DRIFTLENS: Measuring Memory-Induced Reasoning Drift in Personalized Language Models

Abstract: Personalization changes what a model says to a user; we show that it can also change the reasoning trajectory used to justify the response. Modern LLMs personalize interactions by storing user attributes, preferences, and prior context, then injecting this information into future prompts. We study whether such memory reshapes reasoning on open-ended questions where no single ground-truth answer exists. To quantify this effect, we introduce DRIFTLENS, a ground-truth-free framework that maps each expressed reasoning step to a value category and measures divergence between a question's no-memory trajectory and its trajectory under injected user-attribute memory. We first validate that DRIFTLENS distinguishes content-free pragmatic noise from substantive reasoning changes. Across four LLMs and 10 user-attribute categories, including age, occupation, and disability, user-attribute memory induces medium-to-large reasoning drift above each model's pragmatic-noise floor, even when final answers remain fluent, on-topic, and plausible. We then evaluate GRPO- and DPO-based post-training methods for reducing drift. Both reduce drift, but neither uniformly dominates; effects on downstream capability, helpfulness, and instruction following are model-and reward-dependent. These results suggest that memory-induced reasoning drift is a measurable and only partly mitigated failure mode of personalized LLMs.

• The paper introduces DRIFTLENS, which quantifies how user memory impacts reasoning trajectories in LLMs using ontology mapping and divergence-based metrics.
• The framework employs controlled perturbations and dual metrics (DTW and SRI) to detect invisible shifts in model justifications and value trade-offs.
• Post-training methods like GRPO and DPO partially mitigate symbolic drift while preserving overall model performance on standard benchmarks.

Measuring Memory-Induced Reasoning Drift in Personalized LLMs: An Analysis of DRIFTLENS

Introduction

This work introduces DRIFTLENS, a framework for quantifying the impact of user-attribute memory on reasoning trajectories in personalized LLMs (2607.02374). The key assertion is that personalization in LLMs does not only alter output content but also the underlying reasoning process, even when the user's attributes should be logically irrelevant to the prompt. This "symbolic drift" manifests as shifts in the model's priorities, justifications, and value trade-offs, which are not detectable using standard answer-level metrics—especially in unverifiable, open-ended tasks that lack objective ground-truths. DRIFTLENS provides a methodology for surface-level detection and quantification of such drift, regardless of final answer plausibility.

DRIFTLENS Framework and Methodology

The DRIFTLENS framework is ground-truth-free and relies on ontology-based mapping of reasoning steps, paired with divergence-based metrics:

• Benchmark Construction: The authors curate a suite of 422 "persona-indifferent," unverifiable guidance questions, all filtered to ensure that reasonable answers should not logically depend on demographic or psychological attributes of the user.
• Ontology Development: A multi-level value ontology is developed for reasoning step labeling, with high cross-model and within-model agreement rates (≥83%), operationalizing the mapping of each free-text reasoning step into symbolic categories (Social, Practical, Protective, Personal, and 11 subcategories).
• Perturbation Design: Three types of prompt perturbations simulate context changes:
  • Negative controls (pragmatic noise: filler words, whitespace, punctuation) to estimate a "noise floor".
  • Persona cues (user-attribute self-descriptions across 10 categories, e.g., disability, occupation, age, trans status), designed to be extraneous to the question.
  • Positive controls (major life event disclosures) that are known to reshape values.
• Drift Metrics: The framework applies two complementary metrics to measure divergence in reasoning sequences relative to baseline (no-memory) responses:
  • Dynamic Time Warping (DTW): Quantifies structural/order differences in reasoning sequences.
  • Sequence Recurrence Index (SRI): Integrates order-sensitive edit distance and distributional divergence (Jensen-Shannon) over symbolic labels.

Key Findings

Instrument Validity of DRIFTLENS

DRIFTLENS is shown to be sensitive and specific:

• Negative controls (pragmatic noise) do not significantly increase drift over the base level, confirming metric specificity.
• Positive controls (major life events) cause large, significant drift increases—validating metric sensitivity to meaningful context changes.

Both DTW and SRI consistently rank perturbation conditions, with high Spearman ($\rho > 0.93$) and Pearson correlation across metrics and models.

Memory-Induced Reasoning Drift Across Models

Persona cues produce consistent, statistically significant reasoning drift across four open LLMs (Claude Sonnet 4.6, Qwen3-4B, GPT-OSS-120B, DeepSeek-R1):

• For every tested persona category on every model, drift is elevated far above the noise floor for both DTW and SRI, with medium to large effect sizes (Cohen's $d \sim 0.35-0.98$).
• Categories such as trans status, disability, physical traits, and appearance consistently induce the highest drift across metrics and models. Socioeconomic status, occupation, and age also show pronounced, though somewhat lower, effects.
• Final answers remain plausible and fluent, but underlying value trade-offs and justifications shift—a critical, invisible failure mode in reasoning stability for decision-support settings.
• Human evaluation confirms a corresponding decrease in answer helpfulness and increase in distraction, despite answer plausibility.

Mitigation via Post-Training

The study evaluates two post-training mitigation approaches:

• Group Relative Policy Optimization (GRPO): An online RL variant that directly rewards meta-stability of reasoning steps.
• Direct Preference Optimization (DPO): An offline method leveraging preference pairs for perturbation-robust answers, augmented via style-normalizing rewrites for consistency.

Key results:

• Both GRPO and DPO reduce symbolic drift (DTW) across Gemma2-2B, Phi-4-mini, and Qwen3-4B models—best-case reductions in DTW up to 0.12 absolute.
• No single method dominates across all models; the trade-off between stability, helpfulness, and general capability is model- and reward-variant dependent.
• Downstream performance on MMLU, GSM8K, and IFEval remains comparable to base models post-mitigation; prominent drift reductions do not consistently degrade general capabilities.
• DPO more reliably decreases distraction rates according to human labeling, while GRPO can sometimes over-optimize for helpfulness at the expense of robustness.
• Ablation shows that format rewards and high-quality, style-consistent preference pairs further improve drift reduction, especially in weaker models.

Implications and Limitations

Theoretical implications: The study demonstrates that personalization in LLMs causes quantifiable, persistent alterations to the value structure of reasoning, even on persona-orthogonal instructions. This exposes a new axis of model instability (symbolic drift), not previously detectable by answer-level metrics or current evaluation protocols.

Practical implications:

• Any system that persists user attributes for improved personalization—especially in high-stakes, unverifiable settings (e.g., career, health, legal guidance)—requires explicit reasoning stability audits to ensure value-neutrality where appropriate.
• Drift can only be partially mitigated with post-training; neither reward shaping nor preference optimization fully immunizes models against unwanted attribute sensitivity.
• Selection or amplification of training data (e.g., more standard preference pairs) can improve instruction following, but may also increase symbolic drift in stronger models—flagging a critical capacity-alignment tradeoff for practitioners to monitor.

Limitations:

• DRIFTLENS evaluates only externally expressed reasoning, not latent cognition or internal decoder states.
• The fixed ontology may not capture all contextually- or culturally-granular value distinctions.
• Experiments are limited to small/medium open models due to compute constraints; it is unknown if results will generalize to frontier-scale LLMs.
• The framework currently addresses single-turn, open-ended tasks; extension to long-horizon dialogue and agentic systems is an open challenge.

Conclusion

DRIFTLENS provides systematic, ground-truth-free measurement of reasoning drift in memory-equipped LLMs and shows that persisted user attributes fundamentally reshape reasoning trajectories, even in the absence of answer-level changes. Neither policy optimization nor preference-based post-training can fully eliminate this drift; mitigation is only partial and remains entangled with key downstream tradeoffs. These results emphasize the necessity of value-stable reasoning as a core evaluation objective for trustworthy personalized AI, with further work needed to extend symbolic stability metrics and mitigation to larger models and more complex interactive contexts.