Symbolic Reasoning for Spatial Supersensing: A Neuro-Symbolic Hybrid Approach to Video World Models

TL;DR: What if models could reason about space and events in video by using not just neural networks, but also explicit symbolic logic (like flowcharts or equations)? For a first step, couple Cambrian-S with a symbolic reasoning module for 3D inference and event chaining, and compare on VSR/VSC.

Research Question: Can neuro-symbolic architectures, which integrate traditional deep-predictive video models with symbolic logical modules, improve robustness and generalizability in spatial supersensing tasks that require world modeling or multi-step event reasoning?

Hypothesis: Symbolic modules can provide invariance (such as handling 3D relationships or causal event chains) that neural networks alone fail to generalize, especially in out-of-distribution or complex event scenarios as highlighted by GSM-Symbolic and UI2V-Bench.

Experiment Plan: - Architecture: Combine a vision backbone (Cambrian-S) with a symbolic engine (e.g., for spatial logic or temporal reasoning), using interfaces similar to those in GSM-Symbolic or UI2V-Bench’s feedback pipeline.

• Data: Use VSI-SUPER and the new Event-of-Thought dataset, as well as GSM-Symbolic for controlled reasoning sub-tasks.
• Measure: Improvement in generalizability, sample efficiency, and reasoning chain accuracy compared to pure neural approaches.
• Expected Outcome: The hybrid should outperform in tasks needing compositional logic, such as multi-step event recall or world-state inference after perturbations.

References: 1. Yang, S., Yang, J., Huang, P., Brown, E., Yang, Z., Yu, Y., Tong, S., Zheng, Z., Xu, Y., Wang, M., Lu, D., Fergus, R., LeCun, Y., Li, F., & Xie, S. (2025). Cambrian-S: Towards Spatial Supersensing in Video.
2. Mirzadeh, I., Alizadeh-Vahid, K., Shahrokhi, H., Tuzel, O., Bengio, S., & Farajtabar, M. (2024). GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models. International Conference on Learning Representations.
3. Zhang, A., Lei, L., Kong, D., Wang, Z., Xu, J., Song, F., Guo, C.-L., Liu, C., Li, F., & Chen, J. (2025). UI2V-Bench: An Understanding-based Image-to-video Generation Benchmark. arXiv.org.