Song Mei

GSM-Agent: Understanding Agentic Reasoning Using Controllable Environments

How Do Transformers Perform Two-Hop Reasoning in Context?

Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning

Generalization or Hallucination? Understanding Out-of-Context Reasoning in Transformers

Implicit Bias of Gradient Descent for Non-Homogeneous Deep Networks

U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models

A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics

Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs

Improving LLM Safety Alignment with Dual-Objective Optimization

Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearning

Choose Your Anchor Wisely: Effective Unlearning Diffusion Models via Concept Reconditioning