Research map / Large language models / Alignment and preferences

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

NeurIPS 2026 · Large language models › Alignment and preferences · approach: LLM-as-judge and feedback data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces RPEval, a benchmark for rational preference utilization in personalized LLM assistants (intent labels Ignore/Support/Dominate with multi-granularity error-based evaluation), and RP-Reasoner, a Bayesian pragmatic reasoning method that reduces irrational personalization.

Problem

LLM-based personalized assistants insert irrelevant personal memories into context, causing over-personalization that misreads user intent; existing benchmarks assume memories should always be used and do not measure when to use them.

Approach

Frames memory use as pragmatic reasoning: RP-Reasoner ranks candidate intents by combining a query-likelihood estimate (LLM judges closeness of the observed query to simulated queries under each intent, with counterfactual elimination) and an intent-prior estimate from memory, fused via rank sum.

Key results

• RP-Reasoner improves Macro-acc by a relative 258% and Micro-acc by 35% over the best baseline in the multi-preference generative setting.
• RP-Reasoner reduces overall error severity by 26% relative to the best baseline.
• Humans reach 86% accuracy on Ignore cases in single-explicit setting versus 6%-38% for LLMs; multi-preference shows ~40% micro and ~90% macro gaps to humans.
• RP-Reasoner resolves about 80% of bad cases in a deployed commercial personalized assistant.

Surprising findings

• More capable LLMs perform worse at ignoring irrelevant preferences (inverse scaling in rational personalization).
• Failures are attributed to attraction bias, where LLMs reuse and amplify tokens from the context during decoding.

Limitations

• Benchmark does not test long-context memory recall or memory-based QA factual accuracy.
• Intent labels are inherently subjective despite double-blind annotation.
• RP-Reasoner is described as a preliminary exploration with limited component analysis.

Experimental setup

Datasets: RPEval personalized intent reasoning dataset (8,255 generated samples)

Benchmarks: RPEval commercial personalized assistant bad cases

Metrics: intent classification accuracy Micro-acc Macro-acc overall error severity score quadratic-weighted Cohen's kappa

Baselines: Vanilla Reminder CoT CoT-SC Self-Refine PrefEval ImplexCONV LongMemEval MemBench

Builds on: Rational Speech Acts Approximate Bayesian Computation chain-of-thought self-refine

Closest papers by meaning

• Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.90)
• PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.88)
• PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.88)
• AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · similarity 0.88)
• Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction (AAAI 2026 · similarity 0.87)
• Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs (ICLR 2025 · similarity 0.87)