Zhaoxin Feng (Betty)

Hello there

I am a second-year PhD student at The Hong Kong Polytechnic University, advised by Prof. Emmanuele Chersoni and co-advised by Prof. Huang Chu-Ren. I received my B.Sc. in Educational Technology (minor in Data Science) from Beijing Normal University in 2024.

My research asks a central question: can we rely on LLMs and LLM-based agents when they interact with humans? I study this at two levels: how models represent meaning internally, and how they behave when interacting with people.

Internal representations (reliability and interpretability). I probe how LLMs represent human language: whether enabling bidirectional attention in autoregressive LLMs reshapes their lexical semantic representations, and how their embeddings compare to human mental lexicon judgments. The question is whether a model's internal grasp of meaning is faithful enough to be relied on.

Behavioral reliability in interaction. I study how LLMs and agents behave with human users, from sycophancy that chain-of-thought reasoning both mitigates and masks, to reasoning failures on Chinese phonological ambiguities, to memory agents that recall user preferences but fail to act on them.

In the long run, I hope to contribute to personalization: a personal, long-term agent that deepens its memory through continuous interaction with a user, one that truly understands your context and stays reliable as it does so. My PhD thesis, Rethinking Memory for Long-Term LLM Agents: Not Just Remembering, But Acting, takes a step toward this through agentic memory: how preferences, explicit and implicit, are extracted from everyday interactions; how memory is retrieved, updated, and forgotten; and how to evaluate not just what an agent stores but whether it reliably acts on it.