Jump Start or False Start? A Theoretical and Empirical Evaluation of LLM-initialized Bandits

Research examines how Large Language Models can be used to initialize contextual bandits for recommendation systems, finding that LLM-generated preferences remain effective up to 30% data corruption but can harm performance beyond 50% corruption. The study provides theoretical analysis showing when LLM warm-starts outperform cold-start approaches, with implications for AI-driven recommendation systems.

• →LLM-initialized bandits maintain effectiveness up to 30% data corruption but lose advantage around 40% corruption.
• →Systematic misalignment in LLM-generated preferences can lead to worse performance than cold-start bandits even without noise.
• →Researchers developed theoretical conditions for when LLM-based warm starts provably outperform cold-start approaches.
• →The study validates findings across multiple datasets and LLMs, showing alignment estimates reliably predict performance.
• →Results have significant implications for AI recommendation systems that rely on LLM-generated user preference data.