Research map / Vision-language and multimodal / Multimodal LLMs
From Black Boxes to Transparent Minds: Evaluating and Enhancing the Theory of Mind in Multimodal Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds GridToM, a 2D gridworld multimodal ToM benchmark, probes attention heads of MLLMs for belief representations, and applies training-free activation intervention to improve first- and second-order belief accuracy.
Problem
Existing ToM evaluations of multimodal LLMs are black-box QA tests confounded by hallucination and scenario understanding, and real-world video datasets lack per-agent perceptual information.
Approach
Construct a grid-world multimodal ToM dataset with per-agent perspective annotations, use linear probes on attention-head outputs to find belief-sensitive heads, and steer those heads along probe directions at inference time without training.
Key results
- Intervention on LLaVA-Next-Video raises multimodal first-order Both accuracy from 0.8% to 22.0% and second-order Both from 0.2% to 10.2%.
- Intervention on Qwen2-VL-7B-Instruct raises multimodal first-order TB accuracy from 26.6% to 60.4% and second-order Both from 15.4% to 24.2%.
- Human participants reach about 99.8% on multimodal first-order Both, versus at most 23.6% for baseline MLLMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add protagonist-perspective attention steering to LLaVA-NeXT-Video-7B-HF | LLaVA-NeXT-Video-7B-HF · GridToM first-order multimodal | Both accuracy: 0.8 -> 22.0 (+21.2) | Helped |
| Add protagonist-perspective attention steering to LLaVA-NeXT-Video-7B-HF | LLaVA-NeXT-Video-7B-HF · GridToM first-order multimodal | TB accuracy: 53.2 -> 63.8 (+10.6) | Helped |
| Add protagonist-perspective attention steering to LLaVA-NeXT-Video-7B-HF | LLaVA-NeXT-Video-7B-HF · GridToM second-order multimodal | Both accuracy: 0.2 -> 10.2 (+10.0) | Helped |
| Add protagonist-perspective attention steering to LLaVA-NeXT-Video-7B-HF | LLaVA-NeXT-Video-7B-HF · GridToM first-order video-only | Both accuracy: 0.4 -> 16.4 (+16.0) | Helped |
| Add protagonist-perspective attention steering to LLaVA-NeXT-Video-7B-HF | LLaVA-NeXT-Video-7B-HF · GridToM second-order video-only | Both accuracy: 0.3 -> 12.3 (+12.0) | Helped |
| Add protagonist-perspective attention steering to Qwen2-VL-7B-Instruct | Qwen2-VL-7B-Instruct · GridToM first-order multimodal | TB accuracy: 26.6 -> 60.4 (+33.8) | Helped |
| Add protagonist-perspective attention steering to Qwen2-VL-7B-Instruct | Qwen2-VL-7B-Instruct · GridToM first-order multimodal | Both accuracy: 23.6 -> 31.2 (+7.6) | Helped |
| Add protagonist-perspective attention steering to Qwen2-VL-7B-Instruct | Qwen2-VL-7B-Instruct · GridToM second-order multimodal | Both accuracy: 15.4 -> 24.2 (+8.8) | Helped |
| Add protagonist-perspective attention steering to Qwen2-VL-7B-Instruct | Qwen2-VL-7B-Instruct · GridToM first-order video-only | Both accuracy: 5.2 -> 18.6 (+13.4) | Helped |
| Add protagonist-perspective attention steering to Qwen2-VL-7B-Instruct | Qwen2-VL-7B-Instruct · GridToM second-order video-only | Both accuracy: 2.1 -> 19.5 (+17.4) | Helped |
| Vary the intervention hyperparameters K and α within the valid range (ablation) | Attention-steering intervention · GridToM first-order TB and FB tasks | Accuracy: no more than 10% on average | No clear effect |
| Steer in the omniscient-perspective direction instead of the protagonist-perspective direction (ablation) | Attention-steering intervention · GridToM | Accuracy | Hurt |
Surprising findings
- MLLMs achieve near-100% on false-belief questions but very low accuracy on true-belief questions, suggesting overreliance on FB patterns.
- Deepseek-vl2-small produced invalid responses for 89.9% of initial belief tasks.
Limitations
- Only first- and second-order belief tasks within the ATOMS framework are covered.
- Approach validated on only a limited selection of MLLMs due to restricted access to model code.
- Dataset is synthetic 2D gridworld with moderate difficulty, limiting ecological validity.
Experimental setup
Datasets: GridToM MMToM-QA
Benchmarks: GridToM MMToM-QA
Metrics: accuracy TB accuracy FB accuracy both accuracy
Baselines: GPT-4o Doubao-1.5-vision-pro DeepSeek-VL2-small LLaVA-Next-Video-7B Qwen2-VL-7B-Instruct DeepSeek-V3 Llama-3.3-70B-Instruct Mistral-7B-Instruct-v0.3
Builds on: Multigrid Minigrid MMToM-QA unexpected transfer task ATOMS
Closest papers by meaning
- Video-Only ToM: Enhancing Theory of Mind in Multimodal Large Language Models (CVPR 2026 · similarity 0.93)
- Unveiling Visual Perception in Language Models: An Attention Head Analysis Approach (CVPR 2025 · similarity 0.89)
- Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner (ICML 2025 · similarity 0.89)
- Unveiling the Cognitive Compass: Theory-of-Mind–Guided Multimodal Emotion Reasoning (ICLR 2026 · similarity 0.89)
- Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning (ICLR 2024 · similarity 0.88)
- LENS: Multi-level Evaluation of Multimodal Reasoning with Large Language Models (ICLR 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.