The paper introduces STITCH, a training‑free method that partitions videos into semantically meaningful temporal chunks using a frozen video‑text backbone. By detecting changes in the embedding sequence of short video windows, STITCH produces reusable temporal abstractions that can be applied to multiple tasks such as event boundary detection, language‑based moment retrieval, and frame selection for vision‑language models. Experiments show that STITCH performs competitively with specialized methods while requiring no task‑specific training, especially when processing is limited to a few frames or tokens.
By Etienne Casanova, Sevan Brodjian, Pietro Perona
The paper introduces structured video prompting, a training‑free inference‑time technique that augments input videos with lightweight spatial and temporal structure to provide explicit anchors for evidence organization. By applying this method to two video benchmarks and two open video‑language models, the authors demonstrate performance improvements across several tasks, with gains varying by model and task. The study suggests that failures in video‑language models stem not only from reasoning capacity but also from how video evidence is presented during inference.
By Sadegh Mohammadian
Kairos is a new video dataset designed for fine-grained video-language modeling, featuring long-duration videos from ten minutes to half an hour. Each video is annotated with time-resolved labels that capture ongoing actions, entity appearances, attributes, interactions, and evolving contextual cues throughout the timeline. The dataset supports fine-grained evaluation, long-range modeling, reasoning, instruction data construction, representation learning, and video generation.
By Ruibo Ming, Lei Sun, Deheng Zhang, He Zhang, Jialu Li, Jian Wang, Zhendong Li, Mengshun Hu, Danda Pani Paudel, Luc Van Gool, Jinjin Gu
STRAND is a new benchmark that tests multimodal large language models’ ability to track objects, their states, and relationships over time in videos. It evaluates intermediate reasoning by breaking queries into sub‑questions and uses Faithful Accuracy to ensure all parts of an answer are correct. The authors also propose an object‑centric framework that builds structured trajectories and shows reduced hallucinations and better temporal consistency compared to existing models.
By Thong Nguyen, Tri Cao, Khoi Le, Cong-Duy Nguyen, Quynh Vo, See-Kiong Ng, Bryan Hooi Kuen-Yew
TimeBlind is a diagnostic benchmark designed to evaluate fine‑grained spatio‑temporal compositionality in video large language models (LLMs). It categorizes temporal understanding into three levels—atomic event recognition, event property characterization, and reasoning about event interdependencies—and uses a minimal‑pairs paradigm where video pairs share identical static content but differ only in temporal structure. Across 20 state‑of‑the‑art MLLMs tested on 600 curated instances, the best model achieved only 48.2% instance accuracy, far below human performance of 98.2%, highlighting a reliance on static visual shortcuts rather than true temporal reasoning.
By Baiqi Li, Kangyi Zhao, Ce Zhang, Chancharik Mitra, Jean de Dieu Nyandwi, Gedas Bertasius
The paper introduces Seeing Before Synthesizing (SBS), a weakly-supervised dense video captioning framework that uses a vision‑language model to generate frame‑level narratives for gaps between events and detect transitions based on semantic changes. SBS refines temporal masks by aligning transition points with vision‑language cues, rather than relying on rigidly placed synthetic captions. Experiments on ActivityNet Captions and YouCook2 show that SBS achieves state‑of‑the‑art results in both captioning and localization tasks.
By Ye-Chan Kim, Seunghee Choi, SeungJu Cha, Si-Woo Kim, Hwiseon Kim, Hyungee Kim, Dong-Jin Kim
VidOmni-Bench is a new benchmark for fine‑grained video understanding that asks models to verify whether each event in dense video captions is supported by the video. It contains 500 videos covering five complexity types and durations from 4 seconds to 90 minutes, and uses human‑verified sentence‑level labels to create hard negatives. Experiments show that Video‑LLMs often hallucinate events, struggle to detect incorrect descriptions, and exhibit varying weaknesses depending on video complexity and duration.
By Changbeen Kim, Junwon Chang, Kipyo Kim, Risa Shinoda, Kuniaki Saito, Donghyun Kim
VideoTIR introduces a reinforcement‑learning approach to improve long‑video understanding by encouraging multimodal large language models to use comprehensive multi‑level toolkits efficiently. It combines Zero‑RL and SFT cold‑starting strategies to help models retrieve and focus on meaningful video segments, images, and regions, thereby reducing hallucinations. The method includes Toolkit Action Grouped Policy Optimization (TAGPO) to streamline tool‑calling and a sandbox‑based trajectory synthesis framework for high‑quality data, achieving strong results on three long‑video QA benchmarks.
By Zhe Gao, Shiyu Shen, Taifeng Chai, Weinong Wang, Haotian Xu, Xing Wu, Wenbin Li, Qi Fan, Yang Gao, Dacheng Tao
arXiv:2607. 04872v1 Announce Type: cross Abstract: Reasoning temporal localization (RTL) requires a model to generate an answer that itself contains the time interval supporting it, so high-level reasoning and precise temporal grounding must be produced jointly in a single response.
By Youngkil Song, Yoonjae Baek, Dongwon Kim, Inho Kim, Dongkeun Kim, Suha Kwak
EventVL introduces the first generative event-based multimodal large language model (MLLM) designed for explicit semantic understanding of event streams. The framework leverages a newly annotated dataset of nearly 1.4 million event–image/video–text pairs and incorporates an Event Spatiotemporal Representation to capture comprehensive event information, along with Dynamic Semantic Alignment to refine sparse semantic spaces. Experiments demonstrate that EventVL outperforms existing MLLM baselines in event captioning and scene description generation tasks, advancing the field of event vision.
By Pengteng Li, Yunfan Lu, Pinghao Song, Wuyang Li, Huizai Yao, Hui Xiong
TempCloze is a video cloze benchmark designed to evaluate visual temporal reasoning in Video-LLMs. The task presents a video’s beginning and ending clips and asks models to select the correct missing middle from four candidates, focusing on semantic, alignment, and progression aspects while minimizing appearance cues. Evaluation of 31 models shows that temporal alignment is the main challenge, with models performing better on semantic content and event progression but struggling to place events correctly in time.
By Wenqi Pei, Henry Hengyuan Zhao, Yilai Liu, Jiahao Meng, Han Chen, Ziyu Wang, Hongyang Du
arXiv:2606. 01591v1 Announce Type: cross Abstract: The TimeLogic Challenge evaluates formal temporal-logic reasoning over video - 16 operators (before, after, until, since, always, co-occur, ordering, ...
By Ali Alavi