In modern e-commerce search systems, dense retrieval has become an indispensable component. By computing similarities between query and item (product) embeddings, it efficiently selects candidate products from large-scale repositories. With the breakthroughs in large language models (LLMs), mainstream embedding models have gradually shifted from BERT to LLMs for more accurate text modeling. However, these models still adopt direct-embedding methods, and the semantic accuracy of embeddings remains inadequate. Therefore, contrastive learning is heavily employed to achieve tight semantic alignment between positive pairs. Consequently, such models tend to capture statistical co-occurrence patterns in the training data, biasing them toward shallow lexical and semantic matches. For difficult queries exhibiting notable lexical disparity from target items, the performance degrades significantly. In this work, we propose the Large Reasoning Embedding Model (LREM), which novelly integrates reasoning processes into representation learning. For difficult queries, LREM first conducts reasoning to achieve a deep understanding of the original query, and then produces a reasoning-augmented query embedding for retrieval. This reasoning process effectively bridges the semantic gap between original queries and target items, significantly improving retrieval accuracy. Specifically, we adopt a two-stage training process: the first stage optimizes the LLM on carefully curated Query-CoT-Item triplets with SFT and InfoNCE losses to establish preliminary reasoning and embedding capabilities, and the second stage further refines the reasoning trajectories via reinforcement learning (RL). Extensive offline and online experiments validate the effectiveness of LREM, leading to its deployment on China's largest e-commerce platform since August 2025.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Pinterest is a leading visual discovery platform where recommender systems (RecSys) are key to delivering relevant, engaging, and fresh content to our users. In this paper, we study the problem of improving RecSys model predictions for cold-start (CS) items, which appear infrequently in the training data. Although this problem is well-studied in academia, few studies have addressed its root causes effectively at the scale of a platform like Pinterest. By investigating live traffic data, we identified several challenges of the CS problem and developed a corresponding solution for each: First, industrial-scale RecSys models must operate under tight computational constraints. Since CS items are a minority, any related improvements must be highly cost-efficient. To address this, our solutions were designed to be lightweight, collectively increasing the total parameters by only 5%. Second, CS items are represented only by non-historical (e.g., content or attribute) features, which models often treat as less important. To elevate their significance, we introduce a residual connection for the non-historical features. Third, CS items tend to receive lower prediction scores compared to non-CS items, reducing their likelihood of being surfaced. We mitigate this by incorporating a score regularization term into the model. Fourth, the labels associated with CS items are sparse, making it difficult for the model to learn from them. We apply the manifold mixup technique to address this data sparsity. Implemented together, our methods increased fresh content engagement at Pinterest by 10% without negatively impacting overall engagement and cost, and have been deployed to serve over 570 million users on Pinterest.
Large Language Models (LLMs) pose a new paradigm of modeling and computation for information tasks. Recommendation systems are a critical application domain poised to benefit significantly from the sequence modeling capabilities and world knowledge inherent in these large models. In this paper, we introduce PLUM, a framework designed to adapt pre-trained LLMs for industry-scale recommendation tasks. PLUM consists of item tokenization using Semantic IDs, continued pre-training (CPT) on domain-specific data, and task-specific fine-tuning for recommendation objectives. For fine-tuning, we focus particularly on generative retrieval, where the model is directly trained to generate Semantic IDs of recommended items based on user context. We conduct comprehensive experiments on large-scale internal video recommendation datasets. Our results demonstrate that PLUM achieves substantial improvements for retrieval compared to a heavily-optimized production model built with large embedding tables. We also present a scaling study for the model's retrieval performance, our learnings about CPT, a few enhancements to Semantic IDs, along with an overview of the training and inference methods that enable launching this framework to billions of users in YouTube.
Online financial systems for stock prediction, portfolio optimization, and algorithmic trading must remain robust against rare and volatile market events, but historical data often fails to capture diverse unprecedented financial risks, creating a major bottleneck for systematic stress testing. To address this, we propose Financial Wind Tunnel (FWT), a deployable, retrieval-augmented market simulator that generates realistic, controllable, and adaptable financial dynamics for industrial-scale training and testing. FWT integrates both macro- and micro-level market patterns through a retrieval mechanism that conditions diffusion on relevant trends, supporting real-time and interpretable generation. Unlike existing methods that lack transferability across markets and scales, FWT supports large-scale cross-market pattern synthesis and what-if causal generation, enabling generalizable simulation beyond local historical regimes. We further introduce a simulation-driven optimizer for downstream models, enabling continuous improvement of online quantitative strategies. Deployed in production pipelines on CSI300 and HKSE components, FWT demonstrates measurable gains in stock prediction and portfolio optimization performance while maintaining scalability and operational reliability, offering a practical system for quantitative finance and stress testing.
Generative Recommendation (GR), powered by Large Language Models (LLMs), represents a promising new paradigm for industrial recommender systems. However, their practical application is severely hindered by high inference latency, making them infeasible for high-throughput, real-time services and limiting their overall business impact. While Speculative Decoding (SD) has been proposed to accelerate the autoregressive generation process, existing implementations introduce new bottlenecks: they typically require separate draft models and model-based verifiers, which require additional training and increase latency overhead. In this paper, we address these challenges with NEZHA, a novel architecture that achieves hyperspeed decoding for GR systems without sacrificing recommendation quality. Specifically, NEZHA integrates a nimble autoregressive draft head directly into the primary model, enabling efficient self-drafting. This design, combined with a specialized input prompt structure, preserves the integrity of sequence-to-sequence generation. Furthermore, to tackle the critical problem of hallucination—a major source of performance degradation—we introduce an efficient, model-free verifier based on a hash set. We demonstrate the effectiveness of NEZHA through extensive experiments on public datasets and have successfully deployed the system on Taobao since October 2025, achieving 1.2% business improvement, translating to billion-level advertising revenue and serving hundreds of millions of daily active users. The code is available at https://github.com/Applied-Machine-Learning- Lab/WWW2026_NEZHA.
Online Travel Platforms (OTPs) suffer significant revenue loss from supply strikes, where rooms with physical vacancies appear sold out due to delays in manual inventory updates from hotels. While proactively adding inventory is a potential solution, this intervention faces a dual risk: hotels may later reject the booking, and more critically, the intervention might not generate platform-wide revenue, but merely shift sales from a competing hotel. This paper is the first to formalize the inventory decision on OTPs as a causal inference problem. We propose CS2NET, a Causality-Driven, Scarcity- and Service-Aware Network that estimates the platform-wide Individual Treatment Effect of each inventory addition. CS2NET addresses the unique challenges of the OTP environment by integrating: (1) a Room Type Scarcity Representation module for inferring true room availability, (2) a Hotel Service-Engagement Representation module for predicting hotel acceptance, and (3) a bias-corrected causal framework to estimate platform-level uplift while mitigating selection bias. Extensive experiments and an online A/B test on a major OTP, demonstrate that CS2NET significantly increases confirmed bookings and platform revenue, generating over 10 million RMB in additional annual GMV. We also release the first causality dataset for third-party inventory management.
Most large-scale recommender systems follow a multi-stage cascade of retrieval, pre-ranking, ranking, and re-ranking. A key challenge at the pre-ranking stage arises from the heterogeneity of training instances sampled from coarse-grained retrieval results, fine-grained ranking signals, and exposure feedback. Our analysis reveals that prevailing pre-ranking methods, which indiscriminately mix heterogeneous samples, suffer from gradient conflicts: hard samples dominate training while easy ones remain underutilized, leading to suboptimal performance. We further show that the common practice of uniformly scaling model complexity across all samples is inefficient, as it overspends computation on easy cases and slows training without proportional gains. To address these limitations, this paper presents Heterogeneity-Aware Adaptive Pre-ranking (HAP), a unified framework that mitigates gradient conflicts through conflict-sensitive sampling coupled with tailored loss design, while adaptively allocating computational budgets across candidates. Specifically, HAP disentangles easy and hard samples, directing each subset along dedicated optimization paths. Building on this separation, it first applies lightweight models to all candidates for efficient coverage, and further engages stronger models on the hard ones, maintaining accuracy while reducing cost. This approach not only improves pre-ranking effectiveness but also provides a practical perspective on scaling strategies in industrial recommender systems. HAP has been deployed in the Toutiao production system for 9 months, yielding up to 0.4% improvement in user app usage duration and 0.05% in active days, without additional computational cost. We also release a large-scale industrial hybrid-sample dataset to enable the systematic study of source-driven candidate heterogeneity in pre-ranking.
In recent years, the success of large language models (LLMs) has driven the exploration of scaling laws in recommender systems. However, models that demonstrate scaling laws are actually challenging to deploy in industrial settings for modeling long sequences of user behaviors, due to the high computational complexity of the standard self-attention mechanism. Despite various sparse self-attention mechanisms proposed in other fields, they are not fully suited for recommendation scenarios. This is because user behaviors exhibit personalization and temporal characteristics: different users have distinct behavior patterns, and these patterns change over time, with data from these users differing significantly from data in other fields in terms of distribution. To address these challenges, we propose SparseCTR, an efficient and effective model specifically designed for long-term behaviors of users. To be precise, we first segment behavior sequences into chunks in a personalized manner to avoid separating continuous behaviors and enable parallel processing of sequences. Based on these chunks, we propose a three-branch sparse self-attention mechanism to jointly identify users' global interests, interest transitions, and short-term interests. Furthermore, we design a composite relative temporal encoding via learnable, head-specific bias coefficients, better capturing sequential and periodic relationships among user behaviors. Extensive experimental results show that SparseCTR not only improves efficiency but also outperforms state-of-the-art methods. More importantly, it exhibits an obvious scaling law phenomenon, maintaining performance improvements across three orders of magnitude in FLOPs. In online A/B testing, SparseCTR increased CTR by 1.72% and CPM by 1.41%. Our source code is available at https://github.com/laiweijiang/SparseCTR.
Short-video platforms have rapidly become a new generation of information retrieval systems, where users formulate queries to access desired videos. However, user queries, especially long-tail ones, often suffer from spelling errors, incomplete phrasing, and ambiguous intent, resulting in mismatches between user expectations and retrieved results. While large language models (LLMs) have shown success in long-tail query rewriting within e-commerce, they struggle on short-video platforms, where proprietary content such as short videos, live streams, micro dramas, and user social networks falls outside their training distribution. To address this challenge, we introduce CardRewriter, an LLM-based framework that incorporates domain-specific knowledge to enhance long-tail query rewriting. For each query, our method aggregates multi-source knowledge relevant to the query and summarizes it into an informative and query-relevant knowledge card. This card then guides the LLM to better capture user intent and produce more effective query rewrites. We optimize CardRewriter using a two-stage training pipeline: supervised fine-tuning followed by group relative policy optimization, with a tailored reward system balancing query relevance and retrieval effectiveness. Offline experiments show that CardRewriter substantially improves rewriting quality for queries targeting proprietary content. Online A/B testing further confirms significant gains in long-view rate (LVR) and click-through rate (CTR), along with a notable reduction in initiative query reformulation rate (IQRR). Since September 2025, CardRewriter has been deployed on Kuaishou, one of China's largest short-video platforms, serving hundreds of millions of users daily.
Trends on short-video platforms evolve at a rapid pace, with new content issues emerging every day that fall outside the coverage of existing annotation policies. However, traditional human-driven discovery of emerging issues is too slow, which leads to delayed updates of annotation policies and poses a major challenge for effective content governance. In this work, we propose an automatic issue discovery method based on multimodal LLM agents. Our approach automatically recalls short videos containing potential new issues and applies a two-stage clustering strategy to group them, with each cluster corresponding to a newly discovered issue. The agent then generates updated annotation policies from these clusters, thereby extending coverage to these emerging issues. Our agent has been deployed in the real system. Both offline and online experiments demonstrate that this agent-based method significantly improves the effectiveness of emerging-issue discovery (with an F1 score improvement of over 20%) and enhances the performance of subsequent issue governance (reducing the view count of problematic videos by approximately 15%). More importantly, compared to manual issue discovery, it greatly reduces time costs and substantially accelerates the iteration of annotation policies.
Generative Recommenders (GRs), exemplified by the Hierarchical Sequential Transduction Unit (HSTU), have emerged as a powerful paradigm for modeling long user interaction sequences. However, we observe that their ''flat-sequence'' assumption overlooks the rich, intrinsic structure of user behavior. This leads to two key limitations: a failure to capture the temporal hierarchy of session-based engagement, and computational inefficiency, as dense attention introduces significant noise that obscures true preference signals within semantically sparse histories, which deteriorates the quality of the learned representations. To this end, we propose a novel framework named HPGR (Hierarchical and Preference-aware Generative Recommender), built upon a two-stage paradigm that injects these crucial structural priors into the model to handle the drawback. Specifically, HPGR comprises two synergistic stages. First, a structure-aware pre-training stage employs a session-based Masked Item Modeling (MIM) objective to learn a hierarchically-informed and semantically rich item representation space. Second, a preference-aware fine-tuning stage leverages these powerful representations to implement a Preference-Guided Sparse Attention mechanism, which dynamically constrains computation to only the most relevant historical items, enhancing both efficiency and signal-to-noise ratio. Empirical experiments on a large-scale proprietary industrial dataset from APPGallery and an online A/B test verify that HPGR achieves state-of-the-art performance over multiple strong baselines, including HSTU and MTGR.
Short-video recommenders such as Douyin must exploit extremely long user histories without breaking latency or cost budgets. We present an end-to-end system that scales long-sequence modeling to 10k -length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy---train on shorter windows, infer on much longer ones---so the model generalizes to 10k histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end long-sequence recommendation to the 10k regime.
Auto-bidding plays an essential role in online advertising, allowing agents to automatically adjust bids for advertisers. Recently, the rise of Marketing Management service in e-commerce platforms has driven the evolution from auto-bidding to auto-marketing, enabling merchants to delegate their advertising bidding and product's coupon discounting decisions to agents. Auto-marketing requires agents to jointly decide on bidding and coupon discounting. Furthermore, compared to classic static constraints, auto-marketing agent faces a self-funding constraint (where the budget for both bidding and coupon discounting is entirely derived from the agent's commission revenue). Existing rule-based or RL-based methods often struggle with dynamic environments and complex sequential dependencies. To overcome these limitations, we propose a Generative Auto-Marketing framework (GAM), designed for performing joint sequential decisions on bidding and coupon discounting, and optimizing business objectives through post-training alignment. Furthermore, GAM employs a flexible, constraint-aware reward alignment module, and utilizes Group Relative Policy Optimization (GRPO) to align the pre-trained model, thus empirically balancing objective maximization and constraint satisfaction. We construct an offline simulation environment based on large-scale real-world dataset, and demonstrate the effectiveness of GAM through extensive experimental results.
Query-product relevance prediction is fundamental to e-commerce search and has become even more critical in the era of AI-powered shopping, where semantic understanding and complex reasoning directly shape the user experience and exert an indirect, yet substantial impact on business conversion. Large Language Models (LLMs) enable generative, reasoning-based approaches, typically aligned via supervised fine-tuning (SFT) or preference optimization methods like Direct Preference Optimization (DPO). However, the increasing complexity of business rules and user queries exposes the inability of existing methods to endow models with robust reasoning capacity for long-tail and challenging cases. Efforts to address this via reinforcement learning strategies like Group Relative Policy Optimization (GRPO) often suffer from sparse terminal rewards, offering insufficient guidance for multi-step reasoning, which in turn slows convergence. To address these challenges, we propose TaoSR-AGRL, an Adaptive Guided Reinforcement Learning framework for LLM-based relevance prediction in Taobao Search Relevance. TaoSR-AGRL introduces two key innovations: (1) Rule-aware Reward Shaping, which decomposes the final relevance judgment into dense, structured rewards aligned with domain-specific relevance criteria; and (2) Adaptive Guided Replay, which identifies low-accuracy rollouts during training and injects targeted ground-truth guidance to steer the policy away from stagnant, rule-violating reasoning patterns toward compliant trajectories. TaoSR-AGRL was evaluated on large-scale datasets and through online evaluations on Taobao Search. It consistently outperforms DPO and GRPO baselines, improving relevance accuracy and rule adherence, with measurable gains in user engagement and stable training dynamics. The model has been deployed on Taobao, serving hundreds of millions of users.
In e-commerce search, the diverse ways in which users express intentions lead to lexical and semantic gaps between queries and product descriptions, making query rewriting (QR) indispensable for improving matching efficiency. With the development of LLMs, QR has evolved from discriminative approaches to various LLM-based alignment methods. However, these methods typically treat all queries uniformly, without fundamentally distinguishing their rewriting difficulty or underlying linguistic issues, making the rewritten query deviate from human expectations. To address this limitation, we propose AWHCP (Aligning with Human Cognition and Preference), a novel framework that adopts a human-centric perspective and introduces the Problem–Intention–Fix–Rewrite (PIFR) paradigm. Built upon PIFR, AWHCP establishes a multi-granularity alignment training framework that simultaneously aligns with both system retrieval preferences and human rewriting behaviors. First, we construct high-quality PIFR-structured data and perform supervised fine-tuning to enable the model to learn human-like rewriting patterns. Second, we apply beam search to generate multiple candidates and leverage system-side feedback signals to conduct coarse-grained direct preference alignment, endowing the model with initial difficulty-aware reasoning capabilities. Third, we introduce a multi-dimensional rewrite quality judgment model trained via Group Relative Policy Optimization (GRPO), enabling fine-grained alignment with nuanced human rewriting preferences. Deployed on 1688's main search engine since August 2025, AWHCP has demonstrated strong effectiveness through extensive offline evaluations and large-scale online A/B tests, leading to a +3.9% gain in UV-L2O.
We present BanditLP, a scalable multi-stakeholder contextual bandit framework that unifies neural Thompson Sampling (TS) for learning objective-specific outcomes with a large-scale linear program (LP) for constrained action selection at serving time. The methodology is application-agnostic, compatible with arbitrary neural architectures, and deployable at web scale, with an LP solver capable of handling billions of variables. Experiments on public benchmarks and synthetic data show consistent gains over strong baselines. We apply this approach in LinkedIn's email marketing system and demonstrate business win, illustrating the value of integrated exploration and constrained optimization in production.
Web-scale editorial crawling must balance coverage and freshness within tight politeness and request budgets. Nonetheless, in production systems, manual seed management remains common despite being inefficient—oversampling redundant seeds while missing high-yield ones. URLBank replaces manual curation with a label-free controller that infers optimal seed selection directly from temporal crawl telemetry. It identifies candidate entry points, estimates stability—the persistence of links across crawler revolutions—and productivity—the rate of first-seen publications—and ranks them through greedy marginal gain on a shared-credit coverage objective. In a shadow A/B evaluation spanning 5,238 sites, URLBank consistently achieves higher coverage, greater efficiency, and earlier discovery under identical conditions. Gains remain stable across Top-K budgets, approaching near-complete coverage with far fewer seeds. Deployed alongside Meltwater's production crawler Pulitzer, URLBank operates with versioned policies, ranked prefixes for crawl budgets, and integrated health diagnostics, making allocation transparent, auditable, and reversible. Together, these results demonstrate that temporal signals, through an interpretable greedy objective, yield large, measurable improvements in industrial-scale coverage, resource efficiency, and freshness.