AVIC spatial-reasoning work accepted to NeurIPS 2026

The AVIC team says its spatial-reasoning work has been accepted to NeurIPS 2026. It describes a reinforcement-learning policy that selectively uses a video world model. In an earlier update, the team said its AVIC-R extension improved spatial reasoning and that a Qwen2.5-VL-7B policy trained with it outperformed GPT-4o and GPT-4 as policy models.