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

Directly querying large language models across hundreds of millions of product pairs to recommend trade-up alternatives is commercially prohibitive. This paper proposes a two-level architecture that distills RAG-assisted LLM rationales into a compact 15.5M-parameter non-generative classifier. Operating solely on precomputed 768-dimensional embeddings during inference, it further incorporates product-type test-time training (PT-TTT) via lightweight category adapters. Evaluated on an 8,352-pair human benchmark, the system reaches an AUC of 0.941 while cutting estimated inference costs by roughly 10,000x on an eight-GPU machine.