PRIME: Large Language Model Personalization with Cognitive Dual-Memory and Personalized Thought Process

TL;DR

PRIME introduces a cognitive-inspired framework for LLM personalization, combining episodic and semantic memory mechanisms, and demonstrates its effectiveness with a new Reddit-based benchmark.

Contribution

It presents a unified theoretical framework for LLM personalization based on cognitive dual-memory, along with a novel personalized thinking strategy and a new evaluation dataset.

Findings

PRIME outperforms baseline methods in personalization tasks.

The framework effectively captures dynamic user preferences.

The new dataset enables comprehensive long-context personalization evaluation.

Abstract

Large language model (LLM) personalization aims to align model outputs with individuals' unique preferences and opinions. While recent efforts have implemented various personalization methods, a unified theoretical framework that can systematically understand the drivers of effective personalization is still lacking. In this work, we integrate the well-established cognitive dual-memory model into LLM personalization, by mirroring episodic memory to historical user engagements and semantic memory to long-term, evolving user beliefs. Specifically, we systematically investigate memory instantiations and introduce a unified framework, PRIME, using episodic and semantic memory mechanisms. We further augment PRIME with a novel personalized thinking capability inspired by the slow thinking strategy. Moreover, recognizing the absence of suitable benchmarks, we introduce a dataset using Change My View (CMV) from Reddit, specifically designed to evaluate long-context personalization. Extensive experiments validate PRIME's effectiveness across both long- and short-context scenarios. Further analysis confirms that PRIME effectively captures dynamic personalization beyond mere popularity biases.