On Possible Personalization of User Interaction with Large Language Models
Received: 10 Oct 2026 Revised: 20 Oct 2026 Accepted: 22 Oct 2026
Published: 2026, vol. 30, issue 4, pp. 316–319
Abstract
On two LLMs (DeepSeek, MiMo) we study two types of personalization, comparing their effect with the noise level. Passive personalization (user history in the prompt) on LaMP gives a robust average gain (16–50%), but the share of “harmed” users does not exceed the noise level. For active personalization we propose reranking results by user ratings, simulated by LLM personas: personas shift ratings beyond the noise, but change the top real answer no more often than noise (DeepSeek) or only slightly more often (MiMo, +12 pp).
Keywords: large language models, personalization, reranking, LLM personas.
BibTeX
@article{IS-Agazaryan-Ryzhov2026,
author = {Agazaryan, Alexander Anastasovich and Ryzhov, Alexander Pavlovich},
title = {{On Possible Personalization of User Interaction with Large Language Models}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {316--319},
}
AMSBIB
\Bibitem{IS-Agazaryan-Ryzhov2026}
\by A.\,A.~Agazaryan, A.\,P.~Ryzhov
\paper On Possible Personalization of User Interaction with Large Language Models
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 316--319
\lang In Russian
RU