World Knowledge Benchmark for Multimodal Language Models

WorldVQA is a benchmark dataset created by MoonshotAI to evaluate atomic vision-centric world knowledge in Multimodal Large Language Models (MLLMs). It was last updated in February 2026. The dataset decouples visual knowledge retrieval from reasoning to provide a strict measurement of a model's fundamental world knowledge.

Use Cases

• Benchmarking MLLM performance on atomic world knowledge questions using image-question-answer triplets.
• Analyzing the gap between visual knowledge retrieval and reasoning capabilities in vision-language models.
• Training or fine-tuning models to improve performance on specific, factual visual question-answering tasks.

Strengths

• Benchmark designed to decouple and strictly measure atomic world knowledge, a specific evaluation focus.
• Dataset was updated in February 2026, indicating recent maintenance.

Limitations

• Specific dataset size, row count, and column details are unknown.
• The scope of world knowledge (e.g., geographic or temporal coverage) is not specified.

Provenance

Source

MoonshotAI via Hugging Face.

Collection Method

null

Time Range

null

Freshness

Last updated 2026-02-04.

Geography

null