Jian Wu

VLMs can be transformed into pixel-precise structural document parsing experts, achieving state-of-the-art OCR performance by enforcing syntactic validity and structural integrity through reinforcement learning.

LLMs struggle significantly with case-based reasoning in first-order logic, performing far worse than on linear reasoning tasks, exposing a key limitation in their mathematical capabilities.

LLMs can generate unbiased pseudo-labels for unexposed items in pre-ranking, boosting click-through rate by 3.07% in production while improving diversity.

Taobao's recommender system just got a 1.65% CTR boost by compressing ultra-long user behavior sequences with a hierarchical codebook and sparse attention, proving that personalized interest centers can be learned efficiently.