Hugging Face Papers: TriAttention, OpenWorldLib, Vero VLMs, and SkillX Agent Learning Released

Hugging Face published a cluster of significant research papers this week. TriAttention addresses KV cache memory bottlenecks in long-context LLM reasoning using pre-RoPE trigonometric compression, enabling longer reasoning sequences without quality loss. OpenWorldLib provides a standardized unified inference framework and definitions for advanced world models, improving research reproducibility. Vero releases a fully open family of Vision-Language Models achieving broad visual reasoning across charts, science, and spatial tasks — previously locked behind proprietary RL pipelines. SkillX proposes automated collaborative skill knowledge base construction enabling LLM agents to share and reuse learned behaviors rather than rediscovering them independently. MinerU2.5-Pro additionally demonstrates that document parsing SOTA improvements stem from systematic training data engineering rather than architecture innovation.