Research Papers
Relay-OPD: Fixing Prefix Failure in On-Policy Distillation with Teacher-Student Handoffs
A new method called Relay-OPD detects when a student model goes off track during training and lets the teacher briefly take over,...
CodeNib: Multi-View Data System Boosts Coding Agent Efficiency by Up to 25x
Hugging Face researchers introduce CodeNib, a data system that builds reusable lexical, dense, and structural views per repository...
Hugging Face Unveils Wonder: A Real-Time, Camera-Controllable Video World Model
Wonder is a general-purpose video world model that enables real-time, camera-controllable exploration of generated worlds, support...
Mage-VL: Microsoft's Efficient Streaming Multimodal Model Cuts Visual Tokens by 75%
Mage-VL, a new codec-native streaming foundation model from Microsoft, overcomes Moravec's paradox in vision-language models by us...
Hugging Face Researchers Expose 'Implicit-Association Blind Spot' in AI Agent Memory Systems
A new benchmark, InMind, reveals that state-of-the-art agent memory systems fail to apply stored facts when indirect reasoning is...
ReDesign: Agentic Framework Recovers Editable Design Files from Raster Images
Hugging Face researchers introduce ReDesign, an agentic framework that decomposes raster images into editable layer hierarchies wi...
Relevance-Aware RipGrep Agent Boosts Accuracy-Efficiency Frontier in Agentic Search
Hugging Face researchers introduce RARG, a search agent that uses relevance to guide corpus interaction, improving accuracy and ef...
HiFi-UMI: Robot-Free Data Matches Teleoperation for Deployable Manipulation Policies
HiFi-UMI, a high-fidelity robot-free data collection system, enables deployable manipulation policies that match or exceed teleope...
OpenAI Report: AI Agents Transform Scientific Computing in Genomics
A new field report from OpenAI reveals how scientists are leveraging AI coding agents to accelerate software development and disco...
PAJAMA: Distilling LLM Judges into Transparent, Low-Cost Programs
Researchers introduce program distillation to replace expensive LLM-as-a-judge evaluations with a committee of transparent, editab...
REDE: Denoising Reasoning Traces to Boost Hallucination Detection in Large Reasoning Models
Researchers propose REDE, a framework that denoises reasoning traces by removing irrelevant and repetitive steps, improving halluc...
Frozen 12B Model Achieves 100% Accuracy via Verified Memory, Not Raw Reasoning
A new approach from Hugging Face stores verified solutions in a persistent memory, enabling a frozen 12B model to answer 180/180 i...