MindPower: Enabling Theory-of-Mind Reasoning in VLM-based Embodied Agents

AI summary

Overview

Research area: Embodied artificial intelligence, vision–language models (VLMs), and Theory of Mind (ToM) / social reasoning for agents.

Technical level: Advanced — the paper assumes familiarity with VLMs, reinforcement learning from verifiable rewards (GRPO), and embodied simulators.

Scope: The paper introduces a Robot-Centric benchmark (MindPower), a six-layer reasoning hierarchy, and a reinforcement-based reward (Mind-Reward) for training VLM-based embodied agents to infer mental states of both themselves and humans, then act on those inferences.

What This Paper Is About

Current VLM-based embodied agents can describe what they see but remain largely reactive: they do not reason about what a human believes, desires, or intends, and existing ToM benchmarks only test reasoning about the human in a video rather than from the agent's own perspective. The authors build a benchmark where an embodied agent must both infer mental states and produce decisions and executable actions, and they train a model to do so consistently. They report that their model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation.

Key Contributions

1. Robot-Centric benchmark for mental-state-grounded action. MindPower links mental reasoning to embodied action through two tasks — False-Belief Correction and Implicit Goal Inference & Completion — across 590 interactive home scenarios, evaluating whether agents can infer, decide, and assist.
2. A unified MindPower Reasoning Hierarchy. The hierarchy structures reasoning into three levels and six layers (<Perception>, <Belief>, <Desire>, <Intention>, <Decision>, <Action>), providing a standardized way to trace how perception leads to action.
3. Mind-Reward, a reinforcement optimization objective for consistent ToM reasoning. It aligns intermediate reasoning states with final actions to promote coherent, Robot-Centric reasoning.
4. A trained model and evaluation suite. Using Supervised Fine-Tuning plus GRPO with Mind-Reward, the model surpasses GPT-4o by 12.77% in decision accuracy and 12.49% in action generation.

Main Findings

• Humans still lead. Nine trained participants watched the collected videos and produced BDI reasoning, decisions, and actions; human participants achieved the highest scores, clearly outperforming all VLMs.
• Closed-source VLMs lead among untrained models. Gemini-2.5 Pro and GPT-4o achieved the highest scores in Perception, Mental Reasoning, and Decision Making and Action; among open-source VLMs, reasoning-capable models such as Video-R1 and VideoChat-R1 performed best.
• The reasoning hierarchy is necessary. Removing Level-1 and Level-2 and asking GPT-4o to output Decision and Action directly dropped decision-making accuracy by 1.24%, and action generation accuracy fell from 2.91% to 0.82%. Replacing the hierarchy with standard <think>…</think> reasoning dropped decision accuracy by 4.89% and action accuracy from 2.91% to 0.90%.
• VLMs lack a Robot-Centric perspective. At the Perception level, VLMs give general descriptions of clothing or environment instead of focusing on individual actions, missing movements, directions, and appearance sequences; at higher layers they are biased by the environment (for example, predicting cleaning kitchenware when the person is searching for a moved item, or assuming bed-tidying when the person is only retrieving something).
• Zero action scores for several models. InternVL3.5-8B, LLaVA-OV-8B, and Video-ChatGPT obtained zero scores on Success Rate (SR) and Action Correctness (AC) because their outputs contain non-executable expressions such as identify() or scan() rather than concrete goal-directed commands.
• Large gains over the base model. Compared with Qwen2.5-VL-7B-Instruct, the proposed model achieves a +20.04% improvement in Sentence Transformer score for perception and a +23.33% gain in the <Decision> layer; SR increases by 11.6% and AC improves by 15.25%.
• Ablation on training stages. Using only SFT yields improvement, showing the hierarchy helps mental reasoning and decision making; adding Mind-Reward improves further. Without SFT, Mind-Reward alone gives some improvement in decision and action accuracy over the initial Qwen2.5-VL-7B-Instruct, but overall performance remains suboptimal, indicating SFT is needed for cold-start training.

Methodology in Plain English

Benchmark design. The authors extend the Belief–Desire–Intention (BDI) cognitive framework into a six-layer reasoning hierarchy spanning three levels: Level-1 Perception, Level-2 Mental Reasoning (<Belief>, <Desire>, <Intention>), and Level-3 Decision Making and Action (<Decision>, <Action>). Unlike prior benchmarks that infer only the human character's mental state through multiple-choice questions (Role-Centric), MindPower requires the agent to infer both its own and the human's mental states (Robot-Centric) and to produce open-ended decisions and actions.

Data collection. Scenarios are simulated in VirtualHome and ThreeDWorld. Story scripts are generated by GPT-4o based on room type, character setup, goals, and objects, then manually filtered by five annotators for realism, BDI consistency, and diversity. Each script is reenacted in the simulators — a process taking 25–35 minutes per sample in VirtualHome and 50–70 minutes in ThreeDWorld — yielding 590 examples. Five trained annotators label all six layers for every sample. The dataset contains 390 False-Belief Correction examples and 200 Implicit Goal Inference & Completion examples, spanning 2 simulators, 8 home layouts, and 16 humanoid agents including children, adults, and wheelchair users.

Metrics. Perception, Belief, Desire, Intention, and Decision outputs are scored with BERTScore and Sentence Transformer similarity (normalized 0–100). Actions are scored with Success Rate (SR = (2R1 + 3R2 + 5RL)/10) and Action Correctness (AC = floor(|A* ∩ Â| / |Â|)). A BDI and Perspective Consistency (BPC) score from 0 to 10 is judged by GPT-4o on logical coherence, completeness, and whether reasoning adopts the robot's perspective.

Training. The authors use a two-stage paradigm: Supervised Fine-Tuning first, followed by Group Relative Policy Optimization (GRPO). The reward combines Mind-Reward (a weighted sum of ROUGE-1 atomic accuracy, ROUGE-2 local consistency, and ROUGE-L global consistency, with weights α1 = 0.2, α2 = 0.3, α3 = 0.5) and Format-Reward (1 if all six layers appear in the correct order, otherwise 0). Ground-truth and generated outputs are converted into structured atomic actions — in the form action(agent, object) — by Qwen3-Max during training. The base model is Qwen2.5-VL-7B-Instruct; 32 frames are extracted per video, training uses 5 SFT epochs and 400 GRPO iterations with 8 generations, on a single H800 GPU. The dataset is split 8:2 into training and testing sets.

Why This Matters

For research, MindPower shifts ToM evaluation from multiple-choice questions about a video character toward open-ended, self-perspective reasoning that terminates in executable actions. The comparison table positions it against earlier benchmarks such as Hi-ToM, BigToM, FANToM, MuMA-ToM, MMToM-QA, GridToM, and SoMi-ToM, which the authors characterize as Role-Centric and lacking Level-3 ability.

Real-world applications implied by the paper:

• Assistive robotics for users with mobility or reach constraints — the dataset includes wheelchair users and a 1.2-meter-tall child, testing whether an agent infers a hidden goal from context rather than explicit commands.
• Household assistance grounded in false-belief correction — for example, informing a returning person that an object was moved, even when the agent cannot see where it went.
• Proactive human–agent collaboration — agents that infer unstated intentions from searching behavior or repeated failed attempts, such as recognizing that someone is looking for a knife to finish cutting carrots.
• Safety and property monitoring — detecting implicit goals from object properties, such as a person leaving a leaking faucet running, or searching for candles during a power outage.

Industry relevance centers on training VLM-based agents for service and home robotics, where models must produce executable commands (open(fridge), pick_up(milk)) rather than descriptive or non-executable outputs, and on the use of verifiable reward signals to align reasoning with action.

Future Directions

• Extending the MindPower Reasoning Hierarchy to human–robot collaboration and multi-agent coordination.
• Deploying the trained model on real robots to assess performance in practical settings.
• Multi-view extension of the dataset, addressing cases where a fixed camera cannot capture an object's final location.
• Connecting the benchmark to low-level execution models, as discussed in the supplementary material on extensions.

Target Audience

Researchers and engineers working on embodied AI, VLM-based agents, social intelligence, and human–robot interaction, particularly those building benchmarks or reward-based training pipelines for reasoning-grounded action. It is also relevant to practitioners who need agents that infer unstated human goals and act proactively in household or assistive settings. Readers without a background in reinforcement learning and multimodal models will find the method sections demanding.

Authors’ abstract

Theory of Mind (ToM) refers to the ability to infer others' mental states, such as beliefs, desires, and intentions. Current vision-language embodied agents lack ToM-based decision-making, and existing benchmarks focus solely on human mental states while ignoring the agent's own perspective, hindering coherent decision and action generation. To address this, we propose MindPower, a Robot-Centric framework integrating Perception, Mental Reasoning, Decision Making and Action. Given multimodal inputs, MindPower first perceives the environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Furthermore, we introduce Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior. Our model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation.