Temporal Contrastive Learning for Video Temporal Reasoning in Large Vision-Language Models

TL;DR

This work tackles the difficulty of temporal reasoning in video–language understanding by introducing Temporal Semantic Alignment via Dynamic Prompting (TSADP). TSADP combines a Dynamic Prompt Generator (DPG) to encode fine‑grained temporal cues and a Temporal Contrastive Loss (TCL) to align visual and textual embeddings over time, supplemented by a Masked Temporal Prediction objective for robustness. Evaluated on an extended VidSitu dataset, TSADP outperforms state‑of‑the‑art baselines across intra‑video entity association, temporal relationship understanding, and chronology prediction, with human evaluations confirming coherent and semantically accurate temporal descriptions. The method demonstrates robustness and efficiency, offering a practical step forward for real‑world video–language reasoning and description generation, with potential extensions to longer sequences and integration of external knowledge.

Abstract

Temporal reasoning is a critical challenge in video-language understanding, as it requires models to align semantic concepts consistently across time. While existing large vision-language models (LVLMs) and large language models (LLMs) excel at static tasks, they struggle to capture dynamic interactions and temporal dependencies in video sequences. In this work, we propose Temporal Semantic Alignment via Dynamic Prompting (TSADP), a novel framework that enhances temporal reasoning capabilities through dynamic task-specific prompts and temporal contrastive learning. TSADP leverages a Dynamic Prompt Generator (DPG) to encode fine-grained temporal relationships and a Temporal Contrastive Loss (TCL) to align visual and textual embeddings across time. We evaluate our method on the VidSitu dataset, augmented with enriched temporal annotations, and demonstrate significant improvements over state-of-the-art models in tasks such as Intra-Video Entity Association, Temporal Relationship Understanding, and Chronology Prediction. Human evaluations further confirm TSADP's ability to generate coherent and semantically accurate descriptions. Our analysis highlights the robustness, efficiency, and practical utility of TSADP, making it a step forward in the field of video-language understanding.