Paper Detail
Ke Li, Jiayu Chen, Maoliang Li, Zihao Zheng, Hailong Zou, Hengyi Zhang, Xuanzhe Liu, Xiang Chen
Efficient long-video understanding requires vision--language models (VLMs) to reason over a small number of frames selected as sparse visual evidence. Existing relevance-based methods rely on static one-shot selection with fixed frame budgets and candidate pools, while agent-based schedulers achieve adaptivity through costly multi-round reasoning and interactive search. We propose EcoFrame, a training-free framework for low-overhead query-adaptive visual evidence scheduling. EcoFrame leverages the VLM's inference feedback to determine when to increase the frame budget and where to search for additional candidate evidence. Specifically, entropy-gated budget scheduling uses output uncertainty to stop early when the current evidence is sufficient or progressively expand the frame budget otherwise. Meanwhile, attention-guided candidate proposal converts frame-level attention into a temporal prior, enabling dense local search in informative regions while preserving global coverage when attention is diffuse. Experiments on Video-MME, LongVideoBench, and MLVU demonstrate that EcoFrame achieves a better accuracy--efficiency trade-off across multiple VLM backbones. On Qwen2.5-VL, EcoFrame achieves an average accuracy of 64.4, surpassing BOLT at 63.5, while providing a $1.85\times$ speedup over AKS and BOLT. Compared with the agent-based A.I.R., EcoFrame maintains comparable accuracy with up to a $13.5\times$ inference speedup. Code will be available at https://github.com/AK-DREAM/EcoFrame.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@article{li2026when,
title = {When and Where to Look: Adaptive Visual Evidence Scheduling for Efficient Long Video Understanding},
author = {Ke Li and Jiayu Chen and Maoliang Li and Zihao Zheng and Hailong Zou and Hengyi Zhang and Xuanzhe Liu and Xiang Chen},
year = {2026},
abstract = {Efficient long-video understanding requires vision--language models (VLMs) to reason over a small number of frames selected as sparse visual evidence. Existing relevance-based methods rely on static one-shot selection with fixed frame budgets and candidate pools, while agent-based schedulers achieve adaptivity through costly multi-round reasoning and interactive search. We propose EcoFrame, a training-free framework for low-overhead query-adaptive visual evidence scheduling. EcoFrame leverages t},
url = {https://arxiv.org/abs/2608.03918},
keywords = {cs.CV, cs.AI},
eprint = {2608.03918},
archiveprefix = {arXiv},
}
{}