Generating Personalized Responses to the Same Query from Different Queriers

What is it about?

In real interactions, the person who asks—the querier—has large impact on the responses from the LLM responder.The goal is that, given the same query, the LLM should produce different responses that reflect each querier’s characteristics and relationship with the responder.To address the problem, we build the MQDialog dataset as a benchmark and propose a dual-tower LLM with querier-contrastive learning.

Featured Image

Photo by Alexander Shatov on Unsplash

Photo by Alexander Shatov on Unsplash

Why is it important?

Existing work on LLM personalization focus on responder-side roles, for instance, making an LLM act as a doctor, a teacher, or a friend.However, in real interactions, the person who asks—the querier—also matters.Imagine someone asking “Are you busy today?”If it’s your mentor, you might respond formally; if it’s your close friend, you’d answer casually.This observation motivates our new perspective: querier-aware personalization, generating distinct responses to the same query depending on the identity and relationship of the querier.