Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation

A study on scalable upgrade recommendations proposes a two-stage framework aimed at distilling the reasoning capabilities of large language models into efficient non-generative student models and adapting them for specific product types. In the first stage (global distillation), structured labels and reasons are supervised using retrieval enhancement and small-sample teacher generation to train compact classifiers; only pre-computed embedding vectors are used during reasoning, without needing to invoke large models or text generation. Experiments show that the four-classification reasoning distillation student achieved an AUC of 0.924 on a fixed manually labeled dataset, which is better than the student model using only labels (0.912). In the second stage (local adaptation), a training method tailored for product types was used to optimize lightweight adapters with small-sample demonstrations, increasing the AUC from 0.924 to 0.941 and the average accuracy from 0.920 to 0.940. In proxy catalog tests, this distilled student ran approximately 5,000 times faster than direct large model reasoning on a single eight-GPU machine, with estimated costs reduced by 10,000 times.