Hello, I am Ying Zhang

I am a postdoc in the Natural Language Understanding Team, RIKEN Center for Advanced Intelligence Project (AIP), under the supervision of Prof. Kentaro Inui at Tohoku University.

My current research focuses on the mechanistic interpretability of language models: how factual knowledge is acquired during training, how it is stored in model internals, and whether it is genuinely used when models answer. To track my newest publications, you can also visit my ORCID, Semantic Scholar, or Google Scholar.

Keywords: Mechanistic Interpretability · Factual Knowledge in Language Models · Fact Recall

My research asks how factual knowledge lives inside language models — how it is acquired during training, how it is stored in parameters and representations, and whether it is genuinely used when models answer. I approach these questions through mechanistic interpretability: training dynamics, probing, and causal analyses of internal representations.

Our recent work (Special Committee Award, NLP2026) identifies the core mechanism behind generalized recall.

Earlier, my Ph.D. dissertation addressed text degeneration in neural text generation with re-ranking methods: Addressing Text Degeneration of Discriminative Models with Re-ranking Methods.