candidate generation

This work addresses the effectiveness and deployability of generative recommendation systems in large-scale industrial settings—specifically, the Xiaohongshu discovery feed serving hundreds of millions of daily active users—during the ranking stage. We propose RankGPT, a lightweight, production-ready generative ranking architecture. Contrary to prior studies attributing performance gains to training paradigms, we theoretically analyze and empirically validate— for the first time—that the generative modeling paradigm itself is the primary driver of improvement. Methodologically, RankGPT integrates LLM-informed architecture design, low-overhead inference optimization, and an online A/B testing framework to enable efficient industrial deployment. Live experiments demonstrate statistically significant improvements in user satisfaction metrics, while computational overhead remains comparable to the incumbent system. This study establishes a critical pathway and empirical benchmark for industrializing generative recommender systems.

This work addresses the limitations of traditional multi-stage retrieval systems, which suffer from error propagation due to misaligned stage-wise objectives, and end-to-end generative models, whose effectiveness is hindered by the inefficiency of autoregressive decoding. To bridge this gap, the authors propose DaV-Gen, a novel framework that introduces speculative decoding to information retrieval for the first time. DaV-Gen employs a unified “draft-and-verify” mechanism that jointly performs non-autoregressive candidate drafting and generative fine-grained verification. The model is trained with a combined objective integrating contrastive and fusion losses, effectively merging vector similarity and generative likelihood scores while leveraging a structured vector space for enhanced efficiency. This approach preserves the expressive power of generative models while significantly accelerating inference, achieving both the efficiency of sparse retrieval and the accuracy of generative ranking.

This work addresses the inefficiencies arising from the tight coupling between feature engineering and model architecture, which hinders rapid iteration, complicates deployment, and impedes reusability—particularly in low-latency online serving scenarios. To resolve this, we propose the Prompt Generation (PG) framework, which decouples feature processing logic from the model through two declarative JSON configuration files, thereby unifying offline training and online inference pipelines. PG introduces a novel configuration-driven mechanism for high-order tokenization and feature assembly, leveraging four feature categories, three composable processing components, and built-in sequence compression to establish a standardized, general-purpose inference pipeline. Deployed in Taobao Search, PG has delivered statistically significant gains of +0.47% in transaction count and +0.51% in GMV, and has been adopted by multiple search and recommendation teams as the standard iterative paradigm for generative retrieval.

To address the challenge of deteriorating retrieval relevance and generation quality in RAG systems caused by continuous evolution of external data sources, this paper proposes RAGOps—the first operations paradigm spanning the full RAG lifecycle. Methodologically, it introduces a four-dimensional architectural view, a dual-track (data and model) lifecycle model, and a cross-stage quality trade-off framework; integrates the 4+1 architectural style, LLMOps practices, automated data evaluation, and retrieval-generation co-monitoring with feedback mechanisms. Contributions include: (i) systematic identification of six core operational challenges; (ii) establishment of reusable RAGOps design principles and an evaluation methodology; and (iii) empirical validation on two industrial-scale RAG applications, demonstrating a 3.2× improvement in responsiveness to data changes and a 27.6% increase in end-to-end output reliability.

This work addresses the limitations of traditional industrial recommendation systems, which rely on multi-stage cascaded architectures suffering from redundant feature processing, pipeline complexity, and excessively long service chains. The authors propose a unified generate-then-rank framework that encodes user history once and autoregressively generates semantic ID candidates, followed by item-level ranking based on shared encoder states. Innovatively, fine-grained preferences from a high-capacity teacher ranker are distilled into the single model via Rollout distillation, enhancing ranking quality without increasing online inference cost. Evaluated on Yandex Music at scale, the proposed model replaces a legacy cascade of over 15 components with comparable latency while significantly boosting active user count by 1.41%.

This work addresses the limitations of traditional e-commerce recommendation systems, which suffer from fragmented components that hinder page-level personalization, semantic coherence, and flexible adaptation to dynamic operational objectives. To overcome these challenges, the authors propose a cascaded generative product display framework that decomposes homepage construction into two stages: layout-slot theme generation and constrained keyword generation. The approach integrates teacher-student distillation for computational efficiency and incorporates conventional ranking models to preserve the benefits of hybrid architectures. This design enables end-to-end dynamic content generation coupled with AI-driven quality filtering, achieving a balance between personalization, semantic consistency, and deployment safety. Online experiments demonstrate a 2.7% increase in add-to-cart actions per page view compared to strong baselines, with the fine-tuned model attaining performance comparable to that of closed-source large language models.

This work addresses the performance bottleneck in generative recommender systems, which often underperform due to limited ranking capabilities after candidate generation. To bridge this gap, we propose RecoChain, a novel framework that unifies generative retrieval and ranking within a single Transformer backbone for the first time. RecoChain first efficiently generates candidate items through hierarchical semantic ID prediction and then refines their ranking via a Semantic Interaction Modeling (SIM) mechanism that captures fine-grained click likelihood. This end-to-end co-optimization effectively aligns the generative and ranking objectives, closing the performance gap between them. Extensive experiments on large-scale real-world datasets demonstrate that RecoChain significantly improves Top-K recommendation accuracy while preserving strong generative capacity.

Existing generative recommender systems suffer from a disconnect between semantic ID (SID) construction and personalized ranking objectives, which limits retrieval performance. This work proposes DIG, a novel framework that unifies ranking and retrieval through the lens of tokenization for the first time: it embeds a tokenizer within a discriminative ranking model and trains the entire system end-to-end, leveraging user-item cross features to guide codebook boundary optimization. Additionally, a user-to-token (u2t) distillation module is introduced to enable efficient inference. By design, the ranking model inherently acquires retrieval capabilities, leading to significant improvements across ranking, retrieval, and joint tasks on three public benchmarks and two industrial datasets.

This work addresses the inefficiency in end-to-end evaluation of cascaded information retrieval (IR) pipelines caused by redundant computation. It introduces, for the first time, the Trie data structure into IR experimental design to automatically identify and reuse shared sub-pipelines, thereby constructing highly efficient comparative evaluation plans. Implemented within the PyTerrier framework, the approach supports combined evaluation of diverse models, including BM25, MonoT5, and DuoT5. Experiments on the MSMARCO v2 dataset demonstrate a 26% reduction in runtime compared to conventional linear evaluation plans, while user studies confirm the method’s usability and practical utility for IR researchers.

This study addresses the absence of an effective evaluation framework for structured generative search summaries—comprising overviews, titled sections, and cited source documents—that appear at the top of natural search results. It presents the first systematic effort to construct a comprehensive evaluation framework tailored to these summaries, explicitly defining their core components and multidimensional assessment criteria. By integrating large language model generation techniques with established information retrieval evaluation methodologies, the work proposes a practical and scalable evaluation framework and outlines a clear empirical validation pathway. This contribution establishes a foundational methodological basis for future research on generative search summaries and their impact on user experience and information access.