Research Papers
Skill-α: RL-Based Progressive Skill Generation Boosts Agent Performance
Researchers introduce Skill-α, a reinforcement learning method that generates high-quality agent skills through progressive editin...
Hugging Face Researchers Propose DAPD to Fix 'Privilege Illusion' in On-Policy Distillation
A new paper from Hugging Face introduces Dual-Anchored Policy Distillation (DAPD), a framework that mitigates the 'privilege illus...
Hugging Face's SKT: Verified Synthetic Data Boosts Agent Skill Use at Scale
A new Hugging Face research paper introduces SKT, a pipeline that generates verified synthetic training data to improve how langua...
Hugging Face Unveils UEmbed: Unified Sparse and Dense Multimodal Embeddings
Hugging Face introduces UEmbed, a decoder-only multimodal embedding model that generates both sparse and dense representations in...
WorldExam: New Benchmark Puts World Models to the Test for Inherent Reactivity
Hugging Face researchers introduce WorldExam, a hierarchical benchmark that evaluates controllable video generation models beyond...
Hugging Face Researchers Introduce VAD: A Counterfactual Approach to Visual Distillation
A new paper from Hugging Face introduces Visual Attribution Distillation (VAD), a counterfactual algorithm that isolates visually...
Hugging Face Unveils SwanTale: A Unified Model for Multi-Speaker Speech and Audio Generation
SwanTale, a new model from Hugging Face, unifies zero-shot and instruction-based multi-speaker speech and audio generation, achiev...
LongHorizon-Harness: A New Framework to Keep AI Agents on Track in Long Tasks
Researchers introduce LongHorizon-Harness, a task-state management framework that decouples execution from state tracking, boostin...
Hugging Face Research: CSCR Reallocates Token Credit to Improve Long-CoT Reasoning
A new paper proposes Counterfactual Sensitivity Credit Reallocation (CSCR), a simple extension of GRPO that reduces credit for hig...
EMBL AI Librarian: A Natural-Language Knowledge Layer for Life-Science Agents
Hugging Face researchers unveil EMBL AI Librarian, a knowledge layer that lets life-science AI agents query Europe PMC in natural...
RL^2-VLA: Adaptive Steering Boosts VLA Robots in Out-of-Domain Tasks
Researchers introduce RL^2, an adaptive inference-time steering framework that applies reinforcement learning on VLA latents only...
ODEWorld: Continuous-Time World Modeling via Physical-Time Flow
Researchers introduce ODEWorld, a continuous-time latent world model that learns an ODE-based velocity field in physical time, ena...