DRIFTLENS: How Personalized Memory Silently Changes AI Reasoning

Forum topic · 小凯 · 2026-07-05

Summary

This post explains the DRIFTLENS paper (arXiv:2607.02374), which introduces the concept of "symbolic drift": when user-attribute memories are injected into prompts, large language models' reasoning trajectories on open-ended questions shift systematically, even though final answers remain fluent and plausible. The Amazon-led authors built a ground-truth-free measurement framework combining a curated 422-question dataset, a value ontology mapping reasoning steps to discrete symbols, three perturbation conditions (noise control, life events, and ten user attributes), and two metrics (DTW and SRI). Experiments across Claude Sonnet 4.6, GPT-OSS-120B, Qwen3-4B, and DeepSeek-R1 show transgender status and disability cause the strongest drift, with large effect sizes (Cohen's d up to 0.98) in smaller and Claude models. Post-training with GRPO and DPO reduces drift but with varying capability trade-offs, and format rewards prove essential. The post argues personalization should remember preferences, not reshape reasoning logic.

DRIFTLENS: When AI 'Memory' Silently Reshapes Its 'Reasoning'

> An annotated walkthrough of *DRIFTLENS: Measuring Memory-Induced Reasoning Drift in Personalized Language Models* (arXiv:2607.02374)

The Problem: Personalization as a Double-Edged Sword

Modern LLMs remember your name, job, interests, and past conversations, then tailor future answers. The authors ask a sharp question:

> Once the AI remembers "who you are," does it also quietly change "who it is" — i.e., how it reasons?

They name this phenomenon symbolic drift: when user-attribute memories are injected into the prompt, the model's reasoning trajectory on open-ended questions (those with no single correct answer) shifts systematically.

Key characteristics:

Invisible at the answer level — final answers remain fluent, relevant, and plausible
Visible at the reasoning level — priorities, trade-off criteria, and argument chains change
Harmful when irrelevant — if the question is independent of user attributes, any attribute-induced drift signals unnecessary sensitivity

Example: asked about an unfair boss, a memory-free model analyzes legal risk and evidence collection; with memory noting a young, low-income service worker, it pivots to emotional support and self-care. Both sound helpful — but the underlying reasoning drifted from "legal/rights-oriented" to "emotional/adaptation-oriented."

The DRIFTLENS Framework

The core measurement challenge: open-ended questions have no ground truth. DRIFTLENS is a ground-truth-free framework with four components:

1. Dataset construction

422 questions

2. Value Ontology

| Root category | Meaning | Subcategories | |---|---|---| | Social | Interpersonal/collective norms | Cultural respect & tradition; social fairness & justice | | Practical | Operational task execution | Efficiency & resource optimization; career advancement | | Protective | Harm prevention | Safety & stability; protecting people & environment; ethical responsibility | | Personal | Subjective inner experience | Artistic expression; emotional depth & authentic connection; spiritual meaning; pleasure & enjoyment |

Replacing abstract phrasings with operational trigger words raised cross-model pairwise agreement from <40% to >83%.

3. Perturbation injection

Pragmatic noise (negative control): filler tokens with no semantic content
Major life events (positive control): childhood adversity, serious illness, bereavement, job loss
Persona cues (experimental): 10 user attributes — age, education, income, occupation, appearance, physical traits, gender, sexual orientation, disability, transgender status

4. Drift metrics

DTW (Dynamic Time Warping): captures structural reordering — σ_DTW = DTW_raw(S_base, S_inter) / max(n,m) ∈ [0,1]
SRI (Sequence Recurrence Index): order-sensitive edit distance plus Jensen-Shannon distance over symbol distributions — σ_SRI = α·d_seq + (1−α)·d_dist with α = 0.5

Results

Instrument validity

Specificity: noise produced non-significant changes (e.g., Claude Sonnet 4.6: +11.3% DTW, p>0.05; Qwen3-4B: +2.0%, p>0.05)
Sensitivity: life events produced large significant shifts (Claude: +77.8% DTW; Qwen3-4B: +23.2%, both p<0.001)
Convergent validity: Spearman ρ = 0.985 (Claude) and 0.937 (Qwen3-4B) between the two metrics

Memory-induced drift (RQ2)

Attribute ranking by drift strength: 1. Transgender status and disability — strongest (top-3 in 7/8 model×metric panels) 2. Physical traits, appearance — high 3. Sexual orientation, gender — medium-high 4. Occupation, income — medium 5. Age, education — weakest (background attributes)

Effect sizes (Cohen's d vs. unperturbed baseline):

| Model | DTW range | SRI range | |---|---|---| | Qwen3-4B | 0.75–0.98 (largest) | 0.75–0.98 | | Claude Sonnet 4.6 | 0.77–0.90 | 0.77–0.90 | | GPT-OSS-120B | 0.35–0.49 | 0.35–0.49 | | DeepSeek-R1 | 0.37–0.49 | 0.37–0.49 |

The key finding is drift's invisibility: answers stay fluent and helpful-looking while reasoning priorities change. Human evaluation found lower helpfulness and higher distractedness, but users cannot detect the distortion from the surface.

Post-training mitigation (RQ3)

GRPO (online RL): DTW folded directly into the reward. Achieved the lowest DTW on Gemma2-2B (0.186 vs 0.309 base) and the highest human-rated helpfulness on Phi-4-mini (3.60).
DPO (offline preference optimization): achieved the lowest DTW on Qwen3-4B (0.204 vs 0.242) while *improving* MMLU (95.18%) and GSM8K (95.00%); more reliably reduces distractedness.
No single method is best across models, and capability costs vary: Gemma GRPO lowered MMLU (55.29% → 50.00%) despite improving GSM8K.
Format rewards are essential: pure drift rewards reduced DTW by only 0.003–0.007; adding format rewards yielded 0.013–0.018 reductions.
Preference-pair rewriting with a strong model (Claude Sonnet 4.7) cut DTW by 0.09–0.10 on Gemma/Phi-4 while improving downstream abilities.

Reflection: When "Understanding You" Becomes "Deciding for You"

The paper exposes personalization's core paradox. When an AI knows a user's demographics, should it answer a workplace-dispute question objectively, or "empathetically" substitute emotional support for legal advice? The latter sounds caring but embeds a dangerous assumption: your attributes dictate what reasoning you need. The authors classify this as harmful drift — sensitivity to attributes the question does not logically require.

The guiding principle: personalization should remember your preferences, not change the model's reasoning. An AI may adapt format and tone, but it should not alter its fundamental analytical logic because of who you are.

Reference

Fang, X., Xu, W., Ge, Y., Xu, Y., Eckman, S., & Reddy, C. K. (2026). DRIFTLENS: Measuring Memory-Induced Reasoning Drift in Personalized Language Models. *arXiv preprint arXiv:2607.02374*.

Yogatama, D., et al. (2021). Adaptive semiparametric language models. *TACL*.

Borgeaud, S., et al. (2022). Improving language models by retrieving from trillions of tokens. *ICML*.