Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by leveraging external knowledge, where retrieval accuracy directly affects generation quality. However, dense retrievers, commonly employed in RAG, suffer degraded performance in evolving corpora where new documents arrive continuously and distribution shifts accumulate over time. In such settings, continually updating retrievers is crucial, yet conventional retraining is computationally expensive and often impractical. To address this challenge, we propose FlowRAG, a lightweight and effective method for continual retriever adaptation in evolving corpora. FlowRAG augments the encoder with Layer-wise Prompt Embeddings and introduces a Cross-Layer Fusion mechanism to capture hierarchical semantic representations. In addition, a novel Generator-Guided Loss aligns retriever scores and intermediate representations with the LLM's generation likelihoods, encouraging retrieval decisions that are both semantically relevant and beneficial for generation. Experiments on datasets spanning four domains demonstrate that FlowRAG, which updates only about 0.64% of the total model parameters, consistently outperforms strong baselines in retrieval accuracy, generation quality, and robustness to forgetting in non-stationary settings.
论文检索
输入标题、作者或关键词，从 954 篇学术成果中精准定位
Composed Image Retrieval (CIR) aims to retrieve a target image from a query composed of a reference image and modification text. Recent training-free zero-shot methods often employ Multimodal Large Language Models (MLLMs) with Chain-of-Thought (CoT) to compose a target image description for retrieval. However, due to the fuzzy matching nature of ZS-CIR, the generated description is prone to semantic bias relative to the target image. We propose SDR-CIR, a training-free Semantic Debias Ranking method based on CoT reasoning. First, Selective CoT guides the MLLM to extract visual content relevant to the modification text during image understanding, thereby reducing visual noise at the source. We then introduce a Semantic Debias Ranking with two steps, Anchor and Debias, to mitigate semantic bias. In the Anchor step, we fuse reference image features with target description features to reinforce useful semantics and supplement omitted cues. In the Debias step, we explicitly model the visual semantic contribution of the reference image to the description and incorporate it into the similarity score as a penalty term. By supplementing omitted cues while suppressing redundancy, SDR-CIR mitigates semantic bias and improves retrieval performance. Experiments on three standard CIR benchmarks show that SDR-CIR achieves state-of-the-art results among one-stage methods while maintaining high efficiency. The code is publicly available at https://github.com/suny105/SDR-CIR.
Multimodal Retrieval-Augmented Generation (MM-RAG) extends the capabilities of Large Language Models (LLMs) by incorporating external image-text knowledge bases to handle various tasks. However, MM-RAG systems in open environments are highly vulnerable to retrieval poisoning attacks, i.e., adversaries can inject malicious image-text pairs that are retrieved and dominate the generation process, leading to incorrect or harmful outputs. Due to the unique challenges of image-text fusion and cross-modal interference, existing defenses for text-based RAG cannot be directly applied to multimodal scenarios. In this paper, we propose IRAG, the first robust defense framework specifically designed for MM-RAG. The core of IRAG lies in its hazard separation. This structured defense isolates potential contamination sources by leveraging redundancy and consensus, enhancing system robustness and ensuring reliable outputs even when portions of retrieved content are compromised. Extensive experiments conducted under the MMQA and WebQA and the BQI and ROTI poisoning schemes demonstrate that IRAG consistently restores system reliability: the normal answer accuracy improves by 15–30% (restoring it to pre-poisoning levels), while the poisoned answer rate is reduced to below 7%.
Recently, Agentic Reinforcement Learning (Agentic RL) has made significant progress in incentivizing the multi-turn, long-horizon tool-use capabilities of web agents. While mainstream agentic RL algorithms autonomously explore high-uncertainty tool-call steps under the guidance of entropy, excessive reliance on such signals leads to the training collapse. In this paper, we delve into the challenges caused by entropy and propose the Agentic Entropy-Balanced Policy Optimization (AEPO), an agentic RL algorithm designed to balance entropy in both the rollout and policy update phases. AEPO comprises two core components: (1) a dynamic entropy-balanced rollout mechanism that adaptively allocates global and branch sampling budget through entropy pre-monitoring, while imposing a branch penalty on consecutive high-entropy tool-call steps to prevent over-branching issues; and (2) Entropy-Balanced Policy Optimization that inserts a stop-gradient operation into the high-entropy clipping term to preserve and properly rescale gradients on high-entropy tokens, while incorporating entropy-aware advantage estimation to prioritize learning on high-uncertainty tokens. Results across 14 challenging datasets show that AEPO consistently outperforms 7 mainstream RL algorithms. Further analysis reveals that AEPO improves rollout sampling diversity while maintaining stable policy entropy, facilitating scalable training of general web agents. The code is available at https://github.com/RUC-NLPIR/ARPO.
Retrieval-Augmented Generation (RAG) effectively mitigates hallucinations and knowledge gaps in Large Language Models (LLMs) for knowledge-intensive tasks by incorporating external web-based knowledge. However, when integrating diverse yet potentially conflicting web-sourced information, RAG systems are prone to knowledge conflicts that manifest as incorrect or inconsistent model behaviors, ultimately leading to unreliable responses. To address this challenge, this paper proposes Conflict-Aware RAG, a general training framework that leverages the model's inherent conflict-sensing capability to build a more robust RAG system via phased optimization. At the core of this framework lies ConScore, a conflict signal that quantifies the model's awareness of potential knowledge conflicts by comparing generative probabilities across distinct knowledge sources. This signal then guides both the construction of training data and a multi-stage optimization workflow: In the Supervised Fine-Tuning (SFT) stage, conflict features are employed to select representative distracting documents, laying the groundwork for core RAG capabilities; in the Direct Preference Optimization (DPO) stage, high-quality preference pairs are constructed using the conflict signal to boost the model's robustness against distracting knowledge; and in the Reranking stage, conflict confidence and information gain are integrated to synergistically optimize the collaboration mechanism between the retriever and LLM. Experiments on six knowledge-intensive question answering (QA) datasets demonstrate that Conflict-Aware RAG significantly outperforms mainstream baselines. Further ablation studies and quantitative analyses validate the method's stability and generalization, laying the foundation for robust RAG systems.
Composed Image Retrieval (CIR) aims to retrieve target images based on a reference image and modified texts. However, existing methods often struggle to extract the correct semantic cues from the reference image that best reflect the user's intent under textual modification prompts, resulting in interference from irrelevant visual noise. In this paper, we propose a novel Multi-level Vision Selection by Multi-modal Chain-of-Thought Reasoning (MCoT-MVS) for CIR, integrating attention-aware multi-level vision features guided by reasoning cues from a multi-modal large language model (MLLM). Specifically, we leverage an MLLM to perform chain-of-thought reasoning on the multimodal composed input, generating the retained, removed, and target-inferred texts. These textual cues subsequently guide two reference visual attention selection modules to selectively extract discriminative patch-level and instance-level semantics from the reference image. Finally, to effectively fuse these multi-granular visual cues with the modified text and the imagined target description, we design a weighted hierarchical combination module to align the composed query with target images in a unified embedding space. Extensive experiments on two CIR benchmarks, namely CIRR and FashionIQ, demonstrate that our approach consistently outperforms existing methods and achieves new state-of-the-art performance. Code and trained models are publicly released at https://github.com/JJJJerry/WWW2026-MCoT-MVS.
The rapid spread of misinformation on social media has underscored the importance of automatic fact-checking. Existing fact-checking pipelines typically rely on multi-stage frameworks involving evidence retrieval and claim verification. However, these methods face two major challenges: (1) the retrieval process often introduces noisy evidence, which compromises the reliability of the final veracity prediction; and (2) the verification models may overlook critical factual details, resulting in hallucinated conclusions. To address these issues, we propose a fact-checking framework SLED with Self-supervised denoising evidence retrieval and LLM-Enhanced Debate-based verification. In the retrieval stage, SLED leverage trained verifier to assess credibility and necessity of retrieved evidence, enabling the elimination of noisy evidence. In the verification stage, SLED prompts the LLM to generate dual-perspective reasoning and simulates a multi-agent debate, followed by distillation into a lightweight model for final veracity prediction. Experiments on CHEF and HOVER datasets demonstrate that SLED achieves the state-of-the-art results in complex fact verification scenarios.
Large language model (LLM) based answer engines offer real-time, conversational answers but raise concerns about scalability and web sustainability. Unlike traditional search that serves results from cached indices, these systems fetch pages anew for each query, creating redundant network traffic. We present the first measurement-driven study of commercial LLM answer engines, combining automated client-side tracing and controlled server-side audits. Analyzing ChatGPT and Claude across 1,000 queries, we find that both operate as meta-search layers heavily reliant on existing indices, fetching top-ranked pages with minimal caching. A human-equivalent cost model shows their per-query network footprint far exceeds that of human searchers, varying sharply by architecture. These results reveal the infrastructural burden of real-time fetching and motivate cooperative efficiency measures like shared caches, transparent retrieval standards, and publisher controls such as llms.txt, to make AI-augmented search more sustainable.
Retrieval is being redefined by agentic AI, demanding multimodal reasoning beyond conventional similarity-based paradigms. Composed Image Retrieval (CIR) exemplifies this shift as each query combines a reference image with textual modifications, requiring compositional understanding across modalities. While embedding-based CIR methods have achieved progress, they remain narrow in perspective, capturing limited cross-modal cues and lacking semantic reasoning. To address these limitations, we introduce XR, a training-free multi-agent framework that reframes retrieval as a progressively coordinated reasoning process. It orchestrates three specialized types of agents: imagination agents synthesize target representations through cross-modal generation, similarity agents perform coarse filtering via hybrid matching, and question agents verify factual consistency through targeted reasoning for fine filtering. Through progressive multi-agent coordination, XR iteratively refines retrieval to meet both semantic and visual query constraints, achieving up to a 38% gain over strong training-free and training-based baselines on FashionIQ, CIRR, and CIRCO, while ablations show each agent is essential. Code is available: https://01yzzyu.github.io/xr.github.io/.
Large Language Model (LLM)-based agentic systems have shown strong capabilities across various tasks. However, existing multi-agent frameworks often rely on static or task-level workflows, which either over-process simple queries or underperform on complex ones, while also neglecting the efficiency-performance trade-offs across heterogeneous LLMs. To address these limitations, we propose Difficulty-Aware Agentic Orchestration (DAAO), which can dynamically generate query-specific multi-agent workflows guided by predicted query difficulty. DAAO comprises three interdependent modules: a variational autoencoder (VAE) for difficulty estimation, a modular operator allocator, and a cost- and performance-aware LLM router. A self-adjusting policy updates difficulty estimates based on workflow success, enabling simpler workflows for easy queries and more complex strategies for harder ones. Experiments on six benchmarks demonstrate that DAAO surpasses prior multi-agent systems in both accuracy and inference efficiency, validating its effectiveness for adaptive, difficulty-aware reasoning. Our code is open-sourced at https://github.com/AutoAgents-ai/DAAO
Deep search agents, which autonomously iterate through multi-turn web-based reasoning, represent a promising paradigm for complex information-seeking tasks. However, current agents suffer from critical inefficiency: they conduct excessive searches as they cannot accurately judge when to stop searching and start answering. This stems from outcome-centric training that prioritize final results over the search process itself. We identify the root cause as misaligned decision boundaries, the threshold determining when accumulated information suffices to answer. This causes over-search (redundant searching despite sufficient knowledge) and under-search (premature termination yielding incorrect answers). To address these errors, we propose a comprehensive framework comprising two key components. First, we introduce causal intervention-based diagnosis that identifies boundary errors by comparing factual and counterfactual trajectories at each decision point. Second, we develop Decision Boundary Alignment for Deep Search agents (DAS), which constructs preference datasets from causal feedback and aligns policies via preference optimization. Experiments on public datasets demonstrate that decision boundary errors are pervasive across state-of-the-art agents. Our DAS method effectively calibrates these boundaries, mitigating both over-search and under-search to achieve substantial gains in accuracy and efficiency. Our code and data are publicly available at: https://github.com/Applied-Machine-Learning-Lab/WWW2026_DAS.
Existing multimodal document question-answering (QA) systems predominantly rely on flat semantic retrieval, representing documents as a set of disconnected text chunks and largely neglecting their intrinsic hierarchical and relational structures. Such flattening disrupts logical and spatial dependencies—such as section organization, figure-text correspondence, and cross-reference relations—that humans naturally exploit for comprehension. To address this limitation, we introduce a document-level structural Document MAP (DMAP), which explicitly encodes both hierarchical organization and inter-element relationships within multimodal documents. Specifically, we design a Structured-Semantic Understanding Agent to construct DMAP by organizing textual content together with figures, tables, charts, etc into a human-aligned hierarchical schema that captures both semantic and layout dependencies. Building upon this representation, a Reflective Reasoning Agent performs structure-aware and evidence-driven reasoning, dynamically assessing the sufficiency of retrieved context and iteratively refining answers through targeted interactions with DMAP. Extensive experiments on MMDocQA benchmarks demonstrate that DMAP yields document-specific structural representations aligned with human interpretive patterns, substantially enhancing retrieval precision, reasoning consistency, and multimodal comprehension over conventional RAG-based approaches. Code is available at https://github.com/Forlorin/DMAP
Text-to-SQL technology converts natural language queries into SQL statements for database retrieval. Recent advances in large language models (LLMs) have improved Text-to-SQL performance, but generated SQL often contains semantic or syntax errors that degrade user experience and system stability. Existing SQL error detection methods are costly, lack interpretability, and do not support error labeling. To overcome these issues, we propose SQL-Checker a specialized model for Text-to-SQL error detection. We first analyze common error factors in Text-to-SQL, and we design a novel data synthesis framework based on these error factors. This framework simulates error factors to construct a basic error SQL data, and then using an error analysis template to distill high-quality SQL error analysis data from large-scale models. For complex errors, a self-guided iterative distillation strategy further enhances data quality. SQL-Checker is then trained on this distilled dataset. Additionally, we refine SQL error labeling and, integrate error label recognition into the detection task, enabling macro-level cause analysis. Experiments show SQL-Checker achieves state-of-the-art results on multiple error detection datasets. Incorporating SQL-Checker into the Text-to-SQL pipeline also improves execution accuracy.
Approximate nearest neighbor (ANN) searches are commonly employed in various machine learning applications, such as recommendation systems, but traditional ANN searches typically involve only a single reference vector in a query. To broaden the capabilities of ANN search and support multi-reference-vector queries, thereby enabling a wider range of machine learning applications, we introduce all/any-k ANN search. They aim to find vectors that are similar to all or any of the multi-reference vectors in a query, respectively. To effectively and efficiently support all/any-k ANN search, we first propose distance metrics to evaluate the ranking of vectors among those in the dataset for exact all/any-k NN. Building on this, we introduce search algorithms and prove they can search according to the proposed distance metrics on graph indexes designed for traditional ANN. Additionally, we further introduce two-stage search algorithms for all/any-k ANN search to further enhance their search performance. We conduct extensive experiments on real-world datasets to validate the efficiency and effectiveness of our proposed algorithms compared to existing approaches.
Large language models (LLMs) have demonstrated significant potential in listwise document reranking. Due to their limited context length, LLM-based listwise reranking methods often rely on a sliding window strategy that only processes a small subset of documents at a time. While effective, this approach lacks interactions between documents, increases computational overhead, and results in significant API costs. It is crucial to develop long-context LLMs for enabling the full ranking of all documents in one pass. We propose a long-context listwise document reranker, LongRanker, and make two major contributions to enable long-context LLMs for listwise reranking: (i) To improve length extrapolation for listwise inputs, we introduce an intra-inter hierarchical positional encoding approach that combines intra-document encoding to identify token locations within a document with inter-document encoding to specify the document index. (ii) To efficiently model long contexts with reduced computational costs, we introduce a top-k attention pruning mechanism; it ensures that the model attends to the query and the most important documents while disregarding less relevant ones. Extensive empirical results demonstrate that LongRanker achieves performance comparable to state-of-the-art rerankers while significantly improving efficiency through reduced LLM calls.
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. In this paper, we identify the core bottleneck as the decoder's ability to map hidden states to the correct docIDs of newly added documents. Model editing, which enables targeted parameter modifications for docID mapping, represents a promising solution. However, applying model editing to current GR models is not trivial, which is severely hindered by indistinguishable edit vectors across queries, due to the high overlap of shared docIDs in retrieval results. To address this, we propose DOME (docID-oriented model editing), a novel method that effectively and efficiently adapts GR models to unseen documents. DOME comprises three stages: (1) identification of critical layers, (2) optimization of edit vectors, and (3) construction and application of updates. At its core, DOME employs a hybrid-label adaptive training strategy that learns discriminative edit vectors by combining soft labels, which preserve query-specific semantics for distinguishable updates, with hard labels that enforce precise mapping modifications. Experiments on widely used benchmarks, including NQ and MS MARCO, show that our method significantly improves retrieval performance on new documents while maintaining effectiveness on the original collection. Moreover, DOME achieves this with only about 60% of the training time required by incremental training, considerably reducing computational cost and enabling efficient, frequent model updates. Our code is available at https://github.com/zhangzhen-research/DOME
The rapid growth of digital communication has increased the demand for sticker retrieval systems that can match expressive stickers to users' communicative needs. In practice, however, sticker retrieval encounters significant out-of-distribution (OOD) challenges arising from unseen queries and stickers, driven by the diversity of user expression habits and sticker visual representations. These OOD issues often lead to irrelevant or inappropriate retrieval results, undermining the user experience. Drawing on symbolic interactionism in cognition, we propose XAlign-SR, a method that enhances OOD robustness by aligning abstract expressive intent between queries and stickers across modalities. To support this study, we construct OOD benchmarks from sticker datasets that simulate realistic query–sticker scenarios. Experiments demonstrate that our approach significantly outperforms state-of-the-art baselines.
Cross-modal retrieval (CMR) has advanced considerably by mapping image and text features into a shared embedding space; however, these approaches still face two persistent challenges: (1) semantic sparsity, where discriminative cues are confined to localized regions, making it difficult to identify implicit visual evidence; and (2) ranking uncertainty under semantic ambiguity, where models struggle to maintain the correct retrieval order when candidates share similar contexts. To address these issues, we propose the Dual-Branch Multi-Granularity Network (DBMG) with Structured Contrastive Ranking, which enriches visual semantics by leveraging a multimodal large language model to generate auxiliary descriptions, aligns sparse cues through a dual-branch architecture capturing both global and local interactions, and enforces ranking consistency via a three-stage contrastive objective that progressively optimizes category clustering, instance alignment, and margin-based ranking. Extensive experiments on four standard CMR benchmarks demonstrate that DBMG outperforms 12 strong baselines, achieving an average 15.91% improvement in mAP, establishing a new state-of-the-art. The code is available at https://github.com/DMiC-Lab-HFUT/DBMG.
Accurate forecasting of sequential data streams is a cornerstone of modern Web services, supporting applications such as traffic management, user behavior modeling, and online anomaly prevention. However, in many Web environments, new domains emerge rapidly and labeled history data is scarce, which makes zero-shot forecasting particularly challenging. Existing time-series pre-trained models (TSPMs) show promise but they lack the ability to dynamically incorporate external knowledge, while conventional retrieval-augmented generation (RAG) methods are rarely extended beyond text. In this work, we present QuiZSF, a retrieval-augmented forecasting framework that integrates search and forecasting for time series data. The framework performs search by retrieving structurally similar sequences from a large-scale time-series database, and it performs forecasting by integrating the retrieved knowledge into the target sequence. Specifically, QuiZSF introduces a ChronoRAG Base, a hierarchical tree-structured database that enables scalable and domain-aware retrieval, a Multi-grained Series Interaction Learner that captures fine- and coarse-grained dependencies between target and retrieved sequences, and a Model Cooperation Coherer that adapts retrieved knowledge to TSPMs. This design teaches models to actively perform search, align auxiliary information across modalities, and leverage it for more accurate forecasting. Extensive experiments on five public benchmarks demonstrate that QuiZSF consistently outperforms strong baselines, ranking first in up to 87.5% of zero-shot forecasting settings while maintaining high efficiency.