Ang Lv

Coupling Experts and Routers in Mixture-of-Experts via an Auxiliary Loss

Union-of-Experts: Experts in Mixture-of-Experts are Secretly Routers

StepHint: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason

More Expressive Attention with Negative Weights

The Climb Carves Wisdom Deeper Than the Summit: On the Importance of Reasoning Patterns

PolarQuant: Leveraging Polar Transformation for Key Cache Quantization and Decoding Acceleration

Autonomy-of-Experts Models

GETMusic: Generating Music Tracks with a Unified Representation and Diffusion Framework

PEAR: Position-Embedding-Agnostic Attention Re-weighting Enhances Retrieval-Augmented Generation with Zero Inference Overhead