Research map / Vision-language and multimodal / Robot action and navigation
MindPower: Enabling Theory-of-Mind Reasoning in VLM-based Embodied Agents
Ruoxuan Zhang, Qiyun Zheng, Zhiyu Zhou, Ziqi Liao, Siyu Wu, Jian-Yu Jiang-Lin, Bin Wen, Hongxia Xie, Jianlong Fu, Wen-Huang Cheng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SFT on the MindPower Reasoning Hierarchy before GRPO (ablation) | Qwen2.5-VL-7B-Instruct · MindPower Benchmark action | SR: 0.28 -> 8.50 | Improved |
| add Mind-Reward GRPO after SFT (ablation) | SFT model · MindPower Benchmark action | SR: 8.50 -> 11.75 | Improved |
| add Mind-Reward GRPO after SFT (ablation) | SFT model · MindPower Benchmark action | AC: 10.48 -> 15.40 | Improved |
| add Mind-Reward GRPO after SFT (ablation) | SFT model · MindPower Benchmark decision | Sentence Transformer: 43.84 -> 47.12 | Improved |
| add Mind-Reward GRPO after SFT (ablation) | SFT model · MindPower Benchmark perception | Sentence Transformer: 52.72 -> 59.93 | Improved |
| add Mind-Reward GRPO directly to the base model without SFT (ablation) | Qwen2.5-VL-7B-Instruct · MindPower Benchmark decision | Sentence Transformer: 26.56 -> 24.68 | No clear effect |
| add MindPower Reasoning Hierarchy prompting instead of direct decision and action output (ablation) | GPT-4o · MindPower Benchmark action | accuracy: 0.82 -> 2.91 | Improved |
| add MindPower Reasoning Hierarchy instead of standard step-by-step reasoning (ablation) | GPT-4o · MindPower Benchmark action | accuracy: 0.90 -> 2.91 | Improved |
| add MindPower Reasoning Hierarchy instead of standard step-by-step reasoning (ablation) | GPT-4o · MindPower Benchmark decision | accuracy: 4.89% drop without it | Improved |
| add MindPower Reasoning Hierarchy instead of direct output (ablation) | GPT-4o · MindPower Benchmark decision | accuracy: 1.24% drop without it | Improved |
| fine-tune Qwen2.5-VL-7B with SFT plus Mind-Reward | Qwen2.5-VL-7B-Instruct (video) · MindPower Benchmark perception | Sentence Transformer: +20.04% | Improved |
| fine-tune Qwen2.5-VL-7B with SFT plus Mind-Reward | Qwen2.5-VL-7B-Instruct · MindPower Benchmark decision | Sentence Transformer: +23.33% | Improved |
Closest papers by meaning
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.90)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner (ICML 2025 · similarity 0.89)
- Video-Only ToM: Enhancing Theory of Mind in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- Hypothetical Minds: Scaffolding Theory of Mind for Multi-Agent Tasks with Large Language Models (ICLR 2025 · similarity 0.88)
- Adaptive Theory of Mind for LLM-based Multi-Agent Coordination (AAAI 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
