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

Problem Statement

Modern Large Language Models (LLMs) increasingly employ persistent memory to store user attributes, preferences, and prior context, injecting this information into future prompts to personalize interactions. While personalization aims to adapt responses to user needs, this paper identifies a critical, often invisible failure mode: symbolic drift.

Symbolic drift occurs when a model's reasoning trajectory—the chain of priorities, trade-offs, and justifications used to derive an answer—shifts significantly due to injected user attributes, even when the final answer remains fluent, on-topic, and plausible. This is particularly dangerous in "unverifiable" tasks (e.g., career advice, interpersonal conflicts, ethical trade-offs) where no single ground-truth answer exists. In these scenarios, standard accuracy metrics fail to detect that the model has shifted from a legal-risk reasoning framework to an emotional-validation framework simply because the user's profile indicates they are young, disabled, or in a stigmatized occupation. The paper argues that such shifts represent unwanted sensitivity rather than appropriate adaptation, yet they remain undetected by current evaluation methods.

Methodology: The DRIFTLENS Framework

To address the lack of ground-truth answers in reasoning tasks, the authors introduce DRIFTLENS, a ground-truth-free framework designed to measure the stability of reasoning trajectories under controlled memory perturbations.

1. Data Curation

The authors constructed a benchmark of 422 high-agreement questions (and a relaxed set of 1,061) satisfying four criteria:

• Reasoning-invoking: Requires trade-offs, comparisons, or justification.
• Unverifiable: Admits no single objective ground truth.
• Persona-indifferent: The answer should not logically depend on the asker's demographic attributes.
• Standalone: Comprehensible without hidden context.
  Questions were filtered from seven public sources (career, ethics, finance, legal, medical, daily dilemmas) using LLM raters and human annotators to ensure orthogonality between the question content and user attributes.

2. Value Ontology

To quantify reasoning changes, the authors developed a value ontology that maps free-text reasoning steps to discrete symbolic categories. Starting from existing hierarchies, they refined the ontology through 78 rounds of labeling by three LLMs (GPT-OSS-120B, Qwen3-235B, Claude Sonnet 4.5) to achieve >83% cross-model agreement. The ontology consists of four root categories:

• Social: Interpersonal/collective conduct.
• Practical: Operational task execution.
• Protective: Harm prevention and safety.
• Personal: Subjective inner experience.

3. Perturbation Strategy

The framework evaluates drift by comparing a no-memory baseline against responses generated under three conditions:

• Negative Control (Pragmatic Noise): Content-free prefixes (filler, whitespace, punctuation) to establish a noise floor.
• Positive Control (Major Life Events): Disclosures of significant psychological events (e.g., bereavement, serious illness) known to reshape values.
• Experimental Condition (Persona Cues): Specific user-attribute disclosures (gender, age, occupation, disability, etc.) constructed to be independent of the question.

4. Drift Metrics

Two complementary metrics measure the divergence between the baseline symbol sequence ($S_{base}$) and the intervened sequence ($S_{inter}$):

• DTW (Dynamic Time Warping): Captures structural reordering of reasoning steps.
• SRI (Sequence Recurrence Index): Combines an order-sensitive edit distance with a distributional Jensen–Shannon distance over symbol histograms.
  Both metrics are normalized to $[0, 1]$, where lower values indicate more stable reasoning.

Key Contributions

1. DRIFTLENS Framework: A novel, ground-truth-free methodology for measuring reasoning stability under memory perturbations using ontology-based trajectories.
2. Benchmark Construction: A curated dataset of 422 persona-indifferent, unverifiable questions, validated through rigorous human and LLM annotation.
3. Empirical Evidence of Drift: Demonstration that irrelevant persona memory induces medium-to-large reasoning drift across four distinct LLMs (Claude Sonnet 4.6, GPT-OSS-120B, Qwen3-4B, DeepSeek-R1) and 10 user-attribute categories.
4. Mitigation Evaluation: A comparative analysis of post-training methods (GRPO and DPO) for reducing drift, highlighting trade-offs between reasoning stability, helpfulness, and capability.

Results

Validation (RQ1)

DRIFTLENS successfully distinguishes between noise and substantive change. Pragmatic noise produced no significant elevation above the noise floor ($p > 0.05$), while major life events produced large, significant drifts ($p < 0.001$). The two metrics (DTW and SRI) showed high convergent validity (Spearman $\rho > 0.93$).

Memory-Induced Drift (RQ2)

User-attribute memory consistently induced measurable reasoning drift across all tested models.

• Magnitude: Effect sizes ranged from medium to large (Cohen's $d \approx 0.35–0.98$), significantly exceeding the pragmatic noise floor.
• Attributes: Drift was observed across all 10 categories, with Trans status and Disability ranking among the top drift-inducing categories on seven of eight model-metric panels.
• Nature of Drift: While final answers often remained plausible, human studies indicated reduced helpfulness scores and higher distraction rates. The drift primarily manifested in the reasoning trace (e.g., shifting from "legal-risk" to "emotional-validation" reasoning).

Mitigation (RQ3)

The authors evaluated GRPO (incorporating DTW into the reward function) and DPO (using preference pairs of robust vs. distracted responses).

• Drift Reduction: Both methods reduced drift (lower DTW scores) across models. No single method uniformly dominated; performance was model- and reward-dependent.
• Trade-offs:
  • Capability: Post-training did not consistently degrade general capabilities (MMLU, GSM8K), though effects varied by backbone.
  • Helpfulness & Robustness: DPO improved non-distraction accuracy reliably across all backbones without degrading helpfulness. GRPO improved helpfulness on some models but reduced non-distraction accuracy on others (e.g., a 15.4-point drop on Phi-4-mini), suggesting it may over-optimize for perceived helpfulness at the cost of robustness.
  • Format Rewards: Including format rewards in GRPO was critical for drift reduction; drift-only rewards were insufficient.

Significance and Claims

The paper concludes that memory-induced reasoning drift is a measurable and only partly mitigated failure mode of personalized language models.

• Invisibility: The phenomenon is dangerous because it is invisible at the answer level; models can sound reasonable while relying on fundamentally different chains of justification driven by irrelevant user attributes.
• Evaluation Necessity: Standard accuracy metrics are insufficient for unverifiable tasks. Reliability must be measured through reasoning stability.
• Mitigation Limits: While post-training (GRPO/DPO) can reduce drift, it introduces trade-offs. There is no "free lunch" where drift is eliminated without potential impacts on capability, helpfulness, or instruction following.
• Future Direction: The authors emphasize the need to evaluate reasoning stability before deploying memory-equipped LLMs in decision-support settings. They acknowledge limitations, noting that DRIFTLENS measures externally expressed reasoning, not hidden cognition, and that their mitigation experiments were limited to small/medium models due to computational constraints.

The study serves as a call to action for the community to move beyond answer-level evaluation and develop robust mechanisms to ensure that personalization enhances, rather than distorts, the reasoning processes of AI systems.