Research Papers
Hugging Face Paper Proposes Four-Level Framework for Agentic Visual Generation
A new paper on Hugging Face argues that agenticity in visual generation should be defined by what a controller can directly decide...
Kalman Delta Networks Bring Uncertainty Tracking to Linear Attention Memory
A new Hugging Face paper recasts recurrent associative memory as a linear-Gaussian state-space model, letting Kalman-filter update...
NVIDIA-Led System Brings Online Draft Co-Training to Speculative Decoding at 122B Scale
A new end-to-end system enables large-scale online draft co-training for speculative decoding during long-context RL post-training...
VDiff-Bench Exposes Blind Spots in Multimodal AI's Ability to Spot Subtle Image Changes
A new benchmark of 1,756 multiple-choice questions reveals that leading multimodal language models struggle badly with fine-graine...
CVRR Forces Visual Reasoning to Happen in Latent Space, Not Just Text
A new paper introduces Causal Visual Recurrent Reasoning (CVRR), a method that makes recurrent hidden-state computation the only i...
Procedural Graphs Give LLM Agents a Self-Evolving Map for Long-Horizon Tool Use
A new Hugging Face paper proposes Procedural Graphs, which organize procedural knowledge into (procedure, relation, procedure) tri...
BeaconKV Cuts Reasoning-Model Memory Up to 5.8x Without Losing Accuracy
A new training-free KV cache compression method from Hugging Face researchers uses "beacon queries" to predict which past key-valu...
SceneMosaic Blends Image Priors and AI Agents to Generate Diverse 3D Indoor Scenes 24x Faster
A new Hugging Face paper introduces SceneMosaic, a hybrid framework that initializes 3D indoor scenes from learned image priors an...
Mask Forcing Boosts Autoregressive Video Diffusion Distillation
A new Hugging Face paper proposes Mask Forcing, a dual-noise masking rollout strategy that curbs mode collapse in distilled autore...
Marigold V2 Turns Diffusion Transformers into Sharper, Single-Step Depth Estimators
Huawei Bayer Lab, EPFL and University of Bologna researchers repurpose a pretrained diffusion transformer into a monocular depth e...
OpenWAM Turns World-Action Model Pretraining Into an Open, Modular Science
Hugging Face paper introduces OpenWAM, an open research stack that factorizes world-action model pretraining into composable modul...
AuK: Hugging Face Paper Unveils Open-Source Model Unifying Speech Generation and Editing
A new technical report introduces AuK, an open-source foundational model that handles speech generation and editing through natura...