Leading AI knowledge platform
Tencent WeChat AI introduces WeDLM, a diffusion language model that works with standard causal attention and KV caching, achieving up to ten‑fold speedups over autoregressive models while maintaining or improving generation quality across math reasoning and open‑ended tasks.
The article introduces EOC‑Bench, a pioneering benchmark that evaluates multimodal large language models on dynamic first‑person visual tasks across past, present, and future time dimensions, presents its 3,277 questions, novel multi‑scale temporal accuracy metric, extensive model comparisons, and detailed error analysis revealing current models’ limitations in temporal perception and memory.
OmniVGGT introduces a flexible omni‑modality driven transformer that can ingest arbitrary numbers of geometric cues such as depth maps and camera parameters, achieving state‑of‑the‑art performance on diverse 3D tasks while keeping inference speed comparable to its RGB‑only predecessor.
UnityVideo, a new vision framework from HKUST, CUHK, Tsinghua and Kuaishou, unifies training across depth, flow, pose, segmentation and RGB modalities, achieving faster convergence, higher video quality, zero‑shot generalization and stronger physical reasoning compared with existing single‑modality video generators.
A surge of international attention focuses on China's rapid progress in embodied AI, highlighted by the GDPS 2025 competition, mass‑produced robots, and industry‑wide ecosystem support, prompting foreign observers to question why Chinese robotics outpace their own prototypes and demos.
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.
This article analyzes why optimizing sequence‑level rewards for LLMs with token‑level surrogate objectives can improve reinforcement‑learning stability, explains the theoretical conditions required, introduces Routing Replay for MoE models, and presents extensive experiments validating the approach.
Researchers from Kuaishou and Hong Kong City University introduce VANS, a novel Video-as-Answer system that predicts and visualizes the next event in a video by jointly optimizing a visual language model and a video diffusion model, enabling personalized step‑by‑step guidance and future scenario generation.
Meta’s newly released SAM 3 and SAM 3D models enable single‑image 3D reconstruction and promptable segmentation, outperforming prior methods on benchmarks, introducing a shared perception encoder, a Presence Head to reduce hallucinations, and a two‑stage generation pipeline that produces high‑fidelity geometry and texture.
RoMa v2 introduces a two‑stage matching‑then‑refinement pipeline powered by DINOv3 features, custom CUDA kernels, and diverse training data, delivering state‑of‑the‑art accuracy, speed, and pixel‑level uncertainty estimation across a wide range of dense matching benchmarks.
