Hwanjo Yu

RA-VLA achieves superior task adaptation in robotic manipulation by seamlessly integrating context retrieval with execution, outperforming traditional methods.

Grafting trainable components onto frozen teacher models reveals the hidden pitfalls of generalization in sequential recommenders, leading to substantial performance gains.

TRACER balances the Stability-Plasticity-Cognitivity trilemma, leading to a significant performance boost in continual recommendation systems.

Self-attention in recommendation systems overlooks crucial item relations, but PRISM's multi-perspective approach reveals hidden user preferences that boost performance.

Valid reasoning steps can still be inefficient, leading to a 31-53% increase in token usage without improving outcomes.

LLMs waste context on redundant information when making recommendations; selectively augmenting only lesser-known items boosts accuracy and efficiency.

Achieve ensemble-level sequential recommendation performance with a single network at inference time by distilling diversity from a modular ensemble during training.

Naive fine-tuning of VLMs for multimodal sequential recommendation causes catastrophic modality collapse, but can be fixed with gradient rebalancing and cross-modal regularization.