Temporal-Aware Reasoning Optimization for Video Temporal Grounding
2026-06-08 • Computer Vision and Pattern Recognition
Computer Vision and Pattern Recognition
AI summaryⓘ
The authors created a method called TaRO to help multi-modal language models better understand and locate specific times in videos. They noticed current models don't think deeply enough and often guess randomly or only check if the answer is right, not how well they reason. TaRO uses detailed captions to build clear reasoning steps tied to actual video events and rewards the model for focusing on exact moments. Their approach gradually teaches the model to improve its reasoning, leading to better performance on video temporal grounding tests.
Multi-modal Large Language ModelsVideo Temporal GroundingReinforcement LearningReasoning PathsTemporal LocalizationDense CaptionsReward FunctionsCurriculum LearningTemporal-Sensitivity RewardExploration Strategies
Authors
Minghang Zheng, Zihao Yin, Yi Yang, Yuxin Peng, Yang Liu
Abstract
Multi-modal Large Language Models (MLLMs) have achieved remarkable progress in video temporal grounding with reinforcement learning for generating reasoning paths. However, existing models often produce superficial reasoning, which offers limited guidance for precise temporal localization. This limitation stems from (1) inefficient random exploration and (2) reward functions that focus solely on the answer correctness while ignoring reasoning quality. To address these issues, we propose TaRO (Temporal-Aware Reasoning Optimization), a framework that explicitly enhances the model's ability of thinking with time. First, we introduce a Constructive Reasoning Exploration that leverages pre-generated dense captions to construct reasoning paths grounded in explicit visual cues and timestamps, enabling efficient exploration of high-quality time-aware reasoning. Second, to evaluate reasoning quality, we design a Temporal-Sensitivity Reward. High-quality reasoning should be anchored to specific events and timestamps. If the event boundary under thinking is disrupted, such reasoning should become invalid, leading to a drop in the logit of the reasoning path. We utilize this drop as a critique of reasoning quality. Finally, TaRO follows a progressive curriculum, which starts by utilizing this reward to select better constructed reasoning paths, and evolves to a free exploration phase where the model autonomously generates effective reasoning. Experiments demonstrate that TaRO achieves state-of-the-art performance on VTG benchmarks. Code is available at https://github.com/oceanflowlab/TaRO.