Shawn Im

How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability

Normalized Rewards for Preference Learning

Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders

Visual Instruction Bottleneck Tuning

Can DPO Learn Diverse Human Values? A Theoretical Scaling Law

Understanding Multimodal LLMs Under Distribution Shifts: An Information-Theoretic Approach

On the Generalization of Preference Learning with DPO

Understanding Generalization of Preference Optimization Under Noisy Feedback