Research Papers
Hugging Face Paper Explains Why Hybrid Attention LLMs Work — and How to Design Them Better
A new paper, "Mechanics of Long-Context Hybrid Models Part 1.1," analyzes how full attention, sliding-window attention, and linear...
RobotWorld Benchmark Tests How Well AI Agents Handle the Physical World
Hugging Face paper introduces RobotWorld, an 84-task simulation benchmark probing whether general-purpose multimodal agents can tr...
ReSAIL: New Method Prevents Collapse in Iterative Agent Self-Distillation
Hugging Face researchers introduce ReSAIL, a plug-in augmentation that mitigates performance collapse in iterative privileged-info...
WorldSonus Adds Real-Time Spatial Sound to Silent World Models
A new Hugging Face paper introduces WorldSonus, an interactive video-to-audio framework that generates spatially aligned stereo so...
Hugging Face Paper Proposes RunningTab to Keep Long-Horizon Agents on Task
RunningTab adds an environment-side tab that tracks what a task still owes, what files an agent has read, and which files were lis...
Tetris3D Builds 3D Scenes Where Every Object Fits Together
Researchers propose Tetris3D, a framework that reconstructs physically and geometrically coherent 3D scenes from a single image by...
Hugging Face Paper Unveils UniWAM, a Unified World-Action Model for Embodied AI
UniWAM combines a physical reasoner, a world generator, and an action predictor into one architecture trained on over 10,000 hours...
SGF+ Decouples Gradient Flows to Push Autoregressive Video Generation to 24 Hours
A new Hugging Face paper introduces Self Gradient Forcing Plus (SGF+), which separates the parameters used for context writing and...
VepAgent Uses Tool-Augmented RL to Predict Future Video Events
A new Hugging Face paper introduces VepAgent, an agentic framework that combines causal-transition reasoning with tool-augmented r...
DecepEval Benchmark Maps When LLM Agents Turn Deceptive
A new Hugging Face paper introduces DecepEval, a benchmark of 1,532 paired instances that measures how external conditions such as...
UltraText Bench Tests Whether Image Generators Can Render Dense Bilingual Text
A new bilingual benchmark of 432 prompts across 24 real-world scene categories evaluates how well image generators reproduce dense...
Hugging Face Paper: Self-Retrospection Distillation Turns Hindsight into Agent Foresight
A new Hugging Face paper proposes Self-Retrospection Distillation (SRD), a lightweight auxiliary loss that uses post-hoc trajector...