Towards Context-aware Reasoning-enhanced Generative Searching in E-commerce

TL;DR

CRS tackles the limitations of context integration in e-commerce search by unifying heterogeneous user and item signals into a text-based substrate and leveraging large language models for explicit, multi-step reasoning. It introduces a self-evolving post-training loop that alternates supervised fine-tuning with reinforcement learning and a debiased RL algorithm (R-GRPO) to better align optimization with ranking quality. The approach achieves consistent, substantial gains over strong baselines on three real-world datasets, including strong improvements for cold-start and highly similar item scenarios. This work demonstrates a practical path to improving search-based recommendation through context-aware reasoning, scalable inference, and robust training in large-scale e-commerce settings.

Abstract

Search-based recommendation is one of the most critical application scenarios in e-commerce platforms. Users' complex search contexts--such as spatiotemporal factors, historical interactions, and current query's information--constitute an essential part of their decision-making, reflecting implicit preferences that complement explicit query terms. Modeling such rich contextual signals and their intricate associations with candidate items remains a key challenge. Although numerous efforts have been devoted to building more effective search methods, existing approaches still show limitations in integrating contextual information, which hinders their ability to fully capture user intent. To address these challenges, we propose a context-aware reasoning-enhanced generative search framework for better \textbf{understanding the complicated context}. Specifically, the framework first unifies heterogeneous user and item contexts into textual representations or text-based semantic identifiers and aligns them. To overcome the lack of explicit reasoning trajectories, we introduce a self-evolving post-training paradigm that iteratively combines supervised fine-tuning and reinforcement learning to progressively enhance the model's reasoning capability. In addition, we identify potential biases in existing RL algorithms when applied to search scenarios and present a debiased variant of GRPO to improve ranking performance. Extensive experiments on search log data collected from a real-world e-commerce platform demonstrate that our approach achieves superior performance compared with strong baselines, validating its effectiveness for search-based recommendation.

Figures (3)

• Figure 1: The overview of the proposed CRS framework for search-based recommendation in e-commerce. (a) Context representation and alignment are first conducted, where we unify the complex user and item context into a structured JSON format, and introduce a pre-training stage to align the LLM backbone with e-commerce search context. (b) To tackle the challenge of the scarcity of supervision on the reasoning trajectories, we propose a novel self-evolving post-training paradigm that better stimulates the model's context-aware reasoning ability through alternating between exploration with RL and exploitation with SFT. (c) The debiased RL algorithm, R-GRPO, is proposed for the ranking task, which addresses the optimization target misalignment and reward estimation bias issues. (d) After training, our framework conducts inference in a two-step schema.
• Figure 2: Performance evolution of the proposed CRS framework during self-evolving post-training on the All-100K and All-50K datasets, measured by NDCG@10 and HR@10.
• Figure 3: HR@10 evaluation on the three constructed user groups comparing CRS with representative match-based and generative methods.