Enriching Input Context to Endow Ranking Models with Reasoning Capabilities

Determine effective strategies to enrich the input context of industrial Transformer-based ranking models beyond DLRM-style feature engineering so as to better endow these models with reasoning capabilities, specifically by augmenting raw user–item interaction sequences with structured contextual signals.

Background

Most existing industrial Transformer-based models rely primarily on raw user–item interaction sequences and feature engineering practices tailored to DLRM-style architectures, which lack the structured richness of LLM-style prompts.

This open question motivates the paper’s context engineering design (interaction history, preference anchors, situational descriptors, and candidate item set) to provide richer, structured contextual cues that help elicit reasoning capacity in ranking models.

References

Moreover, existing feature engineering practices remain predominantly tailored to DLRM-style architectures, leaving open the question of how to enrich context to better endow ranking models with reasoning capabilities.

— OnePiece: Bringing Context Engineering and Reasoning to Industrial Cascade Ranking System (2509.18091 - Dai et al., 22 Sep 2025) in Section 1, Introduction (bullet: “How to construct an informative input context?”)