LLM Personalization Tests Overstate Effects by 2x

LLM Personalization Tests Overstate Effects by 2x — Analysis Headed to NeurIPS in Paris

A co-author of the paper “Not Your Diary, Just Your Topic: Grading the Wrong-User Control” (arXiv 2611.09609, PALM workshop at NeurIPS 2026) will present in Paris that standard LLM personalization checks via swapping user history for another’s overstate the effect: on LaMP-2, the honest gain from one’s own history is 2.3 p.p. instead of 4.8 p.p., and 86% of the “own vs. other history” gap is explained solely by category labels.

Compiled by Sergey KostenchukPublished 2026-11-09Updated 2026-11-10

2026-11-09 Research

🤖 LLM Personalization Tests Overstate Effects by 2x

A co-author of the paper “Not Your Diary, Just Your Topic: Grading the Wrong-User Control” (arXiv 2611.09609, PALM workshop NeurIPS 2026) is heading to Paris with it. The check via swapping history for another’s is confounded: both “personality” and topics change. On LaMP-2, one’s own history yields only +2.3 p.p., while the “remove history only at evaluation” trick inflates the gain to 4.8 p.p.

🌍 Honest evaluation requires unseen users and topic-matched controls: 86% of the “own vs. other history” gap is explained solely by category labels, and for unseen users, the advantage over similar history drops to a negligible 0.65 p.p.

👤 Building agents with memory? Take the checklist for correct evaluation and read others’ “better by N p.p.” figures with caution — part of the gain may just be topic matching. The 21-page paper can be read in advance or the authors can be found at NeurIPS.

Source 1: https://arxiv.org/abs/2611.09609

Sources