Can Small Language Models be Good Reasoners in Recommender Systems?

Recent advances in large language models (LLMs) have shown strong reasoning capabilities, prompting researchers to explore their use in recommender systems. However, the high computational cost of LLMs makes them unsuitable for industrial recommendation scenarios.

The authors introduce SLIM, a knowledge‑distillation framework that compresses the reasoning power of a teacher LLM (e.g., ChatGPT‑3.5) into a much smaller student model (e.g., LLaMA2‑7B) for sequential recommendation tasks. The distillation follows a step‑by‑step (CoT) prompting strategy: (1) summarize user preferences from historical behavior, (2) generate candidate categories/brands, and (3) recommend items matching those categories/brands.

During distillation, the teacher LLM produces natural‑language reasoning traces that serve as supervision labels for the student model. After fine‑tuning, the student model can generate high‑quality recommendation rationales with only ~4% of the teacher’s parameters, enabling deployment on a single GPU.

The generated rationales are encoded with a text encoder (e.g., BERT) and fused with traditional ID‑based or ID‑agnostic recommendation representations, enriching item and user embeddings with open‑world knowledge and reducing exposure and popularity biases.

Extensive experiments on three datasets (Video Games, Grocery & Gourmet Food, Home & Kitchen) demonstrate that SLIM consistently improves recommendation accuracy, explainability, cold‑start performance, and mitigates popularity bias, while consuming far less resources than state‑of‑the‑art LLM‑based recommenders.

In production at Ant Group’s Alipay platform, the SLIM framework combined with a proprietary large model and vector retrieval has been deployed to serve billions of users, confirming its practical effectiveness.

References: [1] Towards Open‑World Recommendation with Knowledge Augmentation from Large Language Models. [2] How Can Recommender Systems Benefit from Large Language Models: A Survey. [3] Text Matching Improves Sequential Recommendation by Reducing Popularity Biases.