Research Papers
QQWorld: New Regularization Method Sharpens Latent World Models
Researchers propose QQWorld, a quantile-quantile matching objective that replaces the Epps-Pulley regularizer in latent world mode...
N_0-VTLA: First Tactile-Pretrained VTLA Model Boosts Contact-Rich Robot Manipulation
Hugging Face researchers introduce N_0-VTLA, a vision-tactile-language-action foundation model that integrates tactile sensing at...
New AISPA Framework Audits System Prompts Across 88 Commercial AI Products
Researchers introduce AISPA, a user-centric framework for auditing system prompts in AI applications, revealing wide variation in...
Hugging Face Researchers Unveil RLSVR: A New Paradigm for Self-Improving LLMs on Open-Ended Tasks
A new paper from Hugging Face introduces Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation method...
Fairness Pruning: A Surgical Method to Locate Demographic Bias in LLMs
Researchers introduce Fairness Pruning, a lightweight intervention that locates neurons responsible for demographic bias in GLU-ML...
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...