TimeThink: Reasoning with Time for Video LLMs

Video Large Language Models (Video-LLMs) often struggle with reliable reasoning over long temporal sequences, frequently relying on global semantic impressions rather than precise temporal evidence. The authors investigate whether reinforcement learning (RL) can be improved by providing explicit, step-wise guidance for temporal evidence discovery during the reasoning process, rather than relying solely on outcome-based rewards for the final answer.

The authors propose TimeThink, a reinforcement learning framework that structures video reasoning as a sequence of temporal clue steps. Each step requires the model to reference a specific time interval in the video. To train this, they introduce a step-wise temporal process reward based on the Intersection-over-Union (IoU) between the model's referenced interval and ground-truth evidence segments. This is combined with a joint process-outcome optimization objective, where the model is trained using the Group Relative Policy Optimization (GRPO) algorithm. To support this, they constructed TimeThink-RFT-20K, a dataset of 20,000 video question-answer pairs with automatically derived temporal evidence segments.

TimeThink consistently outperforms existing open-source RL-based video models across a wide range of benchmarks, including video reasoning, temporal grounding, and general video understanding. By incorporating step-wise process rewards, the model achieves significantly higher temporal localization accuracy (e.g., on Charades-STA) and demonstrates better visual fidelity in its reasoning trajectories. The authors show that this process-oriented supervision is more effective than outcome-only rewards, as it prevents reward hacking and encourages the model to ground its conclusions in specific, verifiable video segments.

This work demonstrates that the reasoning capabilities of Video-LLMs can be substantially enhanced by moving beyond simple outcome-based reinforcement learning. By forcing the model to explicitly ground its intermediate reasoning steps in temporal evidence, TimeThink bridges the gap between high-level language reasoning and low-level visual localization, providing a scalable and effective path toward more reliable and interpretable video understanding systems.