GrocLM: LLM-powered grocery recommender boosts cart-adds by 7.5%

GrocLM, developed by researchers at Instacart (Yuan Zhong, Chuanwei Ruan, et al.), tackles the unique challenges of online grocery recommendation. Unlike traditional item-level recommenders, GrocLM recommends product categories—a more structured and scalable approach. The model is a fine-tuned LLM that uses a two-stage LoRA-based training strategy to encode cyclical purchasing patterns (e.g., weekly grocery runs) directly into its parameters, making it more effective than prompt-based methods at capturing rebuying signals. To guarantee valid and controllable outputs, the team introduced a trie-based constrained decoding mechanism that restricts the model's predictions to a predefined category space, avoiding nonsensical suggestions.

In a live production restocking task, GrocLM achieved a 7.5% relative improvement in cart-adds per impression, demonstrating real-world impact. The model also maintains efficient inference by generating all recommended categories jointly rather than sequentially. These results, validated on both proprietary Instacart data and a public benchmark, show that integrating LLMs into structured recommendation systems can boost performance while remaining practical for production. The paper is available on arXiv (2607.24764) and represents a significant step toward more intelligent, context-aware grocery e-commerce.