The article introduces Multi-Stream Scene Script (MTSS), a structured JSON‑based video description paradigm that replaces monolithic captions, explains its design principles, compares its advantages, and presents experimental evidence showing significant gains in both video understanding and generation tasks.
The Video‑MME‑v2 benchmark reveals that current multimodal video models, despite high leaderboard scores, struggle with genuine video understanding, thanks to a rigorous three‑layer evaluation, non‑linear scoring, and a meticulously curated 800‑video dataset that exposes their true intelligence limits.
The new Video‑MME‑v2 benchmark reveals that despite saturated high scores on existing video‑understanding tests, the strongest commercial model (Gemini‑3‑Pro) reaches only 49.4 points versus a human expert’s 90.7, highlighting the benchmark’s layered ability system, group‑level non‑linear scoring, and the nuanced impact of "Thinking" features.
MedGRPO, a multimodal large model, achieves a breakthrough in medical video understanding by introducing clinical semantic parsing that aligns visual cues with structured medical knowledge, boosting performance and raising ethical questions about AI’s evolving role from a supportive assistant to a collaborative clinical partner.
This article details how the Qwen‑Video‑8B model, built on Qwen3‑VL‑8B‑Instruct, is fine‑tuned with the LLaMA‑Factory framework using a curated city‑scenery dataset, addresses challenges of domain knowledge, temporal modeling and multimodal fusion, and demonstrates improved video captioning across baseline, English‑fine‑tuned and Chinese‑fine‑tuned versions.
CrossVid is an open‑source benchmark that evaluates multimodal large language models on cross‑video reasoning, offering 5,331 videos and 9,015 high‑quality QA pairs across four reasoning dimensions, and revealing that even the strongest models achieve only about 50% accuracy compared with human performance.
CrossVid is an open‑source benchmark that evaluates multimodal large language models on cross‑video reasoning tasks, providing 5,331 videos, 9,015 QA pairs, four high‑level dimensions and ten specific tasks, and exposing significant performance gaps between current models and humans.
Kwai has open‑sourced its new flagship multimodal model Keye‑VL‑671B‑A37B, which upgrades visual perception, cross‑modal alignment and complex reasoning, achieving top scores on image, video, and mathematical reasoning benchmarks while detailing its architecture, three‑stage pre‑training, post‑training strategies, and future multimodal agent plans.
Introducing Ming‑Flash‑Omni‑Preview, a 103‑billion‑parameter open‑source multimodal model built on a sparse MoE architecture that delivers state‑of‑the‑art performance in controllable image generation, streaming video understanding, and context‑aware speech recognition, surpassing prior models on GenEval and GEdit benchmarks.
Keye‑VL‑1.5, an 8‑billion‑parameter multimodal large language model, introduces a Slow‑Fast video encoding strategy, a four‑stage progressive pre‑training pipeline with 128K context, and a sophisticated post‑training regime that together achieve state‑of‑the‑art performance on video and vision‑language benchmarks while maintaining strong general capabilities.
Fast‑search platform Kwai has open‑sourced the 8‑billion‑parameter multimodal LLM Keye‑VL‑1.5, which introduces a slow‑fast frame encoding, a progressive four‑stage pre‑training pipeline, and an automated data construction workflow, achieving state‑of‑the‑art results on video and vision‑language benchmarks and surpassing many closed‑source models.
Kuaishou's foundational large‑model team secured seven papers at the prestigious ACL 2025 conference, covering alignment bias during model training, safety in inference, decoding strategies, fine‑grained video‑temporal understanding, and new evaluation benchmarks that push the frontier of multimodal large language models.
Kuaishou’s foundational large-model team has secured seven papers at ACL 2025, spanning alignment bias in training, safety defenses during inference, decoding strategies, fine-grained video-temporal understanding, reward fairness in RLHF, multimodal captioning benchmarks, and methods to curb hallucinations in vision-language models.
The paper introduces UNIFIEDREWARD, the first unified reward model for multimodal understanding and generation that supports pairwise ranking and pointwise scoring, builds a 236K human‑preference dataset across image and video tasks, and uses DPO to align VLMs and diffusion models, achieving significant performance gains on both image and video benchmarks.
This article presents the research behind Synchronized Video Storytelling, introducing the E‑SyncVidStory dataset, the VideoNarrator multimodal architecture, and extensive experiments that demonstrate high‑quality, product‑aware video narration generation for e‑commerce applications.
This roundup highlights the latest AI breakthroughs, including Zhipu AI's video‑understanding model for temporal Q&A, Tencent's video‑to‑audio generation system, Vimeo's AI‑content labeling policy, Apple’s Core ML inclusion of ByteDance’s depth model, AMD’s acquisition of Silo AI, Claude’s new editing features, Quark’s all‑in‑one search AI, TikTok’s VR live streaming on Vision Pro, the launch of the "Xinliu" AI search assistant, and Canva’s restrictions on political AI‑generated posters.
The paper presents Action Sensitivity Learning (ASL), a framework that models frame‑wise importance at both class‑level (via learnable Gaussian distributions) and instance‑level (using quality scores), integrates these weights into classification and regression losses, adds a contrastive InfoNCE term, and achieves state‑of‑the‑art temporal action localization performance across six benchmark datasets.
This article presents a comprehensive multi-modal video understanding system for AIGC video generation, detailing technical architecture, GCN-based semi-supervised learning, and practical applications across automotive content scenarios.
This talk presents recent research on integrating dynamic analysis and graph machine learning to generate social relationship graphs from video, detailing hierarchical graph convolution networks, multimodal feature fusion, weakly supervised training, experimental results, and applications such as enhanced video retrieval and storyline understanding.
The paper introduces Action Sequence Verification (ASV), a task that determines whether two videos follow the same ordered actions, provides the Chemical Sequence Verification dataset and re‑annotated COIN‑SV and Diving48‑SV sets, and proposes the CosAlignment Transformer (CAT) with intra‑step feature extraction, a Transformer‑based inter‑step encoder, and a sequence‑alignment loss that outperforms prior baselines and serves as a pre‑training model for video retrieval and classification.
This article presents a multimodal approach that combines dynamic analysis and graph machine learning to generate and apply social relationship graphs in videos, detailing problem background, graph generation modules, applications such as video retrieval, experimental results, and future research directions.
The article introduces pet identity verification using target re‑identification and occluded video instance segmentation, describes recent ICCV VIPriors competitions where Ant Group’s insurance team achieved top ranks, and explains how these computer‑vision techniques are applied to insurance claims, pet identification, and future AI scenarios.
The article shares a Tencent competition champion’s practical TensorFlow‑based video ad solution, detailing data handling, model architecture, optimization tricks, multimodal fusion techniques, and experimental observations to help participants improve performance in the 2021 Tencent Advertising Algorithm Contest.
Youku’s new text‑video alignment system automatically generates short video summaries by extracting multimodal video and linguistic features, matching sentences to clips through embedding and tag‑level models, and enabling AI‑driven auto‑editing that cuts production time from days to minutes.
The Boundary Content Graph Neural Network (BC‑GNN) introduces a bipartite‑graph framework that jointly refines start/end boundary probabilities and segment‑content confidence, enabling more precise temporal action proposals and achieving state‑of‑the‑art results on ActivityNet‑1.3 and THUMOS14.
The paper introduces Boundary Content Graph Neural Network (BC‑GNN), a graph‑based approach that jointly models boundary and content predictions to generate more accurate temporal action proposals and reliable confidence scores, achieving state‑of‑the‑art results on ActivityNet‑1.3 and THUMOS‑14.
This paper introduces a method for video multi-label classification that incorporates label correlation features using graph convolutional networks, significantly improving classification performance.
This article explores how AI is used in iQIYI's video advertising pipeline to analyze video content, generate and recommend ad placement points, create scene‑aware ad creatives, build a video knowledge graph, and support various ad formats, ultimately improving ad relevance and revenue.
This article summarizes Hulu Chief Research Officer Xie Xiaohui’s presentation on why video content understanding is essential, the technical challenges involved, and Hulu’s end‑to‑end solutions—including fine‑grained segmentation, logo and subtitle detection, automated pipelines, tagging taxonomy, content generation, and vector embeddings—to improve recommendation, advertising, and search for massive video libraries.
This article examines the evolution of search, defines multimedia content retrieval, explores user scenarios such as voice, image, and video input, and details key AI techniques—including OCR, face recognition, and content knowledge graphs—that enable semantic understanding and ranking of video content.
