Research Papers
RLVR Narrows Reasoning Diversity at the First Step, Not During Execution
A new study shows that reinforcement learning with verifiable rewards (RLVR) reduces reasoning diversity primarily at the initial...
New Metric OVMI Offers Common Scale for Comparing Speech Brain-Computer Interfaces
Researchers introduce open-vocabulary mutual information (OVMI), an information-theoretic metric that enables fair comparison of s...
VeriPhy: Auditable Physical Reasoning System for World Model Evaluation
Hugging Face researchers introduce VeriPhy, an auditable physical-verification system that compiles prompts into typed physical ob...
QCell: New Query-Based Model Tackles Overlapping Cell Segmentation
Researchers introduce QCell, a query-based model that improves instance segmentation of overlapping microscopy cells via latent-sp...
AI Training as Percolation: New Theory Links SGD Collapse to Phase Transitions
Researchers model stochastic gradient descent as a percolation process, showing that neural networks collapse into simpler subnetw...
New CORD Adapter Preserves Top-1 Predictions in Post-Hoc Calibration
Researchers introduce CORD, a post-fit adapter that repairs calibrated probability vectors to exactly preserve original top-1 pred...
Frame Selection Beats Compression in Long-Video AI Models, Study Finds
A new controlled study from Hugging Face shows that which frames a long-video model selects matters far more than how much it comp...
WorldReward: A Unified Vision-Language Reward Model for Camera-Conditioned World Models
Researchers introduce WorldReward, a vision-language reward model that unifies action-consistency and visual-quality evaluation fo...
PACE: New Dataset and Framework to Help AI Assistants Detect Hidden User Conflicts
Hugging Face researchers introduce PACE, a dataset for evaluating whether personalized AI assistants can identify latent constrain...
RealSWE: Realistic User Requests Cut Coding Agent Performance by 6.4 Points
A new benchmark from Hugging Face reveals that coding agents perform significantly worse on realistic, casual user requests compar...
DRACO: Dynamic Rubrics Boost Long-Horizon Agent Training Without Verifiers
A new method, DRACO, dynamically generates rubrics and redistributes trajectory-level scores into per-step advantages for reinforc...
Hugging Face's CORE Distills Reranker Reasoning into Embedding Models for Better Compositional Retrieval
CORE improves compositional reasoning in multimodal embedding models by distilling ranking knowledge from a cross-attentive rerank...