Latest stories
Coherent Overlap: Rethinking Sparse MoE Routing Beyond Geometric Complementarity
A new study from Hugging Face researchers challenges the geometric view of sparse mixture-of-experts (MoE) routing, showing that c...
AI Tour Meeting: Multi-Agent LLM Framework for Collaborative Group Travel Planning
Researchers propose AI Tour Meeting, a framework where multiple LLM-based agents with distinct personas collaborate through natura...
Study: Deep Research Agents Easily Fooled by Misleading Information
A new framework, MisKnow-Agent, reveals that Deep Research agents adopt false conclusions when exposed to even a single misleading...
OmniScope: Training-Free Token Compression Boosts Omnimodal LLM Efficiency
A new training-free framework, OmniScope, improves token compression for omnimodal LLMs by decoupling salience estimation across a...
Harness-G: Graph-Structured Interface Boosts Search Agent Performance
A new framework from Hugging Face researchers, Harness-G, replaces free-form query generation in RL search agents with a finite me...
Hugging Face Researchers Expand Pedestrian Archetypes to 19 for Safer AV Testing
A new preprint from Hugging Face introduces seven additional pedestrian archetypes, expanding the original taxonomy to 19 to bette...
AMRD: Adaptive Multi-Teacher Relational Distillation for Lightweight Speech Emotion Recognition
Researchers propose AMRD, a novel distillation method that compresses large speech emotion recognition models into lightweight stu...
Hugging Face Study Dissects Lossy Verification in Speculative Decoding, Revealing Failure Modes
A new paper from Hugging Face provides a principled analysis of lossy verification in speculative decoding, classifying methods in...
ReToken: A Single Learnable Token Boosts Vision-Language Models for Visual Retrieval
ReToken introduces a single learnable embedding that acts as an explicit retrieval target, selecting sparse query-relevant visual...
Filesystem Memory for LLM Agents: First Systematic Study Finds Organization Cuts Costs but Not Accuracy
A new study from Hugging Face researchers provides the first systematic exploration of filesystem-based memory for LLM agents, rev...
Multi-Head Attention Residuals: A New Routing Mechanism for Transformers
Hugging Face researchers introduce Multi-Head Attention Residuals (MHAR), a zero-parameter enhancement to attention residuals that...
Echoverse: Evolving Synthetic Environments Boost Computer-Use Agents from 36.5% to 67.1%
Microsoft Research introduces Echoverse, a pipeline that compiles specifications into stateful synthetic applications with grounde...