Dual-Path Parametric Memory for Personalized Language Models
October 9, 2026
The DPPM framework uses an Evidence path to pool interaction history and a Delta path to capture sequential changes, producing history-conditioned LoRA adapters. The method achieves 54.22% on PersonaMem-v2, effectively combining long-term preference preservation with recent state updates.
HOW THIS AFFECTS YOU
●
builderYou can implement more stable user personalization using dual-path adapters rather than expanding context windows.
●
researcherThis approach provides a more robust way to handle the stability-plasticity dilemma in personalized LLMs.