Separate product-type adaptation from category-specific calibration

Determine whether the performance gain attributed to product-type adaptation in the Product-Type Test-Time Training (PT-TTT) method is distinct from gains obtainable through simpler category-specific calibration.

Background

The paper reports that product-type-specific LoRA adaptation improves performance more than a pooled adapter trained on the same expert-labeled support examples, suggesting that category-specific adaptation contributes beyond direct optimization on expert labels. It also reports improved within-product-type AUC, which is intended to rule out explanations based solely on cross-category score rescaling.

However, the authors explicitly state that their controls do not fully distinguish genuine product-type adaptation from simpler calibration procedures conditioned on product type. Resolving this issue would clarify which component of PT-TTT is responsible for the observed improvement and would provide a stronger basis for deploying gradient-based adaptation rather than category-specific calibration.

References

Third, although the pooled-adapter and macro PT-AUC controls address two alternative explanations for the observed gain, we do not fully separate product-type adaptation from simpler category-specific calibration.

— Distill Globally, Adapt Locally: Reasoning Distillation and Product-Type Test-Time Training for Scalable Trade-Up Recommendation  (2609.05363 - Liu et al., 4 Sep 2026) in Section 6, paragraph “Limitations and future work”