Latest stories
LLaDA-UI: Block-Wise Diffusion Brings Parallel Decoding to GUI Agents
A 16.7B-parameter mixture-of-experts diffusion vision-language agent achieves strong multimodal GUI performance while preserving b...
KaiNinja Extends Native 3D Generators to Part-Level Outputs
A new method called KaiNinja upgrades the TRELLIS.2 3D generator to produce part-level assets directly, using a dual-volume repres...
Omni-Streaming Thinking Defers Claims Until Cross-Modal Verification
A new paper proposes Omni-Streaming Thinking (OST), a framework that keeps streaming audio-visual claims pending until evidence fr...
AlayaVista Splits Panoramic World Modeling from Perspective Video
A new streaming world model builds a 360-degree scene prior from a single image, evolves it as a panoramic latent state, and rende...
BVB Benchmark Tests Whether AI Agents Truly Understand Video by Rebuilding It in Blender
A new benchmark, Blender-VideoBench (BVB), asks multimodal agents to reconstruct real-world videos as animated Blender scenes, rev...
Study Questions 'Lossless' Claim of Orthrus Speculative Decoding Under BF16
An independent reproduction of the Orthrus speculative-decoding architecture finds exact trajectory matching in only about 45% of...
Hugging Face Paper Unveils LynnReal-Omni, a 32B Multimodal Video Diffusion Framework for Agentic Visual Workflows
LynnReal-Omni is a unified multimodal video diffusion framework that combines agentic visual controls with high-fidelity generatio...
Hugging Face Paper Proposes Grouped Value Attention to Shrink KV Caches
A new paper introduces Grouped Value Attention (GVA), which stores grouped values and reconstructs content keys with a learned lin...
RSIAgent Lets Frozen Models Self-Improve by Scaling Experience, Not Parameters
A new training-free multi-agent framework from Aether Labs builds reusable environment memory through broad-then-deep exploration,...
Hugging Face Paper Proposes 'Discovery Foundation Models' for Open-Ended Scientific Discovery
A new paper argues that the next frontier for foundation models is Discovery Intelligence: participating in how new problems, repr...
PhysBrain 1.5 Turns Vision-Language Models Into Unified Physical Foundation Models
Hugging Face paper introduces PhysBrain 1.5, an 8B model that jointly learns physical understanding, action generation, and future...
ZGCM-1: A Fully Open 7B Model That Thinks, Searches, and Helps Build Itself
ZGCM-1 is a fully open 7B dense foundation model trained from scratch for extreme data, system, and algorithmic efficiency, pairin...