Latest stories
1% of Tokens Can Be Enough: New Gradient-Estimation Method Sharpens Sparse On-Policy Distillation
A new paper proposes an information-efficiency ratio (IER) that measures how reliably a token's gradient can be estimated, enablin...
CARE Teaches Robots to Recover: Failure-Driven Correction Boosts VLA Success
Hugging Face paper introduces CARE, a data-centric framework that learns from failed robot rollouts to synthesize corrective demon...
Deep Persona: A Three-Layer Psychological Architecture for More Convincing Role-Playing AI Agents
A new paper proposes Deep Persona, a psychologically grounded three-layered architecture that organizes AI personas into observabl...
Harness-Zero Distills Agent Harness Gains Directly Into Model Weights
A new Hugging Face paper proposes Harness-Zero, a method that transfers the benefits of a specialized agent harness into a model's...
Hugging Face Paper Unveils Jev-Mem, a 'System-One' Memory Layer for Faster AI Agents
Jev-Mem splits agentic memory management into a fast System-One control plane and a slower System-Two reasoning plane, reporting b...
Alibaba's Category-Aware Expert Training Boosts SWE Agents to 58% on Pro-618
A new framework from Alibaba trains category-specific software engineering experts via Refresh-Repair-Expand, then consolidates th...
Hugging Face Releases onPanda, a Token-Level Tool That Cuts LLM Annotation Time by 52%
onPanda is an open-source interactive annotation tool that lets humans correct model outputs one token at a time, then resume gene...
D-RAC Brings Retrieval-Aware Chunking to Every Enterprise Document Format
A new Hugging Face paper extends the W-RAC framework into D-RAC, normalizing any document to PDF and converting pages to retrieval...
Grounded Action Models Put 3D Grounding at the Core of Robot Foundation Models
A new Hugging Face paper proposes Grounded Action Models (GAMs), robot foundation models conditioned on language, points, or box p...
Realtime-Venus: Hugging Face Paper Unveils Full-Duplex AI That Talks and Delegates Tasks
A new paper presents Realtime-Venus, a proactive full-duplex interaction system built on two separately trained 9B models that com...
OpenAI Sets Out Principles for Independent Third-Party AI Safety Audits
OpenAI has published its priorities and principles for independent third-party assessments of frontier AI models, emphasizing rigo...
Hugging Face Releases OmniEdu: Open 4B/9B/27B Foundation Models for K-12 Teaching and Learning
OmniEdu is an open family of 4B, 9B, and 27B education foundation models trained on a capability-organized corpus spanning subject...