Human-defined creativity is highly abstract, posing a challenge for multimodal large language models (MLLMs) to comprehend and assess creativity that aligns with human judgments. The absence of an existing benchmark further exacerbates this dilemma. To this end, we propose CreBench, which consists of two key components: 1) an evaluation benchmark covering the multiple dimensions from creative idea to process to products; 2) CreMIT (Creativity Multimodal Instruction Tuning dataset), a multimodal creativity evaluation dataset, consisting of 2.2K diverse-sourced multimodal data, 79.2K human feedbacks and 4.7M multityped instructions. Specifically, to ensure MLLMs can handle diverse creativity-related queries, we prompt GPT to refine the human feedback to activate stronger creativity assessment capabilities. CreBench serves as a foundation for building MLLMs that understand human-aligned creativity. Based on the CreBench, we fine-tune open-source general MLLMs, resulting in CreExpert, a multimodal creativity evaluation expert model. Extensive experiments demonstrate that the proposed CreExpert models achieve significantly better alignment with human creativity evaluation compared to state-ofthe-art MLLMs, including the most advanced GPT-4V and Gemini-Pro-Vision.

The long-lasting global pandemic of Coronavirus disease 2019 (COVID-19) has changed our daily life in many ways and put heavy burden on our mental health. Having a predictive model of negative emotions during COVID-19 is of great importance for identifying potential risky population. To establish a neural predictive model achieving both good interpretability and predictivity, we have utilized a large-scale (n = 542) longitudinal dataset, alongside two independent samples for external validation. We built a predictive model based on psychologically meaningful resting state neural activities. The whole-brain resting-state neural activity and social-psychological profile of the subjects were obtained from Sept. to Dec. 2019 (Time 1). Their negative emotions were tracked and re-assessed twice, on Feb 22 (Time 2) and Apr 24 (Time 3), 2020, respectively. We first applied canonical correlation analysis on both the neural profiles and psychological profiles collected on Time 1, this step selects only the psychological meaningful neural patterns for later model construction. We then trained the neural predictive model using those identified features on data obtained on Time 2. It achieved a good prediction performance (r = 0.44, p = 8.13 × 10-27). The two most important neural predictors are associated with self-control and social interaction. This study established an effective neural prediction model of negative emotions, achieving good interpretability and predictivity. It will be useful for identifying potential risky population of emotional disorders related to COVID-19.