FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting

• URL: http://arxiv.org/abs/2509.24304v2
• Date: Tue, 30 Sep 2025 01:55:37 GMT
• Title: FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting
• Authors: Zefeng He, Xiaoye Qu, Yafu Li, Siyuan Huang, Daizong Liu, Yu Cheng,
• Abstract summary: This paper introduces the concept of thinking with long videos and proposes a novel framework FrameThinker.<n>We show that FrameThinker achieves a significant average improvement of +10.4% over baselines while drastically reducing the number of processed frames.<n>Most notably, our 7B model, FrameThinker establishes a new state-of-the-art on LongVideo-Reason, achieving 76.1% accuracy using an average of only 20.6 frames.
• Score: 62.25888935329454
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: While Large Vision-Language Models (LVLMs) have achieved substantial progress in video understanding, their application to long video reasoning is hindered by uniform frame sampling and static textual reasoning, which are inefficient and struggle to handle visually intensive video tasks. To overcome these challenges, in this paper, we introduce the concept of thinking with long videos and propose a novel framework FrameThinker. Within this framework, LVLMs are able to iteratively interrogate video content. Developing such video reasoning capabilities in LVLMs presents notable challenges, particularly in adapting the model to new video actions (e.g. select frame), and designing reward functions to guide LVLMs to adopt the newly introduced action. To solve these challenges, we propose a two-phase training strategy, first employing Supervised Fine-Tuning (SFT) to instill fundamental action capabilities, followed by Reinforcement Learning (RL) to optimize a strategic decision-making policy. Notably, in this RL phase, we conduct an in-depth and comprehensive exploration of the reward design for each action and format reward. Extensive experiments on reasoning benchmarks like Video-Holmes, LongVideo-Reason, and long-video understanding benchmarks such as LongVideoBench, MLVU, VideoMME, and LVBench, demonstrate that FrameThinker achieves a significant average improvement of +10.4% over baselines while drastically reducing the number of processed frames. Most notably, our 7B model, FrameThinker establishes a new state-of-the-art on LongVideo-Reason, achieving 76.1% accuracy using an average of only 20.6 frames. This not only outperforms the competitive LongVILA-R1 (72.0%) but does so with over 20x fewer frames (vs. 512), demonstrating unparalleled efficiency and effectiveness.

Related papers

• Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning [20.07360876062324]
  Video-Thinker enables MLLMs to autonomously navigate grounding and captioning tasks for video reasoning.<n>Our Video-Thinker-7B substantially outperforms existing baselines such as Video-R1 and establishes state-of-the-art performance among 7B-sized MLLMs.
  arXiv  Detail & Related papers  (2025-10-27T16:10:45Z)
• Video-MTR: Reinforced Multi-Turn Reasoning for Long Video Understanding [33.58579390725519]
  Video-MTR is a reinforced multi-turn reasoning framework designed to enable iterative key video segment selection and question comprehension.<n>Unlike traditional video reasoning pipeline, which generate predictions in a single turn, Video-MTR performs reasoning in multiple turns.<n>To ensure intermediate reasoning process, we introduce a novel gated bi-level reward system.
  arXiv  Detail & Related papers  (2025-08-28T06:55:08Z)
• Episodic Memory Representation for Long-form Video Understanding [52.33907540905242]
  Large Video Language Models excel at general video understanding but struggle with long-form context window limits.<n>We introduce Video-EM, a training free framework inspired by the principles of human memory.<n>Video-EM achieves performance gains of 4-9 percent over respective baselines while utilizing fewer frames.
  arXiv  Detail & Related papers  (2025-08-13T04:33:07Z)
• Scaling RL to Long Videos [107.41198639507255]
  LongVILA-R1-7B achieves strong performance on video benchmarks, reaching 65.1% and 71.1% accuracy on VideoMME without and with subtitles, respectively.<n>LongVILA-R1-7B supports processing up to 8,192 video frames per video, and FPS settings.<n>We release our training system for public availability that supports RL training on various modalities.
  arXiv  Detail & Related papers  (2025-07-10T17:47:40Z)
• BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding [51.49345400300556]
  Large video-language models (VLMs) have demonstrated promising progress in various video understanding tasks.<n>Traditional approaches, such as uniform frame sampling, often inevitably allocate resources to irrelevant content.<n>We introduce BOLT, a method to BOost Large VLMs without additional Training through a comprehensive study of frame selection strategies.
  arXiv  Detail & Related papers  (2025-03-27T13:18:40Z)
• VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling [43.485687038460895]
  Long-context video modeling is critical for multimodal large language models (MLLMs)<n>This paper aims to address this issue from aspects of model architecture, training data, training strategy and evaluation benchmark.<n>We build a powerful video MLLM named VideoChat-Flash, which shows a leading performance on both mainstream long and short video benchmarks.
  arXiv  Detail & Related papers  (2024-12-31T18:01:23Z)
• Free Video-LLM: Prompt-guided Visual Perception for Efficient Training-free Video LLMs [56.040198387038025]
  We present a novel prompt-guided visual perception framework (abbreviated as Free Video-LLM) for efficient inference of training-free video LLMs. Our method effectively reduces the number of visual tokens while maintaining high performance across multiple video question-answering benchmarks.
  arXiv  Detail & Related papers  (2024-10-14T12:35:12Z)
• Koala: Key frame-conditioned long video-LLM [70.52369588364992]
  We propose a lightweight and self-supervised long video-LLM (Koala) to adapt pretrained vLLMs for generalizing to longer videos. Our approach outperforms state-of-the-art large models by 3 - 6% in absolute accuracy across all tasks. Surprisingly, we also empirically show that our approach not only helps a pretrained vLLM to understand long videos but also improves its accuracy on short-term action recognition.
  arXiv  Detail & Related papers  (2024-04-05T18:33:04Z)
• Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought [62.619076257298204]
  We motivate framing video reasoning as the sequential understanding of a small number of video reasonings. We introduce VIP, an inference-time challenge dataset designed to explore models' reasoning capabilities through video chain-of-thought. We benchmark GPT-4, GPT-3, and VICUNA on VIP, demonstrate the performance gap in complex video reasoning tasks, and encourage future work.
  arXiv  Detail & Related papers  (2023-05-23T10:26:42Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.