The Personalization Trap: How User Memory Alters Emotional Reasoning in LLMs

Key Findings

• — Identical emotional scenarios paired with different remembered user profiles produced systematically divergent emotional interpretations across 15 LLMs.
• — In several high-performing models, advantaged profiles received more accurate emotional interpretations; the benchmarks beat's read of the PDF records gaps such as 80.1% against 77.4% for Claude 3.7 Sonnet and 81.6% against 76.6% for DeepSeek-R1.
• — Significant disparities across demographic factors appeared in both emotion-understanding and supportive-recommendation tasks.
• — The authors conclude that personalization mechanisms can embed social hierarchies into models' emotional reasoning, so memory-enhanced assistants may reinforce inequality.

Methodology Notes

Evaluation of 15 models (Claude 3.5 to 4.5, DeepSeek-R1 and V3, Llama 3.x and 4, Mistral Large, Phi-4-mini, Qwen3, Gemma-2 per the beat's PDF read) on validated emotional-intelligence items (STEU and STEM) under varied synthetic user profiles, several thousand questions per experiment; no human participants. Published in the ACL 2026 Short Papers volume (July 2026), pages 511 to 529, DOI 10.18653/v1/2026.acl-short.43; all authors at Amazon. Verified at the ACL Anthology on 2026-09-15.

Sources