Large Language Models (LLMs) have shown strong capabilities in open-domain question answering (QA), but deploying them in real-world online systems introduces critical challenges. These include: (1) handling both simple and complex queries with appropriate levels of reasoning, (2) minimizing latency without compromising answer quality, and (3) maintaining answer consistency under evolving and noisy retrieval contexts. To address these challenges, we propose Retain-to-Refine (ℜ2ℜ), an adaptive agent-based QA framework designed for practical deployment. ℜ2ℜ integrates a Query Critic Agent (QCA) to assess query difficulty and route it accordingly: simple queries are answered directly using fast, prompt-based LLM calls, while complex queries are handled by a Memory Augmented Agent (MAA). MAA performs iterative reasoning guided by a unique long-short memory mechanism. Long-term memory retains and consolidates stable, core facts to ground the reasoning process, while short-term memory identifies transient information gaps to formulate highly focused subsequent queries. To ensure evidence quality, a Supervised Retrospection module validates and filters retrieved documents at each step. This agent-based design enables ℜ2ℜ to dynamically allocate computation based on question complexity, reducing unnecessary overhead while preserving high-quality answers when multi-step reasoning or external knowledge is required. Extensive evaluations across various settings and datasets demonstrate that the efficiency of R2R across diverse question types. In online settings, ℜ2ℜ delivers substantial gains in both response quality and efficiency, making it well-suited for large-scale industrial deployment in real-time QA services.

Generative inference with Large Language Models (LLMs) is the cornerstone of web-scale AI search, where queries are answered using vast, heterogeneous documents retrieved via Retrieval-Augmented Generation (RAG). This paradigm is critically bottlenecked by the cost of self-attention mechanism on long context. The sheer diversity of retrieved web content (multi-sourced, multi-lingual, multi-faceted) makes simple Key-Value (KV) cache optimizations with pre-fixed subsets ineffective, demanding a dynamic, content-aware approach. This challenge, however, introduces a classic chicken-and-egg problem: the model cannot foresee the necessary KV entries for attention without first inferring on the content, yet doing so on the full context is prohibitively expensive. This paper introduces P&F, a unified framework that resolves this dilemma through a core ''probe-and-fetch'' mechanism, which ingeniously integrates with speculative decoding -- an acceleration approach already adopted in web-scale AI search. The probe step repurposes the speculative draft model: while generating candidate tokens, it simultaneously probes the context to predict the most salient KV entries the large model will need for attention. The fetch step immediately acts on this prediction, asynchronously fetching these sparse entries from memory. This synergistic design piggybacks the probing step onto the drafting process, allowing the expensive gathering of a sparse KV cache to be fully masked. Crucially, this co-design breaks the sequential dependency bottleneck that cripples naive integrations of speculative decoding and prefetching due to synchronization issues. Extensive experiments show P&F significantly outperforms state-of-the-art methods in throughput and scalability, offering a practical, drop-in solution. Extensive offline evaluations across various settings and datasets demonstrate that P&F yields superior throughput and scalability compared to advanced baselines, while maintaining model quality across diverse models and scales. In online settings, P&F delivers substantial gains in throughput improvements while preserving response quality, making it well-suited for large-scale industrial deployment in real-time AI Search services.

With the rise of AI in education, knowledge tracing (KT) has become important for modeling students’ knowledge from interaction data. However, existing methods still face three major challenges, including limited modeling of personalized exercise–concept relations, low robustness to noisy interactions, and inefficient training due to suboptimal data selection. To address these issues, we propose R 2 GCurL, a novel KT framework with two key designs. First, we recast KT as a graph classification problem and construct dynamic graphs from student responses, enabling the model to capture structural relations between exercises and concepts for more personalized KT. Second, we introduce a data-centric curriculum learning strategy based on dynamic graph entropy. Under our definition, pairwise dynamic graph entropy measures graph-transition continuity, where larger values indicate stronger structural similarity. Its sequence-level aggregation is used to derive a structure-aware difficulty signal for sample scheduling. On top of this, an RL-based scheduler further adapts batch selection based on model feedback and is especially beneficial under noisier and more unstable training regimes. Theoretical analysis shows that R 2 GCurL has lower computational complexity than existing graph-based KT models. Extensive experiments on five real-world datasets confirm its effectiveness, robustness, and generalizability, including as a plug-and-play enhancement for sequence-based KT models.

Pre-trained language models (PLMs) have been successfully used to build high-performance ranking models for large-scale information retrieval systems. However, traditional PLM-based ranking approaches face two key challenges: (1) these models use both sparse and dense content (such as the query/title and content of documents) as inputs, which may require different attention allocations; and (2) traditional PLM-based ranking approaches have identified multiple objectives to gauge user satisfaction with ranking results, but integrating these objectives into the end-to-end training process and the subsequent feature updates and iterations usually involves significant computational resource overhead. In this paper, we propose a novel PLM-based ranking approach M2oE Rank, Multi-objective Mixture-of-Experts (MoE) enhanced Ranking. Specifically, M2oERank lever-ages a context-aware PLM-based hierarchical encoder to extract semantic relevance between the query and the document title and content, while allowing for separate dense and sparse attention for different inputs. With the extracted semantic relevance repre-sentations, multifacet user satisfaction features and task-specific annotations, M2oERank employs an MoE module to perform multi-objective pre-training of ranking models focused on user satisfaction. Finally, M2oERank uses a weight fusion module that fuses outputs from the above experts to predict ranking scores. Moreover, we present a three-stage offline training strategy and the online system workflow for deploying M2oERank at web-scale search. To demonstrate the effectiveness of our proposed approach, we conduct extensive offline and online evaluations using real-world web traffic from Baidu Search. The comparisons against numbers of advanced baselines confirmed the advantages of M2oERank in producing high-performance ranking models for web-scale search.

As modern learning-to-rank (LTR) systems rely on both textual and behavioral features, it is essential to extend pre-trained language models (PLMs) from text (queries and webpages) understanding to end-to-end ranking score prediction subject to multiple objectives, such as relevance, quality, authority, and recency. While textual inputs encompass a broader array of features than mere relevance and behavioral features are frequently skewed by user feedback with position bias, an integrated solution is required to jointly disentangle and fuse these heterogeneous features, ensuring robust and unbiased ranking predictions. In this work, we introduce RankExpert, a unified framework that holistically models heterogeneous ranking signals by integrating PLM-based semantic extraction with behavioral cues. RankExpert employs a lightweight PLM with hierarchical distillation for efficient query-document representation, a mixture-of-experts (MoE) architecture to disentangle and optimize objectives across relevance, quality, authority, and recency, and a click expert to mitigate position bias in user feedback within a dual-tower design. Then, RankExpert utilizes an adaptive weight fusion layer to dynamically integrate the outputs of these specialized experts dynamically, aligning ranking predictions with diverse user intents. Extensive offline evaluations on two large-scale real-world datasets reveal that RankExpert significantly outperforms strong competitor models and substantial gains in key performance indicators. Moreover, we deploy RankExpert at Baidu Search and conduct comprehensive online evaluations with real-world web traffic, demonstrating substantial improvements in user satisfaction metrics over the online legacy system and underscoring the practicality of the proposed approach.

The exponential growth of online content and increasingly diverse user needs have underscored the necessity for ranking models that go beyond traditional relevance assessments. Although several open-source benchmarks have significantly advanced academic research in Learning-to-Rank (LTR), these datasets predominantly focus on either text-based relevance or user behavior (click-through or dwell time) signals separately. This separation has inadvertently burdened academic progress by limiting the exploration of multifaceted, satisfaction-oriented ranking models. In contrast, industry research has begun to delve into integrated approaches that fuse prior (relevance, authority, recency, and quality) with posterior (user interaction such as clicks and dwell time) signals, thereby better capturing true user satisfaction. In this paper, we introduce FULTR-a large-scale, prior-posterior FUsion LTR dataset. FULTR comprises over 224M queries and 683M documents from Baidu Search, combining both: (1) a rich prior-attribute set with detailed textual relevance, authority, recency, and quality features, and (2) a comprehensive posterior-attribute set enriched by user click data, dwell time, and positional information. By unifying these dual perspectives, FULTR establishes a robust, reproducible benchmark for satisfaction-oriented ranking, enabling researchers to develop models that better capture real-world search behaviors and user satisfaction. In addition, we propose a strong LTR baseline that merges a satisfaction ranker that leverages pre-trained language models to integrate diverse satisfaction signals, with a behavior ranker that captures user interactions using a dual-tower approach. Their outputs are combined via a fusion layer, yielding significant performance gains in multiple evaluation metrics, as confirmed by extensive experiments and ablation studies. We are confident that our contribution not only democratizes access to industrial-grade fusion data for the research community but also paves the way for more effective and holistic LTR model design. FULTR is available to the research community at https://github.com/zhanghao731/FULTR.

The proliferation of complex non-factoid questions in modern information seeking (IS) systems exposes critical limitations in conventional Retrieval-Augmented Generation (RAG) approaches, particularly their static search strategies and the lack of systematic multi-source information integration capabilities. Facing these limitations, we present PASS (Proactive Agent-driven Search System), a novel multi-agent framework that operationalizes human-like proactive search strategies through five specialized agents: Revealer for intent analysis, Navigator for search planning, Seeker/Reader for adaptive retrieval, and Writer for response synthesis, systematically expanding the search space through iterative query refinement and multi-perspective knowledge integration. Crucially, our framework demonstrates remarkable adaptability to mid-sized LLMs, demonstrating its scalability in resource-constrained environments. To comprehensively assess the effectiveness of the proposed framework, we carry out extensive experiments on both mid-sized and proprietary large-scale LLMs, evaluating response quality for complex non-factoid questions using a newly introduced nugget-based assessment. Experimental results from offline nugget-based evaluation and online A/B Tests confirm substantial improvements in answer quality, advancing proactive information seeking methodologies and offering practical pathways for democratizing complex reasoning capabilities to resource-constrained environments.

While representation learning has been used to boost the performance of Learning-to-Rank (LTR) models through distilling key features for webpage ranking, the weak supervision signals extracted from users' sparse click-through data lead to inadequate representation of query-webpage pairs for ranking score prediction. Recent studies in generative LTR pre-training demonstrate the feasibility of incorporating reconstruction loss for enhanced ranking score prediction. However, LTR is afterall a regression task and it might be reasonable to find an alternate route that pre-trains LTR models with discriminative losses. Following the success of Electra in representation learning for natural language processing (NLP), this work proposes RankElectra that pre-trains the LTR model as a discriminator module inside a generative learning framework. Specifically, RankElectra first structures sparsely-annotated query-webpage pairs into a bipartite graph, with query and webpage feature vectors as node types and ranking scores as the connecting edges, and then leverages positive and negative extension strategies to densify the graph by link predictions. Later, this work proposes a novel Electra module that pre-trains the LTR model as a discriminator module for node reconstruction tasks, where node features of selected edges would be randomly masked and reconstructed by a generator, and the discriminator learns to classify whether the reconstructed features are the original or replaced as well as perform correct ranking. Finally, the pre-trained discriminator module, rather than the generator, would be fine-tuned on the labeled graph. We carried out extensive offline and online evaluations using the real-world web traffic of Baidu search engine. The results show that RankElectra could significantly boost the ranking performance of Baidu Search compared with numbers of competitor systems.

Sequential recommendation systems aim to predict the future behaviors of users based on their historical interactions. Despite the success of neural architectures like Transformer and Graph Neural Networks, these models often struggle with the inherent challenge of sparse data in accurately predicting future user behaviors. To alleviate the data sparsity problem, some methods leverage the contrastive learning to generate contrastive views, assuming the items appear discretely at the same time intervals and focusing on the sequence order. However, these approaches neglect the crucial temporal-aware collaborative patterns hidden within the user-item interactions, leading to a limited variety of contrastive pairs and less informative embeddings. The proposed framework,Temporal-awaregraph contrastive learning with theoretical guarantees for sequentialRecommendation (TagRec), integrates temporal-aware collaborative patterns with adaptive data augmentation to generate more informative user and item representations. TagRec employs a temporal-aware graph neural network to embed the original graph, then generates augmented graphs through the addition of interactions via latent user interest mining, the dropping of redundant interaction edges, and the perturbation of temporal information. Theoretical guarantees are provided that these augmentations enhance the graph’s utility. Extensive experiments on real-world datasets demonstrate the superiority of the proposed approach over the state-of-the-art recommendation methods.

While learning to rank (LTR) is widely employed in web searches to prioritize pertinent webpages from the retrieved contents based on input queries, traditional LTR models stumble over two principal stumbling blocks leading to subpar performance: 1) the lack of well-annotated query-webpage pairs with ranking scores to cover search queries of various popularity, debilitating their coverage of search queries across the popularity spectrum, and 2) ill-trained models that are incapable of inducing generalized representations for LTR, culminating in overfitting. To tackle the above challenges, we proposed a Generative Semi-Supervised Pre-trained (GS2P) Learning to Rank model. Specifically, GS2P first generates pseudo-labels for the unlabeled samples using tree-based LTR models after a series of co-training procedures, then learns the representations of query-webpage pairs with self-attentive transformers via both discriminative (LTR) and generative (denoising autoencoding for reconstruction) losses. Finally, GS2P boosts the performance of LTR through incorporating Random Fourier Features to over-parameterize the models into “interpolating regime”, so as to enjoy the further descent of generalization errors with learned representations. We conduct extensive offline experiments on a publicly available dataset and a real-world dataset collected from a large-scale search engine. The results show that GS2P can achieve the best performance on both datasets, compared to baselines. We also deploy GS2P at a large-scale web search engine with realistic traffic, where we can still observe significant improvement in real-world applications. GS2P performs consistently in both online and offline experiments.

Both Transformer and Graph Neural Networks (GNNs) have been used in learning to rank (LTR), they however adhere to two distinct-yet-complementary problem formulations, i.e., ranking score regression based on query-webpage pairs and link prediction within query-webpage bipartite graphs, respectively. Though it is possible to pre-train GNNs or Transformers on source datasets and fine-tune them subject to sparsely annotated LTR datasets separately, the source-target distribution shifts across the pairs and bipartite graphs domains make it extremely difficult to integrate these diverse models into a single LTR framework at a web-scale. We introduce the novel MPGraf model, which utilizes a modular and capsule-based pre-training approach, aiming to incorporate regression capacities from Transformers and link prediction capabilities of GNNs cohesively. Specifically, rather than simply combining these two modules in stacking or parallelizing architectures, MPGraf proposes a three-step method to (1) construct query-webpage graphs from sparsely annotated query-webpage pairs, (2) pre-train the Transformer and GNN modules using source datasets from two domains, and (3) “surgically” fine-tune the integrated models with pretrained weights while handling the distribution shifts of two domains. Extensive experiments have been done to evaluate the performance of MPGraf using real-world datasets collected from large-scale search engines. The results show that MPGraf can outperform baseline algorithms on several major metrics. Further, we deploy and evaluate MPGraf atop a large-scale search engine with realistic web traffics via A/B tests, where we can still observe significant improvement. MPGraf performs consistently in both offline and online evaluations.

While Learning to Rank (LTR) models on top of transformers have been widely adopted to achieve decent performance, it is still challenging to train the model with sufficient data as only an extremely small number of query-webpage pairs could be annotated versus trillions of webpages available online and billions of web search queries everyday. In the meanwhile, industry research communities have released a number of open-source LTR datasets with well annotations but incorporating different designs of LTR features/labels (i.e., heterogeneous domains). In this work, inspired by the recent progress in pre-training transformers for performance advantages, we study the problem of pre-training LTR models using both labeled and unlabeled samples, especially we focus on the use of well-annotated samples in heterogeneous open-source LTR datasets to boost the performance of pre-training. Hereby, we propose S2phere-Semi-Supervised Pre-training with Heterogeneous LTR data strategies for LTR models using both unlabeled and labeled query-webpage pairs across heterogeneous LTR datasets. S2phere consists of a three-step approach: (1) Semi-supervised Feature Extraction Pre-training via Perturbed Contrastive Loss, (2) Cross-domain Ranker Pre-training over Heterogeneous LTR Datasets and (3) End-to-end LTR Fine-tuning via Modular Network Composition. Specifically, given an LTR model composed of a backbone (the feature extractor), a neck (the module to reason the orders) and a head (the predictor of ranking scores), S2phere uses unlabeled/labeled data from the search engine to pre-train the backbone in Step (1) via semi-supervised learning; then Step (2) incorporates multiple open-source heterogeneous LTR datasets to improve pre-training of the neck module as shared parameters of cross-domain learning; and finally, S2phere in Step (3) composes the backbone and neck with a randomly-initialized head into a whole LTR model and fine-tunes the model using search engine data with various learning strategies. Extensive experiments have been done with both offline experiments and online A/B Test on top of Baidu search engine. The comparisons against numbers of baseline algorithms confirmed the advantages of S2phere in producing high-performance LTR models for web-scale search.

Whilelearning to rank(LTR) has been widely used in web search to prioritize most relevant webpages among the retrieved contents subject to the input queries, the traditional LTR models fail to deliver decent performance due to two main reasons: 1) the lack of well-annotated query-webpage pairs with ranking scores to cover search queries of various popularity, and 2) ill-trained models based on a limited number of training samples with poor generalization performance. To improve the performance of LTR models, tremendous efforts have been done from above two aspects, such as enlarging training sets with pseudo-labels of ranking scores by self-training, or refining the features used for LTR through feature extraction and dimension reduction. Though LTR performance has been marginally increased, we still believe these methods could be further improved in the newly-fashioned “interpolating regime”. Specifically, instead of lowering the number of features used for LTR models, our work proposes to transform original data with random Fourier feature, so as to over-parameterize the downstream LTR models (e.g., GBRank or LightGBM) with features in ultra-high dimensionality and achieve superb generalization performance. Furthermore, rather than self-training with pseudo-labels produced by the same LTR model in a “self-tuned” fashion, the proposed method incorporates the diversity of prediction results between the listwise and pointwise LTR models while co-training both models with a cyclic labeling-prediction pipeline in a “ping-pong” manner. We deploy the proposedCo-trained andOver-parameterizedLTRsystemCOLTRat Baidu search and evaluateCOLTRwith a large number of baseline methods. The results show thatCOLTRcould achieve$\Delta NDCG_{4}$= 3.64%$\sim$4.92%, compared to baselines, under various ratios of labeled samples. We also conduct a 7-day A/B Test using the realistic web traffics of Baidu Search, where we can still observe significant performance improvement around$\Delta NDCG_{4}$= 0.17%$\sim$0.92% in real-world applications.COLTRperforms consistently both in online and offline experiments.