CLBench-V: Evaluating Multimodal Context Learning from Grounding to Knowledge Acquisition

Researchers introduced CLBench-V, a new benchmark for multimodal context learning across grounding, applying new information, and acquiring new knowledge.

The benchmark combines public and newly created datasets spanning science, finance, long documents, spatial reasoning, and web VQA.

Six recent multimodal models were tested on 3,443 examples, and overall performance was low.

Results showed no single model dominated; different systems led on different context-learning dimensions.

The findings highlight a major gap in current multimodal AI for reliable real-world use.