ZeroSeedRec: Enhancing Zero-Shot Recommendation with Domain-Specific Instruction-Tuned LLMs: A Pipeline for Synthetic Seed and Self-Instruct Data Generation

Abstract

Large language models (LLMs) have recently shown remarkable promise as zero-shot recommenders, capable of generating meaningful suggestions without task-specific training. Yet, much of the current research underplays the role of contemporary enhancement techniques that could significantly elevate this performance. In response, we present a refined three-step prompting approach that combines strategic data augmentation, efficient fine-tuning, and targeted instruction design. We demonstrate that training on a single, consistent instruction is sufficient, challenging the notion that variety is a prerequisite for effective tuning. While prior studies have highlighted the fragility of LLMs to different prompt formats, our experiments reveal that using a JSON structure enables more reliable data interpretation and improved outcomes. This paper presents a practical blueprint for zero-shot recommendation using large language models (LLMs).

Ähnliche Arbeiten