VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning

TL;DR

VideoZoomer is proposed, a novel agentic framework that enables MLLMs to dynamically control their visual focus during reasoning, and delivers diverse and complex reasoning patterns, yielding strong performance across a broad set of long video understanding and reasoning benchmarks.

Abstract

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language tasks yet remain limited in long video understanding due to the limited context window. Consequently, prevailing approaches tend to rely on uniform frame sampling or static pre-selection, which might overlook critical evidence and unable to correct its initial selection error during its reasoning process. To overcome these limitations, we propose VideoZoomer, a novel agentic framework that enables MLLMs to dynamically control their visual focus during reasoning. Starting from a coarse low-frame-rate overview, VideoZoomer invokes a temporal zoom tool to obtain high-frame-rate clips at autonomously chosen moments, thereby progressively gathering fine-grained evidence in a multi-turn interactive manner. Accordingly, we adopt a two-stage training strategy: a cold-start supervised fine-tuning phase on a curated dataset of distilled exemplar and reflection trajectories, followed by reinforcement learning to further refine the agentic policy. Extensive experiments demonstrate that our 7B model delivers diverse and complex reasoning patterns, yielding strong performance across a broad set of long video understanding and reasoning benchmarks. These emergent capabilities allow it to consistently surpass existing open-source models and even rival proprietary systems on challenging tasks, while achieving superior efficiency under reduced frame budgets.

Figures (13)

• Figure 1: Left: Conceptual comparison of three long video reasoning frameworks: (a) uniform sampling, (b) with frame selector, and (c) VideoZoomer (Ours). Right: Performance comparison of VideoZoomer against various baseline models under different frame budgets on LSDBench.
• Figure 2: VideoZoomer framework for long video reasoning. The process begins with a "Glance" where the model obtains a coarse overview of the video. It then enters an iterative "Zoom" phase, where it can invoke a <video_zoom> tool to request high-fps clips and perform multi-turn reasoning. This process continues until the model procudes a final answer or reaches max turn limit.
• Figure 3: Diverse reasoning patterns demonstrated by our model. (a) Direct-hit Reasoning, (b) Progressive Reasoning, and (c) Self-refine Reasoning.
• Figure 4: The pipeline for curating our cold-start dataset. We first distill exemplar trajectories, then generate reflection data by having an expert model correct the failures of an initial agent. The final dataset combines both verified exemplar and reflection trajectories.
• Figure 5: Training dynamics of ablation baselines. The left panel shows the average number of tool calls per sample during training. The right panel displays the model performance (e.g., accuracy) on the validation set over the course of training.