Generalization of MemRerank across domains and beyond metadata-only signals

Determine whether the MemRerank preference memory framework, which is trained and evaluated only on the Electronics category using product metadata for memory extraction, generalizes to other product domains and to input signals richer than product metadata.

Background

The study builds and evaluates MemRerank primarily on Electronics data and restricts memory extraction to product metadata (titles and descriptions). While results are positive in this setting, the authors point out that this scope may limit external validity.

For broader practical deployment, it is important to know whether the preference memory approach translates to other product categories and whether it can leverage richer signals (e.g., reviews, behavioral logs, or multimodal data). The paper explicitly notes that such generalization remains unclear.

References

First, we evaluate only on the Electronics category and use only product metadata for memory extraction, so generalization to other domains and richer signals remains unclear.

MemRerank: Preference Memory for Personalized Product Reranking  (2603.29247 - Peng et al., 31 Mar 2026) in Limitations