Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2607.02374v1

Read the source

Source

Plain-Language Brief

Researchers developed a framework called DRIFTLENS to measure how personalized language models change their reasoning process based on user attributes, preferences, and prior context. The study found that user-attribute memory induces medium-to-large reasoning drift in four different language models across various user categories. Two post-training methods were evaluated for reducing this drift, but neither uniformly dominated. The results indicate that memory-induced reasoning drift is a measurable failure mode of personalized language models.

What the Document Says

  • The study presents DRIFTLENS, a framework to quantify the effect of user-attribute memory on reasoning trajectory in personalized language models (LLMs).
  • DRIFTLENS 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.
  • The study validates that DRIFTLENS distinguishes content-free pragmatic noise from substantive reasoning changes.
  • User-attribute memory induces medium-to-large reasoning drift in four different LLMs across various user categories including age, occupation, and disability.
  • Final answers remain fluent, on-topic, and plausible even with reasoning drift induced by user-attribute memory.
  • Two post-training methods (GRPO and DPO) were evaluated for reducing drift, but neither uniformly dominates.
  • Effects on downstream capability, helpfulness, and instruction following are model-and reward-dependent.

Key Concepts Mentioned

  • personalized language models
  • DRIFTLENS
  • user-attribute memory
  • reasoning drift
  • pragmatic noise
  • GRPO
  • DPO

Claims Made by the Document

  • DRIFTLENS can quantify the effect of user-attribute memory on reasoning trajectory in personalized language models.
  • User-attribute memory induces medium-to-large reasoning drift in four different LLMs across various user categories.
  • Two post-training methods (GRPO and DPO) can reduce reasoning drift, but neither uniformly dominates.

Evidence Used

  • The text describes DRIFTLENS as a framework to quantify the effect of user-attribute memory on reasoning trajectory.
  • The results indicate medium-to-large reasoning drift in four different LLMs across various user categories.
  • The text mentions that both GRPO and DPO can reduce reasoning drift but neither uniformly dominates.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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 language models.