CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness

CoWorld Reasoner. Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness. By Dongjie Cheng and five others, from the Hong Kong Polytechnic University and Hangzhou Diagens Biotechnology. Published on the fifteenth of October. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

As vision-language models advance, they are increasingly expected to reason about motion, actions and view changes that go beyond what the input images show. Video world models provide a natural, unified engine for that kind of visual imagination, generating task-relevant visual content as additional evidence. Existing approaches, however, typically invoke a video world model through a fixed pipeline, centred on one predefined form of imagination, such as novel-view synthesis. Whatever the task, the imagination on offer is the same. A question about how a scene would look from another side and a question about what happens after an action need different things imagined, and a pipeline built for one does not serve the other.

CoWorld Reasoner is an agentic framework that uses a video world model as a unified visual reasoning engine, so that imagination can take whatever form a given reasoning task needs. The reasoning model decides when and how to imagine, instead of a fixed pipeline deciding for it.

At its centre is the Video World Model-Driven Harness. It exposes multiple visual operations through tools, and uses skills to guide the VLM through a sequence of decisions: whether to imagine at all, which operation to select, how to construct a generation request, and how to interpret the visual evidence that comes back. On top of that sits VLM-VWM Harness Co-Evolution, which jointly revises the tools, the skills and the execution code based on execution feedback. Both models keep their weights fixed throughout. Only the harness around them evolves. What improves is the interface between the two models: the operations on offer, the guidance on when to use them, and the code that carries them out.

The evaluation spans seven benchmarks, covering spatial intelligence, temporal and causal reasoning, and perceptual reasoning. CoWorld Reasoner outperforms direct reasoning with the same VLM. The strongest variant improves on all seven benchmarks, including relative accuracy gains of eleven point eight percent on MMSI and thirteen percent on ERQA. It also outperforms all five evolving baselines that use the same backbone, across five benchmarks.

Two things become flexible here without a single weight changing. Imagination turns into a tool the reasoner chooses to call, with the form of that imagination chosen per task. And the harness that wraps both models improves from its own execution feedback, so the system gets better at using the world model without either model being retrained.