The field of multimodal models is moving towards improving temporal understanding and reasoning capabilities. Recent research has highlighted the limitations of current models in understanding temporal dynamics, such as telling time on analog clocks or understanding causality in videos. To address these limitations, researchers are developing new benchmarks and evaluation methods that can assess a model's ability to reason about time and temporal relationships. These benchmarks, such as VBenchComp and TimeCausality, enable fine-grained evaluation of different capabilities of multimodal models and reveal nuanced model weaknesses that are hidden by traditional overall scores. Furthermore, techniques such as panoramic direct preference optimization (PanoDPO) are being proposed to enhance the robustness of large multimodal models against temporal inconsistency.

Noteworthy papers include: Investigating and Enhancing the Robustness of Large Multimodal Models Against Temporal Inconsistency, which proposes a novel temporal robustness benchmark and a method to enhance model robustness. TimeCausality: Evaluating the Causal Ability in Time Dimension for Vision Language Models, which introduces a novel benchmark to evaluate the causal reasoning ability of vision-language models in the temporal dimension.