Research Papers
nanoMuse: An Open-Source Personal Agent for Every Device You Own
A new report introduces nanoMuse, a GPL-3.0 open-source personal agent that runs across a person's phone, computer and the web, sh...
GRACE Compresses Video Diffusion Latents 8x, Cutting Wan2.1 Latency by 11.1x
A new two-stage framework from KAIST AI and Kakao Corp. compresses a pretrained video autoencoder while keeping it compatible with...
New R-Quest Method Keeps Self-Evolving Reasoning Models From Collapsing
A new paper from Hugging Face examines why self-evolving reasoning models degrade over repeated training rounds, identifying inval...
Hugging Face Paper Unveils Recursive Game Creator, an Agentic Harness for Making Games Fun
A new agentic framework called Recursive Game Creator uses four collaborative roles — Designer, Builder, Player, and Reviewer — to...
Long-WAM Scales Robot Context to 19.2 Seconds for Real-Time Control
Hugging Face paper introduces Long-WAM, a model-system framework that scales visual history in causal world-action models, raising...
STEPQuant Cuts Linear-Attention Memory by Up to 68.7% With 6-Bit Recurrent States
A new post-training quantization framework from Hugging Face researchers targets the recurrent states of Delta-rule linear attenti...
AGO AI Quality Gate Turns RAG Release Calls Into Defensible Decisions
A new evidence-first quality-gate framework for retrieval-augmented generation treats missing data and judge errors as explicit ou...
DiffGate Merges Teacher Guidance With Verifiable Rewards for On-Policy Distillation
A new Hugging Face paper proposes DiffGate, an outcome-gated objective that applies selective, bounded teacher supervision only to...
Hugging Face Paper Proposes Selection-Based Reasoning to Speed Up Multimodal Search Agents
A new Hugging Face paper introduces Selection-based Structured Reasoning (SSR), which replaces free-form reasoning with selection...
TimeBraid Weaves Language and Time-Series Models Into One Unified System
Hugging Face researchers introduce TimeBraid, a family of unified time-series and language models that fuse a pretrained Qwen3 LLM...
RLHND Turns Video Foundation Models Into Physically Grounded Hand Trackers
A new research paper introduces RLHND, a video foundation model-based hand tracker that jointly estimates hand pose and tactile in...
Kinematic MeanFlow Brings One-Step Action Generation to Robotic Foundation Models
A new paper introduces Kinematic MeanFlow (K-MF), a one-step action policy that overcomes the performance collapse of directly app...