Research Papers
Hugging Face's ZimaBlue: Scaling Video Pretraining Boosts Zero-Shot Robot Manipulation
ZimaBlue, a new framework from Hugging Face, learns generalizable world action models from massive egocentric video, improving zer...
SMELT: Looping MoE Transformers Cuts Training FLOPs by Up to 18%
A new Hugging Face study introduces SMELT, a recipe that loops the middle layers of sparse Mixture-of-Experts Transformers twice,...
Hugging Face Unveils UI-Venus-2: A Unified Multimodal GUI Agent for Reliable Digital Automation
UI-Venus-2 is a general-purpose multimodal GUI agent that unifies reasoning and action loops across mobile, web, and desktop envir...
StudentSim: Training Personalized AI Student Simulators from Sparse Data
Hugging Face researchers introduce StudentSim, a framework that trains personalized student simulators from sparse data, outperfor...
New FACE-Eval Benchmark Shows Chain-of-Thought Monitoring Fails When Cues Arrive via Tools
A new 5,100-sample evaluation across 15 open-weight models reveals that chain-of-thought faithfulness drops when preference cues a...
KATok: Adaptive Tokenizer Drops Uninformative Tokens for Compact Video Representation
A new transformer-based video tokenizer, KATok, adaptively drops uninformative tokens to achieve data-dependent compression while...
MineAmongUs: New 3D Benchmark Reveals Non-Verbal Deception as Key to VLM Agent Wins
Hugging Face researchers introduce MineAmongUs, a 3D multimodal environment, and the ARIA harness to study embodied deception in v...
Hugging Face Unveils Matrix-Game 3.5: Geometry-Aware Memory for Real-Time Interactive Worlds
Matrix-Game 3.5 introduces geometry-aware memory, static-dynamic disentanglement, and progressive distillation to enable stable, l...
AutoSciRub: Evaluation-First Rubrics Boost Autonomous Research Agents
A new framework called AutoSciRub generates task-specific executable rubrics before research execution, guiding autonomous agents...
Hugging Face Paper Proposes Five-Level Ladder for Scaling Reasoning Models Beyond Human Supervision
A new paper from Hugging Face outlines a structured framework for scaling large reasoning models beyond human supervision, tracing...
PRISK Framework Reveals Hidden Costs of LLM Personalization
A new study introduces PRISK, a framework that systematically evaluates how personalization in LLMs can lead to irrelevant persona...
Super Library Agent: A New Framework for Joint Multi-App Code Generation and Maintenance
Hugging Face researchers introduce Super Library Agent (SLA), a framework that reduces redundancy and structural erosion when LLM...