Determine whether alignment increases format sensitivity

Determine whether preference alignment increases sensitivity to the difference between structured product text and natural-language product text.

Background

The paper evaluates each preference pair in both a structured attribute representation and a model-rendered natural-language representation. An initial development result suggested that alignment on structured text might make the model more brittle when products are rendered as prose.

However, the benchmark comparison does not test the relevant interaction directly: it compares checkpoints within each format rather than comparing the change in format gaps caused by alignment. The authors therefore leave the effect of alignment on format sensitivity unresolved.

References

Whether alignment increases format sensitivity is a separate question these tests do not answer: they compare checkpoints within each format rather than the difference between format gaps, and the aligned 8B's 1.6-point gap against base-8B's 0.8 (\Cref{tab:main_structured,tab:main_nl}) would need an interaction test we do not run.

— ZooWork-ShopRanker: An Open, Preference-Aligned E-Commerce Reranker  (2609.31002 - Xue et al., 25 Sep 2026) in Appendix, Section 'Mixed-format training' (Section label app:format_training)