Evidence-Driven Dynamic Visual Selector for Efficient Long Video Understanding

• URL: http://arxiv.org/abs/2608.05780v1
• Date: Thu, 06 Aug 2026 09:15:43 GMT
• Title: Evidence-Driven Dynamic Visual Selector for Efficient Long Video Understanding
• Abstract summary: We propose a dynamic visual selection framework named EviSelect grounded in the target MLLM's internal attention evidence.<n>Our method efficiently approximates maps of the target MLLM using highly compressed visual inputs and sparse attention.<n>EviSelect achieves superior performance compared to existing methods while reducing selected visual tokens by about 50% and achieving a 3.9x end-to-end speedup.
• Score: 54.840531495584095
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Recent advancements in MLLM-based long-form video understanding have mitigated inference-time computational cost and limited context lengths by selecting query-relevant frames. However, existing approaches predominantly rely on external proxy scorers and rigid heuristic rules, inevitably suffering from misalignment with the target MLLM's intrinsic evidence and failing to accommodate the non-uniform spatiotemporal information density. In this paper, we propose a fine-grained dynamic visual selection framework named EviSelect, grounded in the target MLLM internal attention evidence. Our method efficiently probes visual evidence via sparse prefilling as a structured prior to guide distribution-aware dynamic sampling. Specifically, we efficiently approximate attention maps of the target MLLM using highly compressed visual inputs and sparse attention, well-aligned to the full counterpart. Conditioned on three complementary attention components derived from this prior, we design a lightweight selector that not only precisely locates query-relevant timestamps but also adaptively adjusts the local sampling rate and spatial resolution. To enable evidence-conditioned spatiotemporal sampling, we formulate the selector as a stochastic policy and optimize it via GRPO under a joint accuracy--efficiency reward. By rewarding correct predictions under lower visual cost through group-relative comparisons, our method encourages the policy to allocate computation dynamically according to the information density of each video. Across three long video understanding benchmarks, EviSelect achieves superior performance compared to existing methods while reducing selected visual tokens by about 50\% and achieving a 3.9x end-to-end speedup.

Related papers

• Efficient Frame Selection for Long Videos at Test Time with Attention-Based MLLM Selectors [69.75895255395702]
  Cross-modal attention at validation-selected extraction layers in MLLMs already provides query-relevant frame evidence without requiring autoregressive generation.<n>A lightweight MLLM selector, even with only 2B parameters, can extract frame-level evidence by converting selected-layer attention into relevance scores through query-conditioned aggregation.<n>Under a 32-frame budget, our selector improves over uniform sampling by up to 6.4 points on Video-MME and outperforms prior training-based selectors under matched frame budgets.
  arXiv  Detail & Related papers  (2026-07-17T07:04:31Z)
• Where to Focus: Query-Modulated Multimodal Keyframe Selection for Long Video Understanding [48.14432643308697]
  Long video understanding is a formidable challenge for Multimodal Large Language Models.<n>Q-Gate is a plug-and-play framework that treats selection as a dynamic modality routing problem.
  arXiv  Detail & Related papers  (2026-04-19T13:04:18Z)
• Query-Conditioned Evidential Keyframe Sampling for MLLM-Based Long-Form Video Understanding [21.306038832082553]
  Multimodal Large Language Models (MLLMs) have shown strong performance on video question answering.<n>Their application to long-form videos is constrained by limited context length and computational cost.<n>We propose an evidence-driven sampling framework grounded in information theory.
  arXiv  Detail & Related papers  (2026-04-01T15:02:50Z)
• VideoBrain: Learning Adaptive Frame Sampling for Long Video Understanding [9.415923244280542]
  VideoBrain is an end-to-end framework that enables Vision-Language Models to adaptively acquire visual information through learned sampling policies.<n>Our approach features dual complementary agents: a CLIP-based agent for semantic retrieval across the video and a Uniform agent for dense temporal sampling within intervals.
  arXiv  Detail & Related papers  (2026-02-04T00:08:35Z)
• Tracking Meets Large Multimodal Models for Driving Scenario Understanding [76.71815464110153]
  Large Multimodal Models (LMMs) have recently gained prominence in autonomous driving research.<n>We propose to integrate tracking information as an additional input to recover 3D spatial and temporal details.<n>We introduce a novel approach for embedding this tracking information into LMMs to enhance their understanding of driving scenarios.
  arXiv  Detail & Related papers  (2025-03-18T17:59:12Z)
• Coarse Correspondences Boost Spatial-Temporal Reasoning in Multimodal Language Model [51.83436609094658]
  We introduce Coarse Correspondences, a simple lightweight method that enhances MLLMs' spatial-temporal reasoning with 2D images as input. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks.
  arXiv  Detail & Related papers  (2024-08-01T17:57:12Z)
• Practical Video Object Detection via Feature Selection and Aggregation [18.15061460125668]
  Video object detection (VOD) needs to concern the high across-frame variation in object appearance, and the diverse deterioration in some frames. Most of contemporary aggregation methods are tailored for two-stage detectors, suffering from high computational costs. This study invents a very simple yet potent strategy of feature selection and aggregation, gaining significant accuracy at marginal computational expense.
  arXiv  Detail & Related papers  (2024-07-29T02:12:11Z)
• You Can Ground Earlier than See: An Effective and Efficient Pipeline for Temporal Sentence Grounding in Compressed Videos [56.676761067861236]
  Given an untrimmed video, temporal sentence grounding aims to locate a target moment semantically according to a sentence query. Previous respectable works have made decent success, but they only focus on high-level visual features extracted from decoded frames. We propose a new setting, compressed-domain TSG, which directly utilizes compressed videos rather than fully-decompressed frames as the visual input.
  arXiv  Detail & Related papers  (2023-03-14T12:53:27Z)
• DFA: Dynamic Feature Aggregation for Efficient Video Object Detection [15.897168900583774]
  We propose a vanilla dynamic aggregation module that adaptively selects the frames for feature enhancement. We extend the vanilla dynamic aggregation module to a more effective and reconfigurable deformable version. On the ImageNet VID benchmark, integrated with our proposed methods, FGFA and SELSA can improve the inference speed by 31% and 76% respectively.
  arXiv  Detail & Related papers  (2022-10-02T17:54:15Z)
• Video Annotation for Visual Tracking via Selection and Refinement [74.08109740917122]
  We present a new framework to facilitate bounding box annotations for video sequences. A temporal assessment network is proposed which is able to capture the temporal coherence of target locations. A visual-geometry refinement network is also designed to further enhance the selected tracking results.
  arXiv  Detail & Related papers  (2021-08-09T05:56:47Z)

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.