Tuo Zhao

ARMOR: High-Performance Semi-Structured Pruning via Adaptive Matrix Factorization

COSMOS: A Hybrid Adaptive Optimizer for Efficient Training of Large Language Models

Convergence of Near-Linear Width ReLU Networks with Unbalanced Initialization

NorMuon: Making Muon more efficient and scalable

On the Efficiency of Structured Pruning in Small Language Model Pretraining

LLMs Can Generate a Better Answer by Aggregating Their Own Responses

AdaSPEC: Selective Knowledge Distillation for Efficient Speculative Decoders

Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models

A Minimalist Example of Edge-of-Stability and Progressive Sharpening

Ask a Strong LLM Judge when Your Reward Model is Uncertain

NoWag: A Unified Framework for Shape Preserving Com- pression of Large Language Models

SlimMoE: Structured Compression of Large MoE Models via Expert Slimming and Distillation

Deep Reinforcement Learning from Hierarchical Preference Design

Adversarial Training of Reward Models

Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data

Self-Rewarding PPO: Aligning Large Language Models with Demonstrations Only

Model Tells Itself Where to Attend: Steerable Prompting for Reliable Reading Comprehension of LLM