Source-linked AI summary
BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs
Sangyeon Yoon, Sunkyoung Kim, Hyesoo Hong, Wonje Jeung, Yongil Kim, Wooseok Seo, Heuiyeen Yeen, Albert No
TL;DR
Persistent-memory LLMs need to personalize responses without applying user preferences when formal communication contexts make them inappropriate. BenchPreS benchmarks this selective application using context–profile evaluations and MR/AAR metrics, finding that even frontier models often over-apply preferences and that reasoning or prompt defenses do not fully resolve the problem.
Problem
Existing personalization benchmarks largely assume user preferences should always be applied, leaving context-sensitive suppression insufficiently evaluated.
Method
BenchPreS pairs persistent-memory user profiles with recipient–task contexts and evaluates preference behavior using Misapplication Rate (MR) and Appropriate Application Rate (AAR).
Results
Even frontier models struggle with context-sensitive preference application: higher AAR consistently accompanies higher MR, indicating broad preference activation with limited contextual filtering.
Takeaways & Limitations
Current models tend to treat personalized preferences as broadly applicable instructions rather than context-dependent signals, so stronger preference adherence does not ensure selectivity.
Takeaways & Limitations
BenchPreS evaluates preference selectivity at the final generation stage and does not cover retrieval or external-tool settings or informal, socially nuanced communication.
Abstract
from arXiv · showhide
Large language models (LLMs) increasingly store user preferences in persistent memory to support personalization across interactions. However, in third-party communication settings governed by social and institutional norms, some user preferences may be inappropriate to apply. We introduce BenchPreS, which evaluates whether memory-based user preferences are appropriately applied or suppressed across communication contexts. Using two complementary metrics, Misapplication Rate (MR) and Appropriate Application Rate (AAR), we find even frontier LLMs struggle to apply preferences in a context-sensitive manner. Models with stronger preference adherence exhibit higher rates of over-application, and neither reasoning capability nor prompt-based defenses fully resolve this issue. These results suggest current LLMs treat personalized preferences as globally enforceable rules rather than as context-dependent normative signals.
1 Introduction
BenchPreS examines whether persistent-memory LLMs can apply user preferences selectively in formal communication contexts. It evaluates the balance between applying appropriate preferences and suppressing inappropriate ones, finding that models often apply preferences globally and that reasoning or prompt defenses do not fully solve the problem.
- Evaluation: Lower MR and higher AAR indicate stronger preference selectivity, with the ideal outcome at MR 0 and AAR 100.Figure 1 uses these two complementary metrics to characterize model behavior across contexts.
- Problem: Context-aware preference selectivity asks whether models apply appropriate persistent-memory preferences while suppressing inappropriate ones for the recipient and task.The paper contrasts everyday preferences such as humor or emojis with formal messages where those preferences may be unsuitable.
- Contribution: BenchPreS evaluates whether language models can distinguish when stored preferences should be applied or suppressed, unlike prior benchmarks that primarily assume preferences should always be followed.The benchmark targets selective personalization rather than preference adherence alone.
- Benchmark design: BenchPreS pairs user profiles with recipient–task contexts spanning 39 formal communication settings and evaluates responses using Misapplication Rate (MR) and Appropriate Application Rate (AAR).Profiles contain factual information and preference attributes, while each context specifies the social setting in which information is shared.
- Findings: Models often scale preference application globally rather than applying or suppressing preferences based on context.This behavior reflects limited selectivity in the benchmark’s formal communication settings.
2 Related Work
Related work has studied persistent-memory retrieval and personalized preference following, but these lines of research primarily focus on remembering or injecting relevant user information. BenchPreS addresses the separate question of whether preferences should be applied in a given context.
- Persistent Memory Systems in LLMs: Early persistent-memory systems selectively retrieve user records relevant to the current query instead of injecting all user information into the prompt.Later retrieval-augmented methods maintain separate memory stores and use retrievers to inject salient personalized information.
- Persistent Memory Systems in LLMs: Long-context LLMs have enabled simpler systems that prefix persistent memory as text, making relevant-information retrieval resemble a needle-in-a-haystack problem.These approaches also raise challenges in controlling how persistent memory is used.
- Personalization and Preference Following: Prior personalization benchmarks mainly measure whether models remember and reflect user-specific information through profile-conditioned generation or role-playing consistency.They generally emphasize preference following rather than deciding when a preference should be suppressed.
3 BenchPreS: Context-Aware Preference Selectivity in Persistent-Memory LLMs
BenchPreS evaluates whether persistent-memory LLMs can distinguish preferences that should be applied from those that should be suppressed across communication contexts. It formalizes this selectivity with context and user profiles, gold applicability labels, LLM-judged responses, and complementary application metrics.
- BenchPreS tests whether LLMs distinguish when persistent-memory preferences should be applied or suppressed across contexts without explicit instructions.
- Data Construction: The benchmark represents each context as a recipient–task pair and uses 39 formal communication contexts spanning five domains.
- Data Construction: It constructs 10 user profiles with persistent memories containing approximately 152 attributes, including five preference attributes per profile.
- Data Construction: Human annotators assign gold labels indicating whether each preference should be applied or suppressed, while excluding socially or culturally ambiguous cases.
- Evaluation Protocols: An LLM-as-Judge determines whether each preference appears in generated responses, producing 1,950 attribute-level evaluation instances.
- Evaluation Protocols: Misapplication Rate measures inappropriate application, whereas Appropriate Application Rate measures correct application of preferences that should be used.
4 Experiments
BenchPreS evaluations show that stronger preference application does not reliably produce selective behavior: models often apply preferences that should be suppressed. Reasoning and prompt-based mitigation alter this trade-off but do not fully eliminate misapplication.
- Main Results: Across 10 LLMs, higher AAR is consistently associated with higher MR, indicating stronger preference application does not translate into improved selectivity.
- Main Results: 88.69% AAR and 86.48% MR make Gemini 3 Pro the highest-scoring model on both appropriate and inappropriate preference application.
- Main Results: 38.49% MR and 49.77% AAR make Mistral 7B Instruct v0.3 the lowest on both metrics, suggesting reduced misapplication may reflect weaker preference application.
- Qualitative Examples: Qualitative cases show models applying playful, stylistic, or role-based preferences in formal communications where those preferences should be suppressed.
- Effect of Reasoning Capability: Reasoning variants increase AAR but also MR, alongside higher IFBench scores, consistent with stronger instruction following that increases overall preference execution.
- Effect of Prompt-Based Defense: Prompt mitigation reduces MR with small AAR decreases across frontier models, but effectiveness varies substantially by system and cannot fully resolve misapplication.
5 Additional Results
Additional analyses show that preference selectivity remains difficult across communication domains and preference categories, while task completeness and selective preference use can diverge.
- Communication Domains: Across five communication domains, MR remains substantial and stronger appropriate application is generally accompanied by higher misapplication.
- Preference Categories: Markers such as emojis and nicknames show consistently high MR across models, whereas GPT-5.2 has particularly low MR for role and style preferences.
- Task Completeness: Task completeness does not necessarily imply strong suppression of inappropriate preferences under personalization.
- Task Completeness: GPT-5.2 preserves task completeness and shows the strongest preference selectivity, while Gemini 3 Pro performs poorly on both.
6 Discussions
The analysis examines judge reliability, unresolved selectivity failures, and reasoning patterns associated with successfully suppressing inappropriate preferences.
- Judge validation: Pairwise agreement among DeepSeek-R1, GPT-5-mini, and a human annotator was high across 100 sampled instances.The sample uniformly covered gold labels g(t, a) = 0 and g(t, a) = 1.
- Future directions: Multi-turn reconfirmation may partially help, but it is poorly suited to automated agent deployments that generate responses without additional user intervention.The discussion identifies structural training signals as a direction for addressing this limitation.
- Reasoning patterns: Successful suppression follows a three-step pattern: enumerate stored preferences, assess contextual appropriateness, and exclude conflicting attributes before generation.The pattern was observed in reasoning traces from successful cases.
7 Conclusion
The conclusion presents BenchPreS as a benchmark for testing selective preference application under formal communication norms. Across models and settings, results indicate that stronger preference adherence is coupled with greater over-application, while current mitigations do not fundamentally resolve the problem.
- 7 Conclusion: BenchPreS evaluates whether persistent-memory LLMs selectively apply user preferences under formal communication norms.The benchmark spans diverse user profiles, contexts, and frontier LLMs.
- 7 Conclusion: Higher Appropriate Application Rate consistently accompanies higher Misapplication Rate, while lower Misapplication Rate tends to accompany lower Appropriate Application Rate.This pattern suggests models modulate overall preference strength rather than selectively suppress inappropriate preferences.
- 7 Conclusion: Neither reasoning capability nor prompt-based mitigation fundamentally resolves context-sensitive preference regulation.The conclusion frames BenchPreS as a diagnostic benchmark for this failure mode.
Limitations
BenchPreS evaluates preference selectivity at final generation and does not cover retrieval-based settings or the full complexity of informal, socially nuanced communication.
- Limitations: BenchPreS excludes settings relying on retrieval or other external tools and may not capture culturally or personally nuanced informal communication.Extending the benchmark to these settings remains future work.
A.1 Data Construction Protocol
The data-construction protocol isolates preference applicability in relatively stable formal communication settings, using filtered contexts, independently assigned human labels, and persistent-memory prompt representations.
- A.1 Data Construction Protocol: The benchmark is designed as a controlled evaluation rather than an exhaustive representation of real-world personalization.It isolates preference applicability where judgments can be made under relatively stable norms.
- A.1 Data Construction Protocol: 39 contexts were retained from 49 recipient–task candidates after excluding cases with substantial interpersonal, social, or cultural ambiguity.The retained contexts came from formal institution-facing and professionally constrained communication scenarios.
- A.1 Data Construction Protocol: Candidate preference instances covered both appropriate and inappropriate cases, while final gold labels were assigned independently by human annotation to prevent construction-process leakage.The candidate instances were not treated as gold labels.
- A.1 Data Construction Protocol: LLM-based labeling was rejected after preliminary experiments showed inconsistent judgments for context-dependent cases.Human annotators assigned g(t, a) = 1 for appropriate and helpful preferences, and g(t, a) = 0 for conflicts with norms, inappropriate tone, or task distraction.
- A.1 Data Construction Protocol: Only instances with unanimous agreement among three annotators were retained, reducing label ambiguity and improving annotation stability.Annotators did not see author-provided labels.
- A.1 Data Construction Protocol: Memory content is represented as natural-language statements prepended using the OpenAI memory template adopted from CIMemories.This setup follows the cited CIMemories representation procedure.
C Failure Analysis via Reasoning Traces
The reasoning traces show that models usually treat stored preferences as requirements rather than evaluating whether contextual norms require withholding them. Even inappropriate preferences are executed, suggesting that reasoning reinforces preference adherence instead of enabling selective application.
- Reasoning-trace failure: Models generally fail to evaluate whether preferences should be withheld under contextual norms, treating them as requirements during reasoning.The analysis describes preference application as requirement following rather than context-sensitive judgment.
- Reasoning-trace failure: Figure 9 shows the model converting preferences into checklist-style constraints and executing them even when they are inappropriate.The example highlights inappropriate preferences being followed as instructions.
- Reasoning-trace failure: The traces suggest reasoning reinforces treating preferences as instructions rather than enabling preference selectivity.This conclusion follows from the observed checklist-based execution of inappropriate preferences.
- Reasoning-trace failure: The example reasoning explicitly begins by checking key requirements, including a school-newsletter format and sarcastic or ironic tone.These requirements illustrate how user preferences enter the task-solving process as constraints.