News
· 2026-08-25
Apple Machine Learning Research has introduced STARFlow2, a novel unified multimodal generation model that bridges autoregressive normalizing flows with Transformer architectures. This innovation addresses critical challenges in existing multimodal models, such as visual fidelity, structural asymmetry, and preservation of pretrained knowledge. STARFlow2 enables deep understanding and generation of interleaved text-image sequences, marking a significant advancement in AI's capacity for complex co
News
· 2026-08-25
The ArXiv team introduces Giraffe, a novel architecture designed to address the efficiency bottlenecks of Multimodal Large Language Models (MLLMs) in generating media content. Unlike traditional methods that use multiple specialized tokens to represent images—resulting in significantly increased input lengths—Giraffe maps hidden token representations to the embedding space of visual models using a single [IMG] token. This approach employs two shallow MLP blocks trained with six distinct loss fun
News
· 2026-08-25
The ArXiv team introduces VisCache, a novel framework designed to enhance the inference efficiency of Vision-Language Large Models (VLLMs). VisCache employs a two-stage process: it first filters temporal redundancies by selectively forwarding semantically informative keyframes, and then utilizes a surgical KV compression algorithm called PruneKV. This approach achieves up to 2.35x speedup in inference while significantly reducing memory consumption, all while maintaining competitive performance
News
· 2026-08-25
The ArXiv team introduces a novel dual-judge evaluation protocol to enhance the performance of vision-language models in legally grounded tasks. This protocol pairs a standard 0-10 quality judge with a strict binary semantic-equivalence judge, using the UK traffic sign interpretation task as a case study. The research reveals discrepancies in model performance under varying visibility and occlusion conditions, demonstrating that high-quality scores are least reliable when input quality degrades.
News
· 2026-08-25
This research explores the significance of language as a design material in architecture, particularly in the context of the growing adoption of text-to-X generative AI models. Starting from the works of Wittgenstein, Chomsky, Lakoff, and Schön's conceptual and generative metaphors, the study analyzes the interplay of three main forms of language—natural language, programming languages, and data annotations—in AI-assisted architectural design. The research proposes a research agenda that connect
News
· 2026-08-25
This research delves into the resource scarcity issue in Persian Natural Language Processing (NLP), arguing that Persian is more accurately described as 'annotation-scarce' rather than simply 'low-resource'. By analyzing 34 representative Persian text resources and conducting quantitative cross-checks, the study highlights the uneven task and domain coverage, incompatible annotation schemes, access and documentation friction, and limited supervision for specialist domains as key factors contribu
News
· 2026-08-25
The ArXiv team proposes MoTE (Mixture of Task Experts), a novel decoder architecture for multi-task video understanding. This architecture transforms the feed-forward networks of large language models into task-specific experts while keeping the multimodal backbone shared. Each sample follows a sample-level task route, ensuring that active task-expert computation remains independent of the number of stored task experts. Evaluated on five COIN benchmarks, the five-expert model activates approxima
News
· 2026-08-25
The ArXiv team introduces a novel product linking method that addresses the challenge of mapping merchant product records to canonical catalog products in e-commerce platforms. The approach employs a retrieve-then-match cascade system, where a lightweight text cross-encoder resolves the majority of high-confidence matches automatically, while an agentic vision-language model (VLM) handles ambiguous cases by inspecting product images and leveraging web search evidence. This method significantly i
News
· 2026-08-25
ArXiv researchers conducted a systematic study to identify the conditions under which auxiliary context can improve multi-modal time series forecasting. They found that context routing is effective only when the target data is not dominated by the last-value shortcut (low autocorrelation) and the context carries information about the target beyond history (non-zero conditional mutual information). Controlled experiments on the MoME model and other fusion mechanisms validated these conditions, an
News
· 2026-08-25
The ArXiv team introduces the Drift Variation Autoencoder, a novel approach that unifies conditional generation and generative representation learning. This method trains a masked encoder and a conditional flow decoder using a single clean-prediction flow matching loss. The research demonstrates its effectiveness in handling complex scenarios such as stochastic masking, cropping, or modality removal, generating multiple plausible completions. The model achieves precise conditional field generati
News
· 2026-08-25
The ArXiv team introduces ADeptS-Bench, a dual-stream trustworthiness benchmark for Computer Use Agents (CUAs). The benchmark, grounded in the ADEPTS capability framework and extensive user studies, features Safety and Disambiguation streams. The Safety stream evaluates agent robustness against threats embedded in visual interfaces, while the Disambiguation stream tests agents' ability to seek clarification when instructions are ambiguous. Experiments reveal that current models struggle to balan
News
· 2026-08-25
Luce is a novel 3D generation model that unifies 3D geometry and physically based rendering (PBR) materials within a voxelized multimodal Gaussian cloud. It employs a variational autoencoder to compress the representation into a unified material-aware latent space and uses a rectified-flow transformer to generate 3D assets from single images, supporting relighting, geometric accuracy, and material fidelity. On the Toys4K dataset, Luce achieves state-of-the-art performance in single-image-to-3D g
News
· 2026-08-25
Tencent has released WeMM-Embedding, a family of universal multimodal embedding models that support text, images, videos, and other modalities with flexible output dimensions. The family includes 2B, 4B, and 9B variants, trained in two stages: large-scale multimodal alignment and refinement using curated data. WeMM-Embedding achieves leading performance on multiple benchmarks, with the 2B variant surpassing the previous 8B open-source baseline on MMEB-v2, and the 9B variant achieving a new state
· 2026-08-25
On August 25, 2026, MiniMax released H3, an AI-powered tool capable of transforming text and images into AI-generated video clips. Designed to streamline the video creation process, H3 offers users an efficient and user-friendly solution for AI-driven video generation. This launch signifies a significant advancement in AI's role in multimedia content creation, providing content creators with a powerful new tool.
· 2026-08-25
ARC-AGI-3 has achieved an impressive score of 100 RHAE (Robust Human-Aligned Evaluation) on the ARC-AGI-Arcade public benchmark, demonstrating significant advancements in AI agent performance for complex task handling and human alignment. This milestone highlights ARC-AGI's ongoing innovation in AI agent technology and sets a new benchmark for performance optimization in diverse AI applications.
· 2026-08-25
OpenGeni has announced the open-sourcing of its Agentic Runtime, a specialized runtime environment designed for AI agents. This runtime aims to optimize the execution efficiency and collaborative capabilities of agents in complex tasks. Featuring a modular architecture and flexible task scheduling mechanisms, Agentic Runtime supports efficient operation of agents in multi-task and multi-modal environments. The release provides AI developers with a more powerful toolset, driving advancements in a
· 2026-08-25
Engram has released the Nemotron-3.5-Lightning-Omni-30B-A3B-GGUF agent model on Hugging Face, enhancing NVIDIA's original architecture with vision and hearing capabilities. This upgrade aims to provide AI agents with more human-like perception, enabling them to operate more effectively in multimodal environments and handle complex tasks with greater proficiency.
News
· 2026-08-24
The study proposes Intention Distillation (INDI), a novel method to enhance the behavior intent modeling of Vision-Language-Action (VLA) models. By distilling behavior-level intent into the action decoder during training, INDI significantly improves model performance in complex tasks. On SimplerEnv-Bridge and RoboCasa Kitchen benchmarks, INDI boosts success rates from 64.3% to 84.7% and from 64.1% to 70.3%, respectively. In real-world tasks, it increases the average success rate from 62.0% to 68
· 2026-08-24
The ArXiv team introduces Object-Uni, a unified model designed to address the limitations of existing visual understanding and generation models in understanding and manipulating the spatial states of object instances. By treating object pose as a shared geometric variable and employing a viewpoint-based orientation abstraction technique, Object-Uni achieves significant improvements in object-level pose understanding and pose-controllable generation. Experimental results demonstrate its superior
· 2026-08-24
The ArXiv team introduces Genda, a temporal generative Danmaku framework that simulates user interactions to generate pseudo Danmaku streams, and DM-FEND, a multimodal model that leverages interactions among video, audio, text, and Danmaku to detect fake news. Experimental results demonstrate that DM-FEND consistently outperforms state-of-the-art baselines across Chinese and English benchmarks, showcasing its potential in enhancing the robustness and accuracy of fake news detection in modern soc
· 2026-08-24
The ArXiv team introduces SPARE, a framework for optimizing reasoning pruning in multimodal agents. Using Kullback-Leibler (KL) divergence, SPARE identifies and removes redundant reasoning text while preserving critical visual evidence, mitigating the impact of 'textual debt' on model performance. Experiments demonstrate that SPARE achieves the highest average accuracy across multiple multimodal tool-use benchmarks while removing 37.89-64.58% of reasoning tokens, showcasing its potential to enha
· 2026-08-24
The ArXiv team has released a novel cross-modal deepfake detection framework named DF-MoE. By integrating multiple pre-trained models to extract audio and visual features and leveraging a Mixture-of-Experts (MoE) backbone to combine unimodal and multimodal cues, DF-MoE significantly enhances the accuracy and generalization capabilities of deepfake detection. Experimental results demonstrate that DF-MoE outperforms existing state-of-the-art methods across multiple benchmarks, showcasing its robus
· 2026-08-24
The ArXiv team introduces MediSkill-Evo, a clinical agent framework designed to address challenges in evidence-grounded, safe interactions under partial observability. By self-evolving clinical, procedural, symbolic, and visual knowledge in four typed knowledge banks and leveraging a process-constrained preference mechanism, MediSkill-Evo achieves significant improvements in diagnosis accuracy and treatment-intent coverage while substantially reducing critical failures. Experimental results demo
· 2026-08-24
ArXiv introduces BasketballBench, a multimodal benchmark designed to address the limitations of existing basketball understanding tasks that evaluate abilities in isolation. The benchmark comprises 7,980 questions across ten tasks in text, image, and video, built from the 2025-2026 NBA season data, including play-by-play, rosters, player profiles, and 2,501 possession-level broadcast clips. Experiments show that current MLLMs struggle with questions requiring the integration of multiple capabili
News
· 2026-08-24
Hugging Face has released EchoWM, an omnimodal world model designed for enterable generative media. EchoWM responds to continuous navigation while jointly generating 720p video, environmental sound, music, and speech. It organizes interactions around camera intent, specifying observer motion in first-person scenes and learning camera-character dynamics in third-person scenes without view-specific controllers. The model employs progressive training and autoregressive post-training for long-horizo
· 2026-08-24
MyBestLook has launched a new platform featuring AI Stylist and Virtual Try-On capabilities for hair, makeup, and outfit combinations. Users can experience real-time styling effects and receive personalized recommendations, enhancing the efficiency and accuracy of fashion and beauty decisions. This innovation provides a novel interactive experience for consumers and opens new digital transformation pathways for the fashion and beauty industries.
News
· 2026-08-24
Hugging Face is reportedly in acquisition talks that value the company at around $13 billion. However, due to the founders' strong sense of responsibility to the community, there are doubts about whether a sale will materialize. As a leading player in the open-source AI space, any significant change in Hugging Face's status could have profound implications for the AI community.
News
· 2026-08-24
arXiv has released a comprehensive survey on multimodal agentic frameworks, examining the role of multimodality in the core functional modules of agentic systems: perception, reasoning, planning, memory, and action. The study highlights the transition from text-centric agents to multimodal frameworks empowered by large multimodal models (LMMs), which can process and integrate diverse data modalities such as images, audio, and video. By analyzing delegated, late-fusion, and early-fusion architect
· 2026-08-22
RingLochid has released ImageSage, an AI-powered tool designed to help users manage and organize photos privately and locally. ImageSage leverages AI for intelligent classification, search, and photo organization without relying on cloud services, ensuring user data privacy and security. This tool is particularly suitable for users with high privacy requirements, offering an efficient and secure solution for photo management.
News
· 2026-08-21
Hugging Face has introduced EviRank, a novel approach to multimodal image re-ranking. EviRank parses queries into a unified evidence package across six semantic slots (e.g., entities, attributes, relations) and employs evidence-conditioned verification combining deterministic rubric scoring and evidence-grounded listwise comparison. This method achieves state-of-the-art performance across benchmarks for text-to-image, image-to-image, and compositional image retrieval. The distilled lightweight s