GrocLM Boosts E-commerce Grocery Recommendations with LLMs

GrocLM is a novel large language model specifically fine-tuned for grocery category recommendation within e-commerce environments. Addressing the challenges of scalability and accuracy faced by traditional item-level recommendation systems, GrocLM focuses on category-level recommendations to better capture the cyclical purchasing behaviors and diverse user intents common in online grocery shopping. The system employs a two-stage LoRA-based training strategy, which effectively encodes recurring purchasing patterns directly into the model's parameters, allowing for more robust utilization of rebuying signals compared to prompt-based conditioning. To ensure the generation of valid and controllable outputs, GrocLM incorporates a trie-based constrained decoding mechanism that operates over a predefined set of grocery categories. Extensive experiments using both proprietary production data and a public benchmark dataset demonstrated GrocLM's superior performance against strong baselines. In a live production restocking task, the model achieved a 7.5% relative improvement in cart-adds per impression, all while maintaining efficient inference by generating all categories simultaneously. These results highlight the practical effectiveness of integrating LLMs into structured recommendation systems for e-commerce.