Research Papers
Co-RL: Multi-Agent RL Enables Unsupervised Reasoning Without Labels
Hugging Face researchers introduce Co-RL, a multi-agent reinforcement learning framework where models learn from peer-derived rewa...
Hugging Face Unveils Zetta: A Closed-Loop Harness for Self-Evolving Robots
Zetta, a new closed-loop embodied harness from Hugging Face, enables robots to evolve runtime critics and recovery skills online,...
HarnessRisk Benchmark Exposes Critical Configuration Flaws in AI Agent Safety
A new benchmark evaluates AI agent harness safety across six operational phases, finding that configuration vulnerabilities and de...
Energy-Guided Flow Matching: Coarse-to-Fine Generation with Moving Endpoints
A new method called Energy-Guided Flow Matching (EG-FM) improves image generation by explicitly modeling a coarse-to-fine trajecto...
SkillForge: Self-Distilling Agents Learn Project-Specific Skills Before Fixing Bugs
Hugging Face researchers introduce SkillForge, a self-distillation framework that synthesizes repository-specific issues to proact...
Hugging Face Researchers Propose Capability-Centric Data Design for Generalist Image Generation
A new paper from Hugging Face introduces a capability-driven data infrastructure with curriculum scheduling and specialized data e...
No One-Size-Fits-All Memory: New Study Calls for Adaptive Substrate Routing in LLM Agents
A comprehensive evaluation of memory substrates for long-horizon LLM agents reveals that no single memory type dominates, highligh...
Hugging Face Unveils MoE-ViE: Efficient Vision Encoders Outperform Dense Models
Researchers at Hugging Face introduce MoE-ViE, a family of Mixture-of-Experts vision encoders that achieve state-of-the-art perfor...
Hugging Face Researchers Unveil Multi-Byte Prediction to Speed Up Byte-Level Language Models
A new paper from Hugging Face introduces multi-byte prediction (MBP), a method that generates multiple bytes in parallel in byte-l...
LEGO-RL: New Framework Bridges Coding-Agent Harnesses with Scalable Reinforcement Learning
Hugging Face researchers introduce LEGO-RL, a framework that enables policy-gradient training directly on native coding-agent harn...
DiSCO: A Black-Box Defense Against Unsafe Text-to-Image Generation
Researchers propose DiSCO, a zero-shot, black-box defense that optimizes prompts to reduce harmful image generation without modify...
EditBridge: A Diffusion Bridge for Faithful, Efficient 4K Image Editing
Researchers propose EditBridge, a diffusion bridge framework that enables faithful and efficient ultra-high-resolution image editi...