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

About this paper

Siliang Liu, Mohammad Ghasemi, Sapan Patel, Amin Banitalebi-Dehkordi

The paper presents a trade-up recommender that uses an LLM offline to label product pairs and provide rationales, trains a compact four-class classifier, and adapts it with product-type-specific expert supervision. On 8,352 human-annotated pairs, the student reaches AUC 0.924 versus 0.912 for label-only training. Adaptation raises AUC to 0.941, evaluated across 29 product types.

Key insights

• Rationale supervision is most beneficial when the student model is trained on fine-grained (four-class) relation labels, allowing it to capture more subtle distinctions in relational geometry compared to binary labels.
• A compact, non-generative student model, through reasoning distillation, can effectively leverage LLM capabilities to achieve high performance while overcoming the scalability challenges of direct LLM inference.
• Product-Type Test-Time Training (PT-TTT) provides additional performance gains by specializing the global model's decision boundary to individual product types, outperforming simpler pooled expert supervision without per-query LLM costs.

Summary by alphaXiv.