Xingwu Sun

Towards a Comprehensive Scaling Law of Mixture-of-Experts

Fast Data Mixture Optimization via Gradient Descent

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

HS-SFT: Hybrid Sparse Supervised Fine-tuning for Offline LLM KV Cache Eviction

PT$^2$-LLM: Post-Training Ternarization for Large Language Models

Proximal Supervised Fine-Tuning

More Expressive Attention with Negative Weights

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

Understanding Data Influence in Reinforcement Finetuning

Scaling Laws for Floating–Point Quantization Training

Autonomy-of-Experts Models