Independent author · NeurIPS 2026 Workshop (NeurReps), Extended Abstract Track · poster
Selective Cross-Modal Alignment in Vision-Language Models: Why Position Grounding Succeeds but Physical Reasoning Fails
Qwen2.5-VL-3B on 6,000 controlled physical videos. Position is linearly decodable (R² = 0.97) and language grounding is 100%; velocity is only implicitly encoded in an entangled manifold and is not used by default. Prompting improves physical reasoning by 62%.
Read the paper