DRIFTLENS measures memory-induced reasoning drift in personalized LLMs

Researchers introduce DRIFTLENS, a ground-truth-free framework that quantifies how user-attribute memory reshapes the reasoning trajectories of large language models. The study validates that this framework distinguishes substantive reasoning changes from pragmatic noise across four LLMs and ten user-attribute categories.

• User-attribute memory induces medium-to-large reasoning drift above the pragmatic-noise floor, even when final answers remain fluent and plausible.
• GRPO- and DPO-based post-training methods reduce this drift, but neither uniformly dominates in effectiveness.
• The impact of mitigation strategies on downstream capability, helpfulness, and instruction following is model- and reward-dependent.

The results indicate that memory-induced reasoning drift is a measurable failure mode of personalized language models that is only partly mitigated by current post-training techniques.