Yi Cao

Interface health in generative recommendation systems is multi-signal, revealing that prefix alignment is critical for candidate exposure but weakens under certain scoring conditions.

Optimizing function evaluations in flow-based generative models can dramatically enhance image restoration quality without retraining.

Achieving near-lossless compression in visual document retrieval, AnchorFold retains over 98% retrieval quality even at 5x compression ratios.

Contextual tunneling in LLMs can be overcome, leading to more reliable and physically grounded materials discovery through the innovative ARIA framework.

SKINNs let you bake in your existing domain knowledge as hard constraints in your neural nets, leading to better generalization and interpretability.

Shrinking visual document retrieval storage by 95% is now possible without sacrificing accuracy, thanks to a layout-aware parsing strategy.

Generative recommenders get a major upgrade: HPGR leverages hierarchical pre-training and sparse attention to dramatically improve performance and efficiency by explicitly modeling the structure of user behavior.

The first comprehensive survey of Visual Document Retrieval reveals how MLLMs are reshaping the field, highlighting the shift towards RAG and agentic systems for complex document understanding.

Multi-vector visual document retrieval gets a speed boost without sacrificing accuracy thanks to a novel "Prune-then-Merge" approach that intelligently compresses visual features.