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

arXiv 2026 40.1 method, system

TLDR

CoWorld Reasoner is an agentic framework that uses video world models as visual reasoning engines with co-evolving tools and skills, improving multimodal reasoning across benchmarks.

Reasoning

The paper presents a strong empirical evaluation across seven benchmarks and an original co-evolution method that updates tools and skills without changing model weights. However, the abstract gives little detail on limitations, implementation costs, or potential instability in the co-evolution process.

Read-first score

Read-first score 40.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 44.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,benchmark,evaluation,result

Topical relevance 25%

62.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Graph prestige 15%

10.9

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0031

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 471.

Keyword Scores

video world model

world model

generative world model

interactive world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Agentic framework using a Video World Model as a unified visual reasoning engine for flexible visual imagination
Video World Model-Driven Harness exposing multiple visual operations as tools and using skills to guide VLM decisions on whether to imagine, operation selection, request construction, and evidence interpretation
VLM-VWM Harness Co-Evolution jointly revising tools, skills, and execution code based on execution feedback while keeping model weights fixed
Supports flexible visual imagination beyond fixed pipelines such as novel-view synthesis

Methodology

CoWorld Reasoner couples a VLM with a Video World Model through a harness of tools and skills, where the VLM decides whether to imagine, selects a visual operation, constructs generation requests, and interprets generated visual evidence. Tools, skills, and execution code are jointly revised via execution feedback while VLM and VWM weights remain fixed. Evaluation uses seven benchmarks spanning spatial intelligence, temporal and causal reasoning, and perceptual reasoning, with baselines including direct VLM reasoning and five evolving baselines using the same backbone.

Key Results

CoWorld Reasoner outperforms direct reasoning with the same VLM; the strongest variant improves on all seven benchmarks, with relative accuracy gains of 11.8% on MMSI and 13.0% on ERQA, and outperforms all five evolving baselines using the same backbone across five benchmarks.

Resources

Paper