Can Small Language Models be Good Reasoners for Sequential Recommendation? This paper proposes a novel step-by-step knowledge distillation framework for recommendation (SLIM) to enable small language models (LLMs) to perform effective reasoning in sequential recommendation tasks. Sequential recommendation aims to predict users' next behavior by capturing evolving and dynamic preferences from historical interactions. However, traditional models suffer from exposure and popularity bias, and large LLMs are resource-intensive and impractical for real-world systems. SLIM addresses these challenges by distilling the reasoning capabilities of large LLMs (e.g., ChatGPT) into a smaller model (e.g., LLaMA2-7B). The framework uses CoT prompting to guide the large model in generating step-by-step reasoning, which is then used to train the smaller model. This allows the smaller model to acquire reasoning capabilities and generate high-quality recommendation rationales. The rationales are encoded into dense vectors and used to enhance both ID-based and ID-agnostic recommendation scenarios. The proposed framework includes two main components: (1) extracting recommendation rationales from LLMs using CoT prompting, and (2) fine-tuning smaller models with these rationales. The rationales are used as labels to train the smaller model, enabling it to generate step-by-step reasoning similar to the larger model. The smaller model is then deployed as a knowledge generator for sequential recommendation, providing reasoning knowledge that can be integrated with any recommendation backbone. Experiments on three datasets show that SLIM outperforms state-of-the-art baselines in both ID-based and ID-agnostic scenarios. SLIM generates meaningful rationales at affordable costs and demonstrates superior performance in handling sparse user data and mitigating popularity bias. The framework is efficient, with a small model size and low computational cost, making it suitable for real-world applications. The results indicate that SLIM can effectively enhance sequential recommendation systems by leveraging the reasoning capabilities of LLMs in a resource-efficient manner.Can Small Language Models be Good Reasoners for Sequential Recommendation? This paper proposes a novel step-by-step knowledge distillation framework for recommendation (SLIM) to enable small language models (LLMs) to perform effective reasoning in sequential recommendation tasks. Sequential recommendation aims to predict users' next behavior by capturing evolving and dynamic preferences from historical interactions. However, traditional models suffer from exposure and popularity bias, and large LLMs are resource-intensive and impractical for real-world systems. SLIM addresses these challenges by distilling the reasoning capabilities of large LLMs (e.g., ChatGPT) into a smaller model (e.g., LLaMA2-7B). The framework uses CoT prompting to guide the large model in generating step-by-step reasoning, which is then used to train the smaller model. This allows the smaller model to acquire reasoning capabilities and generate high-quality recommendation rationales. The rationales are encoded into dense vectors and used to enhance both ID-based and ID-agnostic recommendation scenarios. The proposed framework includes two main components: (1) extracting recommendation rationales from LLMs using CoT prompting, and (2) fine-tuning smaller models with these rationales. The rationales are used as labels to train the smaller model, enabling it to generate step-by-step reasoning similar to the larger model. The smaller model is then deployed as a knowledge generator for sequential recommendation, providing reasoning knowledge that can be integrated with any recommendation backbone. Experiments on three datasets show that SLIM outperforms state-of-the-art baselines in both ID-based and ID-agnostic scenarios. SLIM generates meaningful rationales at affordable costs and demonstrates superior performance in handling sparse user data and mitigating popularity bias. The framework is efficient, with a small model size and low computational cost, making it suitable for real-world applications. The results indicate that SLIM can effectively enhance sequential recommendation systems by leveraging the reasoning capabilities of LLMs in a resource-efficient manner.