SCOReD optimizes teacher CoTs for 27% shorter student reasoning and better recommendations

Large teacher models in recommendation tasks exhibit high reasoning uncertainty, often rechecking answers without revision. When distilled via chain-of-thought (CoT) supervision, smaller student models learn to mimic this verbose, uncertain behavior—producing lengthy rationales that never revise initial guesses. This problem is compounded by the novelty of the recommendation domain, making teacher traces highly out-of-distribution for student LLMs. To address these issues, researchers introduce SCOReD (Student-Aware CoT Optimization for Recommendation Distillation). The framework first parses each teacher reasoning trace into typed segments, then uses the student model's attention mechanism to score the importance of each segment. Based on output length and comparative log probability lift of the answer given the edit, SCOReD dynamically selects per-segment actions: KEEP, REWRITE, FUSE, or PRUNE. This prunes redundant sections while preserving information-dense parts, effectively adapting raw teacher traces to the student's output distribution.

Training on SCOReD-optimized CoTs provides a much cleaner learning signal to the student model. In experiments, the approach improves over standard supervised fine-tuning by 1.56% in NDCG and 1.9% in Recall@5 for recommendation tasks, while simultaneously reducing the reasoning length by 27.3%. This means students generate shorter, more focused rationales without sacrificing—and even slightly improving—recommendation accuracy. The work has significant practical implications for deploying efficient yet accurate recommendation systems powered by smaller language models. By addressing the core issue of out-of-distribution teacher traces and redundant reasoning, SCOReD offers a generalizable framework for CoT distillation in specialized domains.