Gaze Attention is a new mechanism for multimodal large language models that selectively focuses on relevant visual regions during each generation step, rather than attending to all visual tokens. By grouping tokens into spatial regions and using learnable context tokens to retain global information, it reduces attention computation and visual key‑value entries by up to 90%. Experiments on 13 image and 6 video benchmarks show that Gaze Attention matches or outperforms dense‑attention baselines while using fewer visual resources.
By Junha Song, Byeongho Heo, Geonmo Gu, Jaegul Choo, Dongyoon Han, Sangdoo Yun
FocusGraph is a framework for selecting keyframes in egocentric long‑video question answering. It uses a lightweight Scene‑Graph LLM Selector to identify query‑relevant clips from graph‑based captions, then extracts keyframes with Patch‑wise Sparse‑Flow Retention (PSFR) before feeding them to a multimodal large language model for answer generation. The method achieves state‑of‑the‑art performance on FindingDory and HourVideo while reducing question‑time inference cost.
By Tatiana Zemskova, Solomon Andryushenko, Ilya Obrubov, Viktoriia Khoruzhaia, Ekaterina Eroshenko, Ekaterina Derevyanka, Dmitry Yudin
arXiv:2605. 20950v2 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) face a bottleneck of prohibitive computational costs arising from massive visual token sequences during inference.
By Yulin Zhao, Zheng Zhang
MemoryCard is a video-memory-based augmentation framework designed to improve long-video question answering for Vision‑Language Models. It segments lengthy videos into semantically coherent units—each representing a distinct topic or event—by performing a self‑reading process over the video and aligned utterances. For each unit, the framework generates an event‑level video gist and selects representative visual moments, which are compiled into unified Memory Cards that are used for retrieval and answering questions, yielding up to a 21.8% relative accuracy improvement under comparable visual‑token budgets.
By Qing Yang, Pengcheng Huang, Xinze Li, Zhenghao Liu, Yukun Yan, Yu Gu, Ge Yu, Gang Li, Maosong Sun
BuddyVQA is a new benchmark for companion‑style question answering on egocentric streaming video, comprising 21.6K questions tied to 6K highlight moments across 1,012 long first‑person videos. It emphasizes two often overlooked aspects of daily first‑person QA: ego‑deictic expressions and interactively chained questions, requiring models to resolve visual pronouns and infer user intent within a long‑form streaming context. The authors propose MyBuddy, a multimodal chain‑of‑thought QA assistant that uses a question filter and multi‑level memory to efficiently retrieve visual and QA information, achieving significant performance gains on BuddyVQA and generalizing to other streaming and common video QA benchmarks.
By Hangyu Qin, Junbin Xiao, Shenglang Zhang, Angela Yao
arXiv:2608. 15614v1 Announce Type: cross Abstract: The use of multimodal LLMs (MLLMs) for egocentric video understanding with wearable devices is constrained by the token budget.
By Matteo Stoiber, Niels Buus Lassen
The paper introduces a training‑free visual token pruning strategy for vision‑language models that separates early vision‑guided pruning from later text‑guided reselection. By first pruning tokens with vision‑encoder attention, retaining candidates until the decoder midpoint, and then applying text‑to‑visual attention, the method preserves task‑relevant visual information. Across eight benchmarks and three models, it achieves an average performance recovery of 11.10 and 16.84 percentage points at 80% and 90% pruning, respectively, while maintaining comparable or lower LLM‑prefill latency.
GazeFlow is a new framework for egocentric gaze prediction that models gaze as a joint distribution of temporal positions conditioned on both top‑down task cues and bottom‑up visual saliency. It employs conditional flow matching to iteratively transform Gaussian noise into realistic gaze trajectories, using a velocity field informed by video‑encoded visual features and global task queries. On standard benchmarks, GazeFlow outperforms existing methods on per‑frame metrics and produces trajectories that better reflect human gaze dynamics.
By Sheng Zhao, Weikai Lin, Yuhao Zhu
arXiv:2604. 08342v2 Announce Type: replace Abstract: Long context egocentric video understanding has recently attracted significant research attention, with augmented reality (AR) highlighted as one of its most important application domains.
By Qiance Tang, Ziqi Wang, Jieyu Lin, Ziyun Li, Barbara De Salvo, Sai Qian Zhang
arXiv:2607. 08489v1 Announce Type: cross Abstract: Vision-language models excel at video captioning, yet typically generate descriptions that fail to capture individual viewers' attention.
By Shenghui Chen, Po-han Li, Ximeng Sun, Shijia Yang, Emad Barsoum, Zicheng Liu, Sandeep Chinchali, Ufuk Topcu
Estimating human gaze targets from images in-the-wild is an important and formidable task. Existing approaches primarily employ brittle, multi-stage pipelines that require explicit inputs, like head bounding boxes and human pose, in order to identify the subject of gaze analysis.
VISTA (Visually Impaired Scene and Task Annotations) is a multimodal egocentric benchmark designed to evaluate goal‑oriented assistance for blind and low‑vision users. It comprises about 1,000 recordings from Project Aria glasses, each providing six synchronized sensor streams and three types of annotations across ten task categories, from object localization to hazard detection. The benchmark tests vision‑language models in zero‑shot settings and introduces a lightweight multimodal baseline that integrates all six modalities, revealing that current models can describe scenes but struggle to generate effective assistive guidance.
By No\'e Mosca, Mohamed Rayan Barhdadi, Muhammad Hamza, Joshua Yeh, Suryodaya Bikram Shahi, Zhiyan Li, Selim-Antoine Lali, Weiyue Li, Hasan Kurban, Erchin Serpedin, Gang Luo, Yue Zhang, Minghan Li, Mengyu Wang