Stimulating Temporal Awareness in Egocentric Video Comprehension Models
Multimodal linguistic models (MLLMs) have recently shown strong performance in visual comprehension, however they often lack temporal awareness, especially in cognitive settings where reasoning depends on the fine-tuning and evolution of events. This deficit is due in part to training objectives that fail to explicitly reward temporal thinking and instead rely on frame-level local shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with a proven reward algorithm (RLVR) designed to promote temporal awareness in MLLMs. The results of the TGPO comparison model generated from temporal ordering versus scrambled video frames to obtain standardized, globally standardized reward signals clearly favor temporal coherence reasoning. Combined with GRPO and GSPO, TGPO supports non-preemptive RL training and effectively suppresses the behavior of local shortcuts learned by existing MLLMs. Tests across five egocentric video benchmarks show that TGPO consistently improves temporal support and causal coherence, outperforming previous RL-based video reasoning methods. Our results suggest that TGPO provides a simple and accessible approach to temporally robust MLLMs for highly focused video understanding.
- † Virginia Tech
- ‡ Harvard University
- § University of Illinois Urbana-Champaign
- UC Davis
- ** Work done while at Apple
