Sangdoo Yun

RL makes MLLMs see better than SFT

Aligned Novel View Image and Geometry Synthesis via Cross-modal Attention Instillation

Dr.LLM: Dynamic Layer Routing in LLMs

Less is Not Worse: Effective Reasoning Without Complete Reasoning Traces

Vision Transformers Secretly Crave Noise

KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction

DaWin: Training-free Dynamic Weight Interpolation for Robust Adaptation

Probabilistic Language-Image Pre-Training

Token Bottleneck: One Token to Remember Dynamics

Versatile Motion-Language Models for Multi-turn Interactive Agents