General-Level & General-Bench

A capability taxonomy and mega-benchmark for multimodal generalists (ICML 2025 Oral): General-Level defines a 5-level scale for MLLM generality — scoring whether understanding and generation, across modalities, reinforce each other (with "synergy" as the core criterion) — and General-Bench operationalizes it with 700+ tasks across comprehension, generation, and modality-spanning skills, one of the broadest multimodal eval suites assembled.

Led by NUS NExT++ (Hao Fei, Tat-Seng Chua) with a large multi-institution author list — the evaluation counterpart of the lab's NExT-GPT/Vitron generalist-model agenda.