Fine-Tuning Small Language Models with Low-Rank Adapters to Mimic User Behaviors (arxiv.org)

Researchers propose a practical method for simulating realistic user behavior for recommender systems by combining frozen large language models with fine-tuned small language models (SLMs). Instead of expensive end-to-end LLM fine-tuning on raw tabular interaction logs, they use a frozen LLM to convert large-scale user-item interaction data into robust textual user representations, then train lightweight SLM “user agents” on those representations. To scale to millions of users while retaining personalization, the team trains multiple low-rank adapters (akin to LoRA) tied to user groups or personas, enabling efficient specialization without full-model updates. This approach is significant because it addresses three key barriers in prior LLM-based simulators: parsing tabular data at scale, overcoming pretraining biases to capture user-specific preferences, and doing so cost-effectively for many users. Low-rank adapters provide a middle ground that maintains per-person behavior fidelity while keeping compute, memory, and deployment costs low. Empirical results show these persona-driven SLM agents better align offline metrics with likely real-world recommender performance, suggesting a scalable path for more accurate simulator-driven evaluation and continual user modeling in production systems.