Multimodal models face finer-grained understanding and execution tests

What Generation Knows reports gains from generation-side visual representations, while Code2Video-Compass finds that MLLM performance declines with task complexity and visual quality does not guarantee execution reliability. AgentVQA adds 1,000 interactive visual-understanding examples; its best model reaches 58.5%, with grounding and state tracking prominent failure modes. TreeOfLife-200M may support fine-grained biological vision, but available coverage lacks methods and benchmark evidence.