Abstract

Despite the growing interest in replicating the scaled success of largelanguage models (LLMs) in industrial search and recommender systems, mostexisting industrial efforts remain limited to transplanting Transformerarchitectures, which bring only incremental improvements over strong DeepLearning Recommendation Models (DLRMs). From a first principle perspective, thebreakthroughs of LLMs stem not only from their architectures but also from twocomplementary mechanisms: context engineering, which enriches raw input querieswith contextual cues to better elicit model capabilities, and multi-stepreasoning, which iteratively refines model outputs through intermediatereasoning paths. However, these two mechanisms and their potential to unlocksubstantial improvements remain largely underexplored in industrial rankingsystems. In this paper, we propose OnePiece, a unified framework that seamlesslyintegrates LLM-style context engineering and reasoning into both retrieval andranking models of industrial cascaded pipelines. OnePiece is built on a pureTransformer backbone and further introduces three key innovations: (1)structured context engineering, which augments interaction history withpreference and scenario signals and unifies them into a structured tokenizedinput sequence for both retrieval and ranking; (2) block-wise latent reasoning,which equips the model with multi-step refinement of representations and scalesreasoning bandwidth via block size; (3) progressive multi-task training, whichleverages user feedback chains to effectively supervise reasoning steps duringtraining. OnePiece has been deployed in the main personalized search scenarioof Shopee and achieves consistent online gains across different key businessmetrics, including over +2% GMV/UU and a +2.90% increase in advertisingrevenue.