VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning

Yang Ding, Yizhen Zhang, Xin Lai, Ruihang Chu, Yujiu Yang

Introduction

Multimodal Large Language Models (MLLMs) have shown remarkable capabilities in image and short video understanding. However, long video reasoning remains a significant challenge. The primary bottleneck is the limited context window of these models, which necessitates aggressive downsampling of video content. Traditional "passive" approaches, such as uniform frame sampling, suffer from significant information loss—often missing critical, transient events—and lack a mechanism for iterative refinement. If the initial sampling misses the key evidence, the model cannot "go back and look again."

To address these limitations, we present VideoZoomer, an agentic framework that enables MLLMs to dynamically control their visual focus. By reframing video reasoning as an interactive, multi-turn task, VideoZoomer allows models to "zoom in" on specific temporal segments, achieving high-resolution understanding with minimal computational overhead.

The VideoZoomer Framework: Glance and Zoom

VideoZoomer operates on a "First Glance, Then Zoom" strategy. Instead of processing a massive number of frames in a single pass, the model follows an iterative reasoning loop:

  1. Glance: The model begins with a coarse, low-frame-rate overview of the entire video to establish context.
  2. Zoom: Based on the initial glance and the specific question, the model identifies segments requiring closer inspection. It invokes a specialized <video_zoom> tool to retrieve high-frame-rate clips for these specific time intervals (e.g., [12.0s, 14.0s]).
  3. Reason: Through multi-turn Chain-of-Thought (CoT) reasoning, the model aggregates evidence turn-by-turn. It can perform multiple zooms if the first attempt is insufficient, continuing until it can confidently provide an answer.
The multi-turn 'Glance and Zoom' reasoning pipeline.

This active agency allows the model to allocate its "frame budget" intelligently, focusing on critical moments rather than wasting resources on redundant background footage.

Training Methodology

Developing a robust video agent requires more than standard supervised fine-tuning (SFT). VideoZoomer employs a two-stage training strategy:

Cold-Start SFT with Reflection

The model is initially trained on 11,000 "golden" trajectories distilled from proprietary models like GPT-4o and Gemini. A key innovation here is the inclusion of reflection data. These trajectories include instances where an agent initially failed (e.g., zooming into the wrong segment), reflected on the error, and corrected its tool call. This teaches the model that failure is part of the search process and enables robust, multi-turn investigation.

Reinforcement Learning (RL)

To transform the model into a truly adaptive agent, we utilize Group Relative Policy Optimization (GRPO). The reward function is designed to balance three objectives: accuracy in the reasoning task, adherence to precise tool-use syntax, and exploration efficiency (finding evidence without excessive redundancy).

Experimental Evaluation and Results

VideoZoomer was evaluated across a wide range of benchmarks, including MLVU, LongVideoBench, VideoMME, and a specialized reasoning benchmark, LongVideoReason-eval.

Superior Reasoning Performance

On the LongVideoReason-eval benchmark, VideoZoomer (7B) achieved a score of 80.3, significantly outperforming much larger proprietary models such as GPT-4o (60.7) and Gemini-1.5-Pro (67.3). Compared to its base model (Qwen2.5-VL-7B), VideoZoomer showed a massive +10.5 point improvement on the MLVU benchmark.

Frame Efficiency and Temporal Precision

One of the most striking results is the model's efficiency. VideoZoomer achieves higher accuracy with significantly fewer input frames than traditional uniform sampling methods. On MLVU, it reached peak performance using an average of only 48 frames, whereas baseline models required 128 frames or more to achieve lower scores.

Performance comparison on LSDBench showing accuracy vs. input frames.

This temporal precision is best exemplified in the "Action Count" task (MLVU), where the model's ability to zoom into fast-moving events led to a score increase from 13.6 to 50.5.

Conclusion and Future Work

VideoZoomer demonstrates that active, agentic control over visual input is a more effective paradigm for long video understanding than passive sampling. By learning to "zoom" and reflect on its own search process, the model achieves state-of-the-art reasoning performance with high efficiency. Future extensions include incorporating spatial zooming to allow the model to crop specific regions of a frame, further enhancing its investigative capabilities.

Summary updated

BibTeX
@inproceedings{ding2026videozoomer,
  author = {Yang Ding and Yizhen Zhang and Xin Lai and Ruihang Chu and Yujiu Yang},
  title = {{VideoZoomer: Reinforcement-Learned Temporal Focusing for Long Video Reasoning}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=ARHCFvgx6G},
}
Figure