CoWorld Reasoner: Agentic Multimodal Reasoning with the Evolving Video World Model-Driven Harness
The strongest variant improved accuracy across seven benchmarks without changing either model’s weights.
CoWorld Reasoner lets a vision-language model request generated visual evidence from a video world model to help solve reasoning tasks. Tools and skills guide whether to generate that evidence, which operation to use, and how to interpret the result. Execution feedback drives revisions to those tools, skills and execution code. The authors report relative accuracy gains of 11.8% on MMSI and 13.0% on ERQA over direct reasoning with the same vision-language model.
HF Daily Papers' note
CoWorld Reasoner lets a vision-language model request generated visual evidence from a video world model to help solve reasoning tasks. Tools and skills guide whether to generate that evidence, which operation to use, and how to interpret the result. Execution feedback drives revisions to those tools, skills and execution code. The authors report relative accuracy gains of 11.8% on MMSI and 13.0% on ERQA over direct reasoning with the same vision-language model.
HF Daily Papers' note
score 5