BenchPreS: A Benchmark for Context-Aware Personalized Preference Selectivity of Persistent-Memory LLMs¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://huggingface.co/datasets/sangyon/BenchPreS
Area: Information Retrieval & RAG
Keywords: Persistent Memory, Personalized Preference Selectivity, Large Language Models, Contextual Suppression, Normative Alignment
TL;DR¶
BenchPreS is the first benchmark assessing the context-aware preference selectivity of persistent-memory LLMs across formal communication scenarios, revealing that frontier models indiscriminately treat stored user preferences as globally enforceable commands—achieving misapplication rates of 40% to 86% across institutional communication tasks and demonstrating a fundamental lack of context-sensitive normative suppression.
Background & Motivation¶
As long-context reasoning capabilities advance and persistent memory architectures become standard infrastructure in modern AI assistants, large language models increasingly store personalized user profiles, historical behavioral traits, and stylistic preferences to maintain coherence across extended interactions. In production environments, systems routinely inject these persistent memories as textual prefixes into input prompts, enabling models to retrieve and condition responses on user-specific attributes. However, when autonomous LLMs transition from conversational companions into active agents conducting external communications—such as drafting formal communications to regulatory agencies, legal mediators, admissions committees, or lending institutions—unregulated personalization introduces severe normative conflicts and pragmatic hazards.
Existing personalization and user modeling benchmarks, including LaMP, PrefEval, and FollowBench, suffer from a fundamental conceptual blind spot: they universally presume that whenever a user preference exists within memory, maximum adherence constitutes optimal behavior. Consequently, evaluation paradigms focus exclusively on positive preference following. Yet in formal, institutionally governed communication, directly applying user habits such as sarcastic humor, eccentric pseudonyms, unsolicited emoji markers, elementary teacher perspectives, or newsletter-style typography is not merely unprofessional—it actively compromises communicative objectives, impairs institutional credibility, and invites punitive legal or financial repercussions. A truly capable personalized assistant must not act as a naive echo chamber of user habits; it must possess context-aware selectivity to determine when a preference should be actively manifested and when it must be strictly suppressed.
The core tension underlying these systemic failures is that current post-training and alignment paradigms treat stored preferences as globally mandatory execution constraints, providing zero negative supervision for context-dependent withholding. Models excel at surface-level instruction following but lack the normative reasoning required to subordinate personal stylistic preferences to external institutional demands. To address this gap, BenchPreS formalizes the diagnostic challenge of context-aware preference selectivity in persistent-memory LLMs, examining whether models can selectively reflect appropriate preferences while suppressing inappropriate ones under strict communication norms. Core idea: formalize context-aware preference selectivity for persistent-memory LLMs, establishing a controlled diagnostic benchmark BenchPreS featuring 39 institutional recipient-task contexts and 10 multi-attribute synthetic profiles to decouple preference adherence from contextual suppression via Misapplication Rate (MR) and Appropriate Application Rate (AAR).
Method¶
Overall Architecture¶
BenchPreS provides a rigorous diagnostic framework to evaluate whether persistent-memory LLMs can balance personalization with communicative appropriateness when generating task-oriented messages for external recipients. The architecture encompasses a formal communication context space, multi-attribute persistent memory profiles, unanimous human gold annotations, and an LLM-as-Judge evaluation pipeline powered by DeepSeek-R1.
During evaluation, an LLM is presented with a prompt combining persistent memory statements (encompassing background facts and five core preference attributes) and a task instruction specifying the target recipient and purpose. The model's generated response is analyzed by an independent judge model to detect the presence of each preference attribute, which is then compared against human-validated gold labels to compute dual selectivity metrics.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input Formulation<br/>Persistent Memory + Context Prompt"] --> B["Context Space Construction<br/>5 Domains / 39 Recipient-Task Pairs"]
B --> C["Persistent Memory Modeling<br/>10 Profiles / 5 Preference Categories"]
C --> D["Gold Label Annotation<br/>Unanimous Agreement / Ambiguity Filtering"]
D --> E["Decoupled Selectivity Evaluation<br/>Misapplication Rate vs Appropriate Application Rate"]
E --> F["Selectivity Prompting & Reasoning Intervention<br/>Context-Aware Normative Reasoning Analysis"]
Key Designs¶
1. Context Space Construction: Standardizing Institution-Facing Communication Scenarios
To ensure unambiguous normative expectations, BenchPreS systematically refines candidate scenarios from CIMemories into 39 standardized recipient-task pairs (\(|T|=39\)) across five core formal domains: Finance, Legal/Taxation, Health, Education, Employment, and Housing. Each context \(t = (\text{recipient}, \text{task})\) represents an institutional interaction where communicative norms are codified and socially invariant—such as resolving tax discrepancies with an Internal Revenue Service (IRS) agent, submitting evidentiary records to a civil litigation attorney, or detailing housing history to a tenant screening service. Interpersonal contexts where appropriateness depends on subjective relationship dynamics (e.g., negotiating with an ex-partner) were deliberately excluded to maintain strict evaluative objectivity.
2. Persistent Memory Modeling: Decoupling Factual Knowledge from Multi-Tier Preferences
To faithfully simulate production-grade persistent memory architectures, BenchPreS constructs 10 synthetic user profiles (\(|U|=10\)), each containing an average of 152 natural language memory records prepended using the official OpenAI memory template format. Each profile combines extensive factual task-solving information (e.g., employment history, financial assets, health records) with exactly \(k=5\) preference attributes \(A_{\text{pref}}^u\). These preferences span five orthogonal categories aligned with modern commercial personalization interfaces: Role (e.g., speaking from an elementary school teacher or stand-up comedian persona), Style (e.g., formatting documents as school newsletters), Tone (e.g., sarcastic or ironic register), Marker (e.g., inserting emojis or bolding text for emphasis), and Nickname (e.g., requesting to be addressed as "Joker" or "Rambo").
3. Gold Label Annotation: Strict Unanimous Agreement Filtering
To avoid noise and subjectivity in normative ground truth, ground truth labels \(g(t, a) \in \{0, 1\}\)—indicating whether preference \(a\) is appropriate to apply (\(g=1\)) or must be withheld (\(g=0\)) in context \(t\)—were assigned by three independent expert annotators spanning diverse age groups and cultural backgrounds. The protocol enforced a 100% unanimous agreement threshold; candidate pairs exhibiting any interpretive discrepancy (e.g., whether subtle bolding enhances readability or violates formal register in specific memos) were systematically pruned. The annotators reached initial consensus on 94.3% of candidates, with the 5.7% borderline instances excised to yield an uncontroversial evaluation corpus of 1,950 attribute-level test instances.
4. Decoupled Selectivity Evaluation: Orthogonal MR and AAR Metric Matrix
Recognizing that conventional adherence metrics mask over-personalization failure modes, BenchPreS introduces two complementary metrics evaluated via DeepSeek-R1 as an automated LLM-as-Judge (validated against human judgment with 88.67% agreement and a Cohen's \(\kappa\) of 0.773). Given generated response \(y_{u,t}\), the judge outputs binary decision \(\hat{z}(y_{u,t}, a) \in \{0, 1\}\) denoting whether preference \(a\) is manifested. Misapplication Rate (MR) measures the fraction of normatively forbidden preferences that are erroneously applied:
Appropriate Application Rate (AAR) quantifies the proportion of contextually warranted preferences that are successfully applied:
An optimal system achieves \(\text{MR} \to 0\%\) and \(\text{AAR} \to 100\%\). The difference metric \((\text{AAR} - \text{MR})\) provides a direct summary measure of preference discrimination; models near or below zero simply treat preferences as indiscriminate operational commands.
5. Selectivity Prompting & Reasoning Intervention: Uncovering Execution vs. Suppression Dynamics
To probe whether explicit guidance mitigates indiscriminate application, the benchmark assesses both selectivity-oriented system prompting (instructing models to apply task-appropriate preferences while actively suppressing norm-violating traits) and explicit reasoning chain dynamics. Tracing reasoning behavior reveals that unguided reasoning models treat user preferences as obligatory checklist constraints, thereby exacerbating misapplication. In contrast, successful suppression requires structural reasoning: enumerating memories, auditing contextual appropriateness against institutional norms, explicitly flagging conflicting preferences for exclusion, and only then generating the final message.
Key Experimental Results¶
Main Results¶
Evaluations were conducted across 10 prominent commercial and open-weight LLMs spanning diverse parameter scales and reasoning variants. Responses were generated at temperature 1.0 across three runs per user-context pair. Preference recall checks confirmed that all 10 models achieve 100% attribute recall from persistent memory, verifying that failures stem entirely from context-dependent selectivity rather than retrieval degradation.
| Model Category / Model Name | Parameters | Misapplication Rate (MR) ↓ | Appropriate Application Rate (AAR) ↑ | Selectivity Gap (AAR - MR) ↑ |
|---|---|---|---|---|
| Frontier Frontier Models (≥ 500B) | ||||
| Gemini 3 Pro | Frontier API | 86.48% ±0.72 | 88.69% ±0.58 | +2.21 |
| DeepSeek V3.2 | 671B (MoE) | 61.15% ±1.14 | 87.63% ±1.26 | +26.48 |
| Claude-4.5 Sonnet | Frontier API | 52.99% ±1.37 | 87.93% ±0.81 | +34.94 |
| GPT-5.2 | Frontier API | 40.95% ±2.54 | 87.33% ±0.34 | +46.38 |
| Qwen3 235B A22B Thinking 2507 | 235B (MoE) | 84.62% ±1.06 | 82.85% ±1.44 | -1.77 |
| Mid-Scale & Open Baselines (< 500B) | ||||
| gpt-oss-120b | 120B | 74.28% ±0.63 | 87.40% ±0.89 | +13.12 |
| Qwen-3 32B* | 32B | 68.22% ±1.52 | 81.45% ±1.18 | +13.23 |
| K-EXAONE-236B-A23B | 236B | 60.30% ±1.09 | 86.12% ±0.97 | +25.82 |
| Llama-3.3 70B Instruct* | 70B | 58.04% ±1.84 | 68.78% ±1.62 | +10.74 |
| Mistral 7B Instruct v0.3* | 7B | 38.49% ±2.11 | 49.77% ±1.95 | +11.28 |
Note: Asterisk () denotes standard instruction-tuned models without explicit reasoning architectures.*
Ablation Study: Selectivity-Oriented Prompting (Mitigation) and Task Completeness Impact¶
The following table summarizes the impact of selectivity-oriented mitigation prompting on reducing MR and preserving AAR across leading models, alongside task completeness evaluations on a 1–5 scale with and without persistent memory preferences:
| Model Name | Prompt Strategy | Misapplication Rate (MR) ↓ | Appropriate Application Rate (AAR) ↑ | Task Score (w/o Prefs) ↑ | Task Score (w/ Prefs) ↑ | Task Score Diff (Δ) |
|---|---|---|---|---|---|---|
| Gemini 3 Pro | Default Prompt | 86.48% | 88.69% | 4.855 | 3.746 | -1.109 |
| Mitigation Prompt | 12.80% (-73.68 pp) | 84.87% (-3.82 pp) | — | — | — | |
| DeepSeek V3.2 | Default Prompt | 61.15% | 87.63% | 4.734 | 4.734 | 0.000 |
| Mitigation Prompt | 40.68% (-20.47 pp) | 84.84% (-2.79 pp) | — | — | — | |
| Claude-4.5 Sonnet | Default Prompt | 52.99% | 87.93% | 4.637 | 4.183 | -0.454 |
| Mitigation Prompt | 14.04% (-38.95 pp) | 84.75% (-3.18 pp) | — | — | — | |
| GPT-5.2 | Default Prompt | 40.95% | 87.33% | 4.925 | 4.957 | +0.032 |
| Mitigation Prompt | 21.52% (-19.43 pp) | 86.55% (-0.78 pp) | — | — | — |
Key Findings¶
- Pathological Coupling of Adherence and Misapplication: Across models, performance clusters along the diagonal \(y = x\). High preference adherence (e.g., Gemini 3 Pro reaching 88.69% AAR) directly translates into disastrous misapplication (86.48% MR). Conversely, models exhibiting lower MR, such as Mistral 7B (38.49%), do so due to weak instruction following (49.77% AAR) rather than contextual discernment. Current LLMs merely scale global preference responsiveness rather than exercising contextual selectivity.
- Reasoning Enhances Instruction Following at the Expense of Selectivity: Comparing Instruct and Thinking variants of Qwen3 235B reveals that enabling explicit reasoning improves general instruction-following scores (IFBench) while simultaneously inflating both AAR and MR, leading to a negative selectivity gap (-1.77). In reasoning traces, models decompose user preferences into rigid subgoals to be executed, blind to normative friction.
- Surface Markers and Nicknames are Harder to Suppress than Roles: Category-wise analysis indicates that models such as GPT-5.2 show moderate suppression for Role and Style attributes, but all models suffer elevated MR on Markers (e.g., emojis) and Nicknames. LLMs treat lexical markers and pseudonyms as cosmetic syntax rather than contextual pragmatic violations.
- Indiscriminate Personalization Severely Degrades Task Completeness: Injecting inappropriate preferences into memory causes Gemini 3 Pro's task fulfillment score to plunge from 4.855 to 3.746 (-1.109). In prioritizing comic personas or newsletter formatting, the model omitted critical financial and legal information necessary to achieve the task.
Highlights & Insights¶
- Reframing Personalization Alignment: Demonstrates that personalization cannot simply be equated with maximizing preference adherence; negative contextual regulation and selective withholding are equally essential pillars of safety and alignment for autonomous agents.
- The Paradox of Unconstrained CoT Reasoning: Shows that unguided chain-of-thought reasoning can paradoxically aggravate compliance failures by framing contextual preferences as mandatory checklist constraints, explaining why scaling inference-time reasoning alone does not solve social norm alignment.
- Structural Reasoning Patterns for Post-Training: Pinpoints the explicit cognitive trajectory found in successful runs (enumeration \(\to\) recipient norm evaluation \(\to\) explicit exclusion \(\to\) generation), providing an actionable recipe for synthesizing SFT data and designing reward functions for RL post-training.
Limitations & Future Work¶
- Constrained to Final Generation with Static Memory: BenchPreS evaluates end-to-end generation where preferences are prepended as text, leaving interactive multi-turn disambiguation and upstream retrieval filtering unaddressed.
- Focus on Strict Institutional Norms: To ensure unanimous gold-standard labels, the benchmark focuses on high-stakes institutional interactions, omitting ambiguous interpersonal scenarios where appropriateness depends on subjective cultural and interpersonal interpretations.
- Future Directions: Extending the framework to multi-agent interactive preference negotiation and integrating context-aware negative preference rewards into reinforcement learning from human feedback (RLHF) pipelines.
Related Work & Insights¶
- vs LaMP / PrefEval: Prior benchmarks measure how accurately models recall and reflect user history across multi-turn interactions, viewing preference manifestation strictly as a virtue; BenchPreS introduces contextual suppression as a necessary counterpart, exposing severe over-application blind spots.
- vs CIMemories / AgentDAM: CIMemories and AgentDAM study the unintended leakage of private or sensitive personal records; BenchPreS investigates benign, user-authorized stylistic and pragmatic preferences that turn toxic due to contextual normative mismatch.
- vs FollowBench / AdvancedIF: Instruction-following benchmarks evaluate whether models comply with explicit negative constraints provided in the prompt; BenchPreS evaluates tacit normative awareness where the prompt does not explicitly forbid the preference, testing whether models can autonomously infer social appropriateness.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ [Pioneers the formal conceptualization and benchmark design for context-aware preference selectivity in persistent-memory LLMs]
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ [Evaluates 10 frontier models, incorporating fine-grained preference taxonomies, prompt ablations, reasoning trace analyses, and task efficacy impacts]
- Writing Quality: ⭐⭐⭐⭐⭐ [Exceptionally clear conceptualization, rigorous experimental exposition, and dense analytical insight]
- Value: ⭐⭐⭐⭐⭐ [Crucial diagnostic benchmark for memory-augmented LLMs, providing foundational guidance for the safe deployment of autonomous personalized agents]