Latest stories
USDCraft Uses LLMs to Turn Real Objects Into Simulation-Ready Articulated 3D Assets
A new Hugging Face paper introduces USDCraft, a framework where a pretrained LLM writes and revises executable programs to build a...
SparseEngine: A Sparse-First Inference Engine for Long-Context LLM Agents
Hugging Face paper introduces SparseEngine, a sparse-first inference engine that supports 15 sparse methods and delivers over 10x...
Hugging Face Paper Proposes Evidence-Grounded Oversight for Long-Horizon AI Agents
A new paper introduces AgentMonBench and the Evidence-Grounded Behavior Graph (EBG), a training-free method that helps oversight m...
Hugging Face Paper: OneSearch-VL Unifies Image and Video Deep Research
A new research paper introduces OneSearch-VL, a unified multimodal agent built around a Visually Grounded Evidence Graph, reportin...
SparseDecoding Targets the Real Bottleneck in LLM Inference: The Decode Stage
A new Hugging Face paper introduces SparseDecoding, a decoding-aware pruning framework that calibrates pruning on self-generated t...
Hugging Face Paper Combines Preference and Rubric Rewards to Post-Train Text-to-Image Models
A new post-training recipe composes human-preference rewards with rubric-based rewards to improve frontier text-to-image models, b...
Memento 3 Lets Frozen LLMs Self-Improve via Reflective Rulebooks
Hugging Face paper introduces Memento 3, a model-based approach to recursive self-improvement in which a frozen LLM agent builds a...
TestPrism Challenges Single-Reference Test Evaluation for LLM Coding Agents
A new benchmark, TestPrism, shows that judging AI-generated tests against a single reference solution inflates quality scores, wit...
MC-Sparse Closes the Dense-Sparse Attention Gap in Diffusion Transformers
A new training-free sparse-attention framework from Hugging Face researchers traces quality loss in diffusion transformers to thre...
Asana Slashes Browser Agent Costs 76x with OpenAI's GPT-6.1 Sol
Asana used GPT-6 Astra in Codex to make its browser agent 76 times cheaper and 5 times faster in tests, aiming to offer customers...
Sophos Cuts Threat Investigation Time by 96% Using OpenAI's Daybreak
Sophos has reduced cyber-threat investigation time by 96% and automated 52% of its MDR cases using OpenAI's Daybreak, while keepin...
VibeEdit Lets You Edit Images by Drawing on Them
Hugging Face researchers introduce VibeEdit, a canvas-instruction interface that lets users mark regions and add short notes direc...