Research Papers
TurnSight: A New Framework for Fine-Grained Credit Assignment in Tool-Integrated Reasoning
Hugging Face researchers introduce TurnSight, a turn-level hindsight self-distillation framework that improves reinforcement learn...
Hugging Face Unveils UniWorld-Design: Layer-Native Image Generation
UniWorld-Design redefines image generation by using semantic RGBA layers as atomic units, enabling structured composition and inst...
SkillJack: New Attack Turns Self-Evolving Agents' Learning into Persistent Backdoors
Researchers unveil SkillJack, the first attack targeting the experience-to-skill pipeline of self-evolving agents, implanting mali...
CAPEval: New Benchmark Decouples Caption Quality into Coverage and Precision
Researchers introduce CAPEval, a benchmark that separates caption quality into Coverage and Precision, revealing that Coverage pre...
Any-OPD: New Framework Enables On-Policy Distillation Between Any Flow-Matching Models
Researchers introduce Any-OPD, the first framework for on-policy distillation between arbitrary pairs of latent flow-matching gene...
OmniPack: Training-Free Token Compression Boosts Omni-Modal LLM Efficiency
A new training-free framework, OmniPack, coordinates structural and semantic token compression to cut computational costs in omni-...
Hugging Face Unveils LLaDA MoE v2: Scaling Laws for Diffusion Language Models
A new paper from Hugging Face introduces LLaDA MoE v2, a 30B-A3B diffusion language model trained on 23.5T tokens, and reveals sca...
PAST-Bench: New Benchmark Tests Whether Personal AI Agents Really Improve from Experience
Researchers introduce PAST-Bench, a benchmark with 26 scenarios and 204 episodes to isolate whether personal AI agents improve fro...
Hugging Face Paper Proposes Agent-Centric World Proxies to Rethink World Modeling
A new Hugging Face paper introduces Agent-Centric Interactive World Proxies, shifting world modeling from physical state predictio...
Hugging Face Unveils Video-DeepResearch: A Multimodal Agent That Watches Before It Searches
Hugging Face introduces Video-DeepResearch, a framework that extends multimodal agents from static images to continuous video stre...
Hugging Face Researchers Propose PCSD to Boost Agentic RL with Persistent Consistency Self-Distillation
A new method, Persistent Consistency Self-Distillation (PCSD), improves reinforcement learning for LLM agents by providing dense,...
Hugging Face Paper: Knowledge-Geometry Decoupling Boosts Streaming Recommendations by 4-12%
A new paper from Hugging Face introduces Knowledge-Geometry Decoupling (KGD), a method that improves streaming recommendation syst...