Research Papers
VCSD: Visual Contrastive Self-Distillation Boosts VLMs Without External Teachers
Researchers propose Visual Contrastive Self-Distillation (VCSD), a novel on-policy self-distillation method that improves vision-l...
ReferTrack: A New Paradigm for Embodied Visual Tracking Achieves SOTA on EVT-Bench
Hugging Face researchers introduce ReferTrack, a referring-then-tracking paradigm that grounds embodied visual tracking using a si...
K12-KGraph: Curriculum-Aligned Knowledge Graph Boosts Educational LLMs
Researchers introduce K12-KGraph, a curriculum-aligned knowledge graph from Chinese textbooks, along with a 23,640-question benchm...
Hugging Face Unveils AREX: A Recursively Self-Improving AI Agent for Deep Research
Hugging Face introduces AREX, a family of recursively self-improving agents that alternate between evidence gathering and constrai...
Anchor-Align: New Method Boosts VLA Robot Generalization by Preserving Pretrained Representations
Researchers propose Anchor-Align, a finetuning method for vision-language-action (VLA) policies that prevents representation drift...
New Scaling Laws Reveal Hypernetworks as a Scalable Alternative for Knowledge Injection in LLMs
A new study from Hugging Face introduces scaling laws for hypernetwork-based knowledge injection in LLMs, demonstrating that hyper...
ActiveVision Benchmark Reveals MLLMs Collapse on Active Observation Tasks
A new benchmark, ActiveVision, tests whether multimodal large language models (MLLMs) can perform active visual observation—repeat...
Hugging Face Paper Proposes Rubric-Oriented Document Set Selection Beyond Relevance
A new framework, SetwiseEvalKit, evaluates document sets as a whole—measuring coverage, conflict, and complementarity—rather than...
Self Gradient Forcing: New Method Boosts Long Video Generation from Short Training Clips
Researchers propose Self Gradient Forcing (SGF), a two-pass training strategy that enables autoregressive video diffusion models t...
SLAI T-Rex Achieves 34.22% MFU on Ascend NPU for DeepSeek-V4 Post-Training
A new hierarchical optimization framework enables full-parameter post-training of trillion-parameter MoE models on Ascend NPU Supe...
Hugging Face Unveils AlayaWorld: Open-Source Interactive Video World Model
AlayaWorld is an interactive long-horizon video world model that generates 24-fps video at 540p and 720p from text, image, or vide...
Hugging Face Unveils Mage-Flow: A Compact 4B Model for Efficient Image Generation and Editing
Mage-Flow is a 4B-parameter generative stack that achieves competitive quality in text-to-image generation and instruction-based e...