Official Review · ICLR 2026

This paper presents HUMANVBENCH, a benchmark for evaluating human-centric video understanding in multimodal large language models (MLLMs). It focuses on assessing fine-grained abilities such as emotion perception, behavior analysis, and audio-visual alignment, which are often neglected in existing benchmarks.The benchmark includes 16 tasks covering multiple aspects of human understanding and is used to evaluate 27 leading MLLMs.Results show that current models, including strong proprietary ones like Gemini-2.5-Pro and GPT-5, still fall short of human-level comprehension, especially in emotional and cross-modal reasoning. Overall, HUMANVBENCH offers a scalable and fine-grained evaluation framework that advances the study of human-centric intelligence in MLLMs, though some methodological details could be clarified further. 1.The paper investigates a meaningful research problem, focusing on the fine-grained evaluation of MLLMs in understanding human emotions, behaviors, and audio-visual alignment. 2.It provides a detailed introduction to the automated annotation and QA generation framework. 3.The benchmark is comprehensive, covering 16 tasks and evaluating 27 MLLMs, offering systematic and broad assessment. 1.The paper introduces an automated annotation and QA generation pipeline, but its scalability and effectiveness for large-scale training data or model performance improvement are not empirically demonstrated. 2.The related work section discusses several existing benchmarks, yet the comparison remains qualitative and lacks a systematic or quantitative analysis to clearly position HUMANVBENCH. 3.The discussion of model failure cases is brief and lacks deeper insights into the underlying causes, reducing the benchmark’s diagnostic and interpretive value. Please refer to the reference.