As Vision‑Language Models (VLMs) become more capable, they are expected to reason about motion, actions, and viewpoint changes over time, beyond what static images provide. Video World Models (VWMs) serve as a unified engine for visual imagination by generating task‑relevant visual content as supplemental evidence. Existing approaches typically invoke a VWM through fixed pipelines that focus on a single form of imagination, such as novel‑view synthesis, limiting flexibility across diverse reasoning tasks.
CoWorld Reasoner introduces an agentic framework that treats a VWM as a universal visual reasoning engine. Its Video World Model‑Driven Harness exposes multiple visual operations as tools and employs a set of skills to guide the VLM in deciding whether to imagine, selecting an operation, constructing a generation request, and interpreting the resulting visual evidence.
To further improve adaptability, the authors propose VLM‑VWM Harness Co‑Evolution: execution feedback is used to iteratively revise tools, skills, and execution code while keeping model weights frozen.
Across seven benchmarks covering spatial intelligence, temporal‑causal reasoning, and perceptual reasoning, CoWorld Reasoner consistently outperforms direct reasoning with the same VLM. The strongest variant gains 11.8% relative accuracy on MMSI and 13.0% on ERQA, and surpasses all five evolving baselines that share the same backbone on five benchmarks.
Review