Customizing In-context Learning for Dynamic Interest Adaption in LLM-based Recommendation

Periodically updating Large Language Model (LLM)-based recommender systems to adapt to dynamic user interests-as is done for traditional ones-is impractical due to high training costs. This work explores the possibility of achieving interest adaptation without any model-level updates via In-context Learning (ICL), which enables adaptation through fewshot examples within input prompts. Using recent interactions as ICL few-shot examples, LLMs can directly learn the new interest in prompt without needing model updates. While pre-trained LLMs possess strong in-context learning capabilities, these are often diminished after task-specific fine-tuning in recommender systems, and the original ICL mechanism lacks specialization for recommendation tasks. To address this, we propose RecICL, a framework for recommendation-oriented ICL. It performs tuning in an ICL manner, structuring new-interest interactions as few-shot examples to capture dynamic interests during data fitting. Extensive experiments across multiple benchmarks demonstrate RecICL's superior performance, achieving better results without model updates. Our implementation is publicly available at https://github.com/ym689/rec_icl.