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.

Deep-research agents are capable of executing multi-step web exploration, targeted retrieval, and sophisticated question answering. Despite their powerful capabilities, deep-research agents face two critical bottlenecks: (1) the lack of large-scale, challenging datasets with real-world difficulty, and (2) the absence of accessible, open-source frameworks for data synthesis and agent training. To bridge these gaps, we first construct DeepResearch-9K, a large-scale challenging dataset specifically designed for deep-research scenarios built from open-source multi-hop question-answering (QA) datasets via a low-cost autonomous pipeline. Notably, it consists of (1) 9000 questions spanning three difficulty levels from L1 to L3 (2) high-quality search trajectories with reasoning chains from Tongyi-DeepResearch-30B-A3B, a state-of-the-art deep-research agent, and (3) verifiable answers. Furthermore, we develop an open-source training framework DeepResearch-R1 that supports (1) multi-turn web interactions, (2) different reinforcement learning (RL) approaches, and (3) different reward models such as rule-based outcome reward and LLM-as-Judge feedback. Finally, empirical results demonstrate that agents trained on DeepResearch-9K under our DeepResearch-R1 achieve state-of-the-art results on challenging deep-research benchmarks. We release the DeepResearch-9K dataset on https://huggingface.co/datasets/artillerywu/DeepResearch-9K and the code of DeepResearch-R1 on https://github.com/Applied-Machine-Learning-Lab/SIGIR2026_DeepResearch-R1.

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.

Generative retrieval (GR) reformulates the Information Retrieval (IR) task as the generation of document identifiers (docIDs). Despite its promise, existing GR models exhibit poor generalization to newly added documents, often failing to generate the correct docIDs. While incremental training offers a straightforward remedy, it is computationally expensive, resource-intensive, and prone to catastrophic forgetting, thereby limiting the scalability and practicality of GR.

Retrieval-augmented Generation (RAG) integrates Large Language Models (LLMs) with retrievers to access external knowledge, improving the factuality of LLM generation in knowledge-grounded tasks. To optimize the RAG performance, most previous work independently fine-tunes the retriever to adapt to frozen LLMs or trains the LLMs to use documents retrieved by off-the-shelf retrievers, lacking end-to-end training supervision. Recent work addresses this limitation by jointly training these two components but relies on overly simplifying assumptions of document independence, which has been criticized for being far from real-world scenarios. Thus, effectively optimizing the overall RAG performance remains a critical challenge. We propose a Direct Retrieval-augmented Optimization ( DRO ) framework that enables end-to-end training of two key components: (i) a generative knowledge selection model and (ii) an LLM generator. DRO alternates between two phases: (i) document permutation estimation and (ii) re-weighted maximization, progressively improving RAG components through a variational approach. In the estimation step, we treat document permutation as a latent variable and directly estimate its distribution from the selection model by applying an importance sampling strategy. In the maximization step, we calibrate the optimization expectation using importance weights and jointly train the selection model and LLM generator. Our theoretical analysis reveals that DRO is analogous to policy-gradient methods in reinforcement learning. Extensive experiments conducted on five datasets illustrate that DRO outperforms the best baseline with 5–15% improvements in EM and F1. We also qualitatively analyze the stability, convergence, and variance of DRO. (Code is available on DRO GitHub ).

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.

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.

Visual Document Retrieval (VDR) is an emerging research area that focuses on encoding and retrieving document images directly, bypassing the dependence on Optical Character Recognition (OCR) for document search. A recent advance in VDR was introduced by ColPali, which significantly improved retrieval effectiveness through a late interaction mechanism. ColPali's approach demonstrated substantial performance gains over existing baselines that do not use late interaction on an established benchmark. In this study, we investigate the reproducibility and replicability of VDR methods with and without late interaction mechanisms by systematically evaluating their performance across multiple pre-trained vision-language models. Our findings confirm that late interaction yields considerable improvements in retrieval effectiveness; however, it also introduces computational inefficiencies during inference. Additionally, we examine the adaptability of VDR models to textual inputs and assess their robustness across text-intensive datasets within the proposed benchmark, particularly when scaling the indexing mechanism. Furthermore, our research investigates the specific contributions of late interaction by looking into query-patch matching in the context of visual document retrieval. We find that although query tokens cannot explicitly match image patches as in the text retrieval scenario, they tend to match the patch contains visually similar tokens or their surrounding patches.

Generative retrieval (GR) has emerged as a promising paradigm in information retrieval (IR). However, most existing GR models are developed and evaluated using a static document collection, and their performance in dynamic corpora where document collections evolve continuously is rarely studied. In this paper, we first reproduce and systematically evaluate various representative GR approaches over dynamic corpora. Through extensive experiments, we reveal that existing GR models with text-based docids show superior generalization to unseen documents. We observe that the more fine-grained the docid design in the GR model, the better its performance over dynamic corpora, surpassing BM25 and even being comparable to dense retrieval methods. While GR models with numeric-based docids show high efficiency, their performance drops significantly over dynamic corpora. Furthermore, our experiments find that the underperformance of numeric-based docids is partly due to their excessive tendency toward the initial document set, which likely results from overfitting on the training set. We then conduct an in-depth analysis of the best-performing GR methods. We identify three critical advantages of text-based docids in dynamic corpora: (i) Semantic alignment with language models' pretrained knowledge (ii) Fine-grained docid design, and (iii) High lexical diversity. Building on these insights, we finally propose a novel multi-docid design that leverages both the efficiency of numeric-based docids and the effectiveness of text-based docids, achieving improved performance in dynamic corpus without requiring additional retraining. Our work offers empirical evidence for advancing GR methods over dynamic corpora and paves the way for developing more generalized yet efficient GR models in real-world search engines.

Large Language Models (LLMs) are increasingly employed in zero-shot documents ranking, yielding commendable results. However, several significant challenges still persist in LLMs for ranking: (1) LLMs are constrained by limited input length, precluding them from processing a large number of documents simultaneously; (2) The output document sequence is influenced by the input order of documents, resulting in inconsistent ranking outcomes; (3) Achieving a balance between cost and ranking performance is challenging. To tackle these issues, we introduce a novel documents ranking method called TourRank1. which is inspired by the sport tournaments, such as FIFA World Cup. Specifically, we 1) overcome the limitation in input length and reduce the ranking latency by incorporating a multi-stage grouping strategy similar to the parallel group stage of sport tournaments; 2) improve the ranking performance and robustness to input orders by using a points system to ensemble multiple ranking results. We test TourRank with different LLMs on the TREC DL datasets and the BEIR benchmark. The experimental results demonstrate that TourRank delivers state-of-the-art performance at a modest cost.

Grounded on pre-trained language models (PLMs), dense retrieval has been studied extensively on plain text. In contrast, there has been little research on retrieving data with multiple aspects using dense models. In the scenarios such as product search, the aspect information plays an essential role in relevance matching, e.g., category: Electronics, Computers, and Pet Supplies. A common way of leveraging aspect information for multi-aspect retrieval is to introduce an auxiliary classification objective, i.e., using item contents to predict the annotated value IDs of item aspects. However, by learning the value embeddings from scratch, this approach may not capture the various semantic similarities between the values sufficiently. To address this limitation, we leverage the aspect information as text strings rather than class IDs during pre-training so that their semantic similarities can be naturally captured in the PLMs. To facilitate effective retrieval with the aspect strings, we propose mutual prediction objectives between the text of the item aspect and content. In this way, our model makes more sufficient use of aspect information than conducting undifferentiated masked language modeling (MLM) on the concatenated text of aspects and content. Extensive experiments on two real-world datasets (product and mini-program search) show that our approach can outperform competitive baselines both treating aspect values as classes and conducting the same MLM for aspect and content strings. Code and related dataset will be available at the URL \footnotehttps://github.com/sunxiaojie99/ATTEMPT.

Pre-training and fine-tuning have achieved significant advances in the information retrieval (IR). A typical approach is to fine-tune all the parameters of large-scale pre-trained models (PTMs) on downstream tasks. As the model size and the number of tasks increase greatly, such approach becomes less feasible and prohibitively expensive. Recently, a variety of parameter-efficient tuning methods have been proposed in natural language processing (NLP) that only fine-tune a small number of parameters while still attaining strong performance. Yet there has been little effort to explore parameter-efficient tuning for IR.

Dense retrieval (DR) has shown promising results in information retrieval. In essence, DR requires high-quality text representations to support effective search in the representation space. Recent studies have shown that pre-trained autoencoder-based language models with a weak decoder can provide high-quality text representations, boosting the effectiveness and few-shot ability of DR models. However, even a weak autoregressive decoder has the bypass effect on the encoder. More importantly, the discriminative ability of learned representations may be limited since each token is treated equally important in decoding the input texts. To address the above problems, in this paper, we propose a contrastive pre-training approach to learn a discriminative autoencoder with a lightweight multi-layer perception (MLP) decoder. The basic idea is to generate word distributions of input text in a non-autoregressive fashion and pull the word distributions of two masked versions of one text close while pushing away from others. We theoretically show that our contrastive strategy can suppress the common words and highlight the representative words in decoding, leading to discriminative representations. Empirical results show that our method can significantly outperform the state-of-the-art autoencoder-based language models and other pre-trained models for dense retrieval.

Dense retrieval has shown promising results in many information retrieval (IR) related tasks, whose foundation is high-quality text representation learning for effective search. Some recent studies have shown that autoencoder-based language models are able to boost the dense retrieval performance using a weak decoder. However, we argue that 1) it is not discriminative to decode all the input texts and, 2) even a weak decoder has the bypass effect on the encoder. Therefore, in this work, we introduce a novel contrastive span prediction task to pre-train the encoder alone, but still retain the bottleneck ability of the autoencoder. In this way, we can 1) learn discriminative text representations efficiently with the group-wise contrastive learning over spans and, 2) avoid the bypass effect of the decoder thoroughly. Comprehensive experiments over publicly available retrieval benchmark datasets show that our approach can outperform existing pre-training methods for dense retrieval significantly.

Pre-training and fine-tuning have achieved remarkable success in many downstream natural language processing (NLP) tasks. Recently, pre-training methods tailored for information retrieval (IR) have also been explored, and the latest success is the PROP method which has reached new SOTA on a variety of ad-hoc retrieval benchmarks. The basic idea of PROP is to construct therepresentative words prediction (ROP) task for pre-training inspired by the query likelihood model. Despite its exciting performance, the effectiveness of PROP might be bounded by the classical unigram language model adopted in the ROP task construction process. To tackle this problem, we propose a bootstrapped pre-training method (namely B-PROP) based on BERT for ad-hoc retrieval. The key idea is to use the powerful contextual language model BERT to replace the classical unigram language model for the ROP task construction, and re-train BERT itself towards the tailored objective for IR. Specifically, we introduce a novel contrastive method, inspired by the divergence-from-randomness idea, to leverage BERT's self-attention mechanism to sample representative words from the document. By further fine-tuning on downstream ad-hoc retrieval tasks, our method achieves significant improvements over PROP and other baselines, and further pushes forward the SOTA on a variety of ad-hoc retrieval tasks.

Recently pre-trained language representation models such as BERT have shown great success when fine-tuned on downstream tasks including information retrieval (IR). However, pre-training objectives tailored for ad-hoc retrieval have not been well explored. In this paper, we propose Pre-training with Representative wOrds Prediction (PROP) for ad-hoc retrieval. PROP is inspired by the classical statistical language model for IR, specifically the query likelihood model, which assumes that the query is generated as the piece of text representative of the "ideal" document. Based on this idea, we construct the representative words prediction (ROP) task for pre-training. Given an input document, we sample a pair of word sets according to the document language model, where the set with higher likelihood is deemed as more representative of the document. We then pre-train the Transformer model to predict the pairwise preference between the two word sets, jointly with the Masked Language Model (MLM) objective. By further fine-tuning on a variety of representative downstream ad-hoc retrieval tasks, PROP achieves significant improvements over baselines without pre-training or with other pre-training methods. We also show that PROP can achieve exciting performance under both the zero- and low-resource IR settings.

Relevance plays a central role in information retrieval (IR), which has received extensive studies starting from the 20th century. The definition and the modeling of relevance has always been critical challenges in both information science and computer science research areas. Along with the debate and exploration on relevance, IR has already become a core task in many real-world applications, such as Web search engines, question answering systems, conversational bots, and so on. While relevance acts as a unified concept in all these retrieval tasks, the inherent definitions are quite different due to the heterogeneity of these tasks. This raises a question to us: Do these different forms of relevance really lead to different modeling focuses? To answer this question, in this work, we conduct an empirical study on relevance modeling in three representative IR tasks, i.e., document retrieval, answer retrieval, and response retrieval. Specifically, we attempt to study the following two questions: 1) Does relevance modeling in these tasks really show differences in terms of natural language understanding (NLU)? We employ 16 linguistic tasks to probe a unified retrieval model over these three retrieval tasks to answer this question. 2) If there do exist differences, how can we leverage the findings to enhance the relevance modeling? We proposed three intervention methods to investigate how to leverage different modeling focuses of relevance to improve these IR tasks. We believe the way we study the problem as well as our findings would be beneficial to the IR community.