This paper was accepted on the Workshop on Massive Language Mannequin Memorization (L2M2) 2025.
Massive Language Fashions (LLMs) have shortly turn into a useful assistant for a wide range of duties. Nevertheless, their effectiveness is constrained by their means to tailor responses to human preferences and behaviors through personalization. Prior work in LLM personalization has largely targeted on fashion switch or incorporating small factoids in regards to the consumer, as data injection stays an open problem. On this paper, we discover injecting data of prior conversations into LLMs to allow future work on much less redundant, customized conversations. We determine two real-world constraints: (1) conversations are sequential in time and have to be handled as such throughout coaching, and (2) per-user personalization is simply viable in parameter-efficient settings. To this goal, we suggest PLUM, a pipeline performing information augmentation for up-sampling conversations as question-answer pairs, which are then used to finetune a low-rank adaptation adapter with a weighted cross entropy loss. Even on this first exploration of the issue, we carry out competitively with baselines corresponding to RAG, attaining an accuracy of 81.5% throughout 100 conversations.
- * Work finished whereas at Apple
- † College of Cambridge
