Director–Planner–Critic: Multi-Agent Deliberation for Subject-Consistent Video Reasoning

TL;DR: Give the model a film crew: a Director plans the shots, a Planner grounds entities and their roles, and a Critic checks global–local coherence, so the final “thought video” stays subject-consistent and logically tight. In a first study, we combine MAGUS-style agents, BindWeave grounding, and GLUS global–local checks; we expect improved reasoning and subject coherence on OpenS2V and VideoThinkBench.

Research Question: Can a modular multi-agent system that separates planning, subject grounding, and global–local verification improve the coherence and utility of video-of-thought for complex, multi-entity reasoning?

Hypothesis: Decoupling cognition (planning and grounding) from deliberation (generation plus verification) reduces entity drift and temporal incoherence, boosting both generative fidelity and downstream reasoning accuracy.

Experiment Plan: - Setup:

• Cognition: MAGUS-like Perceiver–Planner–Reflector agents co-write a structured storyboard; VG-TVP bridges text↔video via fused captioning and prompting.
• Grounding: BindWeave-style MLLM grounding disentangles roles/attributes/interactions to produce subject-aware hidden states.
• Deliberation: Video generator renders; GLUS-style Critic enforces global context (sparse keyframes) and local tracking (query frames) and flags inconsistencies for iterative revision.
• Data/Materials:
  • OpenS2V from BindWeave for subject consistency.
  • VideoThinkBench for reasoning transfer.
  • VSI-Bench for spatial memory and cognitive map evaluation.
• Measurements:
  • Subject consistency, text relevance, and naturalness (BindWeave metrics); reasoning accuracy (VideoThinkBench); spatial distance ability (VSI-Bench); ablations on each agent.
• Expected Outcomes:
  • Significant gains in subject consistency and spatial-temporal coherence with multi-agent deliberation.
  • Better generalization on multi-entity, long-horizon reasoning tasks.

References: ['Li, J., Huang, P., Li, Y., Chen, S., Hu, J., & Tian, Y. (2025). A Unified Multi-Agent Framework for Universal Multimodal Understanding and Generation. arXiv.org.', 'Li, Z., Qian, D., Su, K., Diao, Q., Xia, X., Liu, C., Yang, W., Zhang, T., & Yuan, Z. (2025). BindWeave: Subject-Consistent Video Generation via Cross-Modal Integration. arXiv.org.', 'Lin, L., Yu, X., Pang, Z., & Wang, Y.-X. (2025). GLUS: Global-Local Reasoning Unified into A Single Large Language Model for Video Segmentation. Computer Vision and Pattern Recognition.', 'Ilaslan, M., Koksal, A., Lin, K. Q., Satar, B., Shou, M. Z., & Xu, Q. (2024). VG-TVP: Multimodal Procedural Planning via Visually Grounded Text-Video Prompting. AAAI Conference on Artificial Intelligence.', 'Tong, J., Mou, Y., Li, H., Li, M., Yang, Y., Zhang, M., Chen, Q., Liang, T., Hu, X., Zheng, Y., Chen, X., Zhao, J., Huang, X., & Qiu, X. (2025). Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm.', 'Yang, J., Yang, S., Gupta, A. W., Han, R., Li, F.-F., & Xie, S. (2024). Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces. Computer Vision and Pattern Recognition.']