Latest stories
Hugging Face Releases Zing-0.5, a 5B Real-Time Playable World Model
Zing-0.5 is a 5B autoregressive world model that lets users explore generated worlds through joint keyboard and online text contro...
Hugging Face Paper Proposes XConf: Confidence Estimation Grounded in a Model's Own Experience
A new paper introduces XConf, an experiential confidence estimator that retrieves a model's own graded past episodes to judge how...
Hugging Face Paper Exposes 'Value Flattening' in PPO Critics, Proposes SP3O Fix
A new paper identifies Value Flattening, a systematic failure mode in PPO critics used for LLM reinforcement learning, and introdu...
LimiX-2 Debuts Contextual Mechanism Networks for Structured-Data Intelligence
Hugging Face researchers unveil LimiX-2, a pretrained foundation model built on Contextual Mechanism Networks and Context-Conditio...
EvolveTrade Lets LLM Trading Agents Rewrite Their Own Playbooks
A new Hugging Face paper introduces EvolveTrade, a framework that treats an LLM trading agent's system prompt as an evolving polic...
Hugging Face Paper Unveils VC-Attention for Faster Low-bit Video Generation
VC-Attention is a training-free low-bit attention framework that pairs value smoothing with a fused probability cast, delivering u...
Agora Turns Git Into Shared Memory So AI Research Agents Stop Repeating Each Other
A new Hugging Face paper proposes Agora, a Git-backed append-only DAG that records every research claim as a rerunnable commit, an...
ActionPiece Rethinks Action Tokenization for Vision-Language-Action Models
A new Hugging Face paper introduces physical rank consistency (PRC) to measure how well action tokenizers preserve relational fide...
ProgramDistill Benchmark Tests Coding Agents on Features Inferred from Working Apps
A new benchmark, ProgramDistill, evaluates coding agents on features they must discover by interacting with fully functional refer...
ScienceIDE Turns Scientific Code Repositories into Training Grounds for AI Agents
A new Hugging Face paper introduces ScienceIDE, infrastructure that converts scientific code repositories into executable, verifia...
ImpossibleRubrics Benchmark Exposes How Easily LLM Reward Rubrics Are Gamed
A new benchmark of 169 impossible tasks shows that rubric-based reward signals can be exploited by adversarial answers, with tailo...
Gavel Reads Skill Routing Signals Inside a Frozen LLM's Hidden States
A new Hugging Face paper proposes Gavel, a method that extracts skill-routing signals from a frozen LLM's hidden states using two...