How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants

TL;DR

The paper investigates rational personalization in LLM-powered assistants by formalizing memory utilization as a pragmatic reasoning task and introducing RPEval, a benchmark that jointly tests memory applicability and intent understanding. It defines a three-level personalization framework (L0/L1/L2) and develops a data-generation pipeline and a multi-granularity evaluation protocol to reveal widespread irrational personalization in current LLMs. The authors propose RP-Reasoner, a pragmatic, Bayesian-inspired module that estimates intent priors from memory and uses query-facing cues to decide memory integration, achieving notable gains over baselines and resolving most problematic cases in real-world deployments. The work highlights significant practical implications for safer, more user-aligned personalization and calls for privacy-preserving, responsible design as memory-enabled assistants scale.

Abstract

Large language model (LLM)-powered assistants have recently integrated memory mechanisms that record user preferences, leading to more personalized and user-aligned responses. However, irrelevant personalized memories are often introduced into the context, interfering with the LLM's intent understanding. To comprehensively investigate the dual effects of personalization, we develop RPEval, a benchmark comprising a personalized intent reasoning dataset and a multi-granularity evaluation protocol. RPEval reveals the widespread phenomenon of irrational personalization in existing LLMs and, through error pattern analysis, illustrates its negative impact on user experience. Finally, we introduce RP-Reasoner, which treats memory utilization as a pragmatic reasoning process, enabling the selective integration of personalized information. Experimental results demonstrate that our method significantly outperforms carefully designed baselines on RPEval, and resolves 80% of the bad cases observed in a large-scale commercial personalized assistant, highlighting the potential of pragmatic reasoning to mitigate irrational personalization. Our benchmark is publicly available at https://github.com/XueyangFeng/RPEval.

Figures (23)

• Figure 1: Different levels of PAs. In $L_1$, memory is directly concatenated with the query, whereas in $L_2$, the PA infers implicit cues from the user’s query to determine memory utilization strategy.
• Figure 2: An illustration of our proposed RPEval: (a) Personalized Intent Reasoning Data Generation; (b) Dataset Expansion Strategies; (c) Multi-Granularity Evaluation Protocol.
• Figure 3: Performance of major LLMs on the discriminative intent matching accuracy in RPEval.
• Figure 4: RP-Reasoner: An Implementation of Pragmatic Personalized Assistant.
• Figure 5: RP-Reasoner achieves notable gains in multi-preference generative settings.