluates them not only in question answering but also in proactive assistance. This suggests a broader methodological direction: training large multimodal models to maintain structured distributions over latent human intentions, using only behavior and simulator- or model-derived consistency signals, rather than relying on direct annotation of mental states (Zhang et al., 29 May 2026).