Long-context processing remains a significant challenge for large language models (LLMs). Retrieval-augmented generation (RAG) has recently emerged as a promising approach, enabling LLMs to selectively access relevant information from extended contexts to improve efficiency. However, existing RAG approaches often lag behind other efficient long-context processing methods primarily due to inherent limitations on inaccurate retrieval and fragmented contexts. To address these limitations, we propose RetroLM, a novel RAG framework designed for effective long-context processing. Unlike traditional approaches, RetroLM introduces KV-level retrieval augmentation, which partitions the LLM's KV cache into contiguous pages and performs encoding and decoding operations based on the retrieved KV pages. Built upon this framework, we further develop a specialized retriever for precise retrieval of critical pages and conduct unsupervised post-training to optimize the model’s ability to leverage retrieved information. Compared with traditional RAG, the new approach enhances robustness to retrieval inaccuracy, facilitates effective utilization of fragmented contexts, and saves the cost from repeated context-encoding operations. We conduct extensive evaluations across several popular benchmarks, including LongBench, InfiniteBench, and RULER. RetroLM consistently outperforms existing long-LLMs and RAG-based methods, especially in tasks requiring deep reasoning or extreme context lengths.

Professional financial reports are the cornerstone of investment decision-making, demanding deep analytical reasoning and multimodal synthesis.While recent deep research systems excel in open-domain search tasks, they struggle with financial reporting, specifically in processing structured financial data, ensuring analytical depth, and integrating professional visualizations.To address this gap, we introduce FinSight (Financial InSight), the first multi-agent framework for end-to-end automation of professional multimodal financial reports.At its core, we propose a Code Agent with Variable Memory architecture, unifying financial data, domain tools, and agent modules into a programmable variable space to enable flexible data manipulation and reasoning via executable code.To guarantee report quality, FinSight adopts a Two-Stage Writing Framework with Generative Retrieval.It first distills raw data into structured Chain-of-Analysis segments, then synthesizes them into a coherent, citation-aware, and multimodal narrative complying with financial reporting norms.Additionally, an Iterative Vision-Enhanced Mechanism uses visual feedback to refine codegenerated charts to expert standards.Experiments on company and industry-level tasks demonstrate that FinSight significantly outperforms state-of-the-art deep research systems in factual accuracy, analytical depth, and presentation quality, validating the effectiveness of our framework.

Kun Luo, Hongjin Qian, Zheng Liu, Ziyi Xia, Shitao Xiao, Zhao Cao, Siqi Bao, Jun Zhao, Kang Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Processing long contexts poses a significant challenge for large language models (LLMs) due to their inherent context-window limitations and the computational burden of extensive key-value (KV) activations, which severely impact efficiency.For information-seeking tasks, full context perception is often unnecessary, as a query's information needs can dynamically range from localized details to a global perspective, depending on its complexity.However, existing methods struggle to adapt effectively to these dynamic information needs.In the paper, we propose a method for processing long-context information-seeking tasks via query-guided ACtivation REfilling (ACRE).ACRE constructs a Bi-layer KV Cache for long contexts, where the layer-1 (L1) cache compactly captures global information, and the layer-2 (L2) cache provides detailed and localized information.ACRE establishes a proxying relationship between the two caches, allowing the input query to attend to the L1 cache and dynamically refill it with relevant entries from the L2 cache.This mechanism integrates global understanding with query-specific local details, thus improving answer decoding.Experiments on a variety of long-context informationseeking datasets demonstrate ACRE's effectiveness, achieving improvements in both performance and efficiency.We will release our source codes in this repository.

Conversational Retrieval-Augmented Generation (RAG) systems enhance user interactions by integrating large language models (LLMs) with external knowledge retrieval. However, multi-turn conversations present significant challenges, including implicit user intent and noisy context, which hinder accurate retrieval and response generation. Existing approaches often struggle with the unstructured conversational context and fail to model explicit relations among conversational turns. Moreover, they do not leverage historically relevant passages effectively. To overcome these limitations, we propose EvoRAG, a novel framework that maintains an evolving knowledge graph aligned with the unstructured conversational context. This graph explicitly captures relations among user queries, system responses, and relevant passages across conversational turns, serving as a structured representation of the context. EvoRAG includes three key components: (1) a dual-path retrieval module for context denoising, (2) a unified knowledge integration module for query rewriting and summarization, and (3) a graph-enhanced RAG module for accurate retrieval and response generation. Experiments on four public conversational RAG datasets show that EvoRAG significantly outperforms strong baselines, particularly in handling topic shifts and long dialogue contexts.

Existing studies have optimized retrievalaugmented generation (RAG) across various sub-tasks, such as query understanding and retrieval refinement, but integrating these optimizations into a unified framework remains challenging.To tackle this problem, this work proposes RoleRAG, a unified RAG framework that achieves efficient multi-task processing through role-specific token optimization.RoleRAG comprises six modules, each handling a specific sub-task within the RAG process.Additionally, we introduce a query graph to represent the decomposition of the query, which can be dynamically resolved according to the decomposing state.All modules are driven by the same underlying LLM, distinguished by task-specific role tokens that are individually optimized.This design allows RoleRAG to dynamically activate different modules within a single LLM instance, thereby streamlining deployment and reducing resource consumption.Experimental results on five open-domain question-answering datasets demonstrate the effectiveness, generalizability, and flexibility of our framework.

Large Language Models (LLMs), built on Transformer architectures, exhibit remarkable generalization across a wide range of tasks. However, fine-tuning these models for specific tasks remains resource-intensive due to their extensive parameterization. In this paper, we explore two remarkable phenomena related to the attention mechanism during the fine-tuning of LLMs (where Wq, Wk, and Wv denote the weights of the query, key, and value layers, respectively). The first phenomenon, termed “Unequal Importance of Attention Matrices”, highlights the impact of fine-tuning different weight matrices. It shows that optimizing the Wv matrix yields significantly better performance than optimizing the Wk matrix. Fine-tuning only the Wq and Wv matrices is computationally efficient while delivering results comparable to, or even better than fine-tuning all three matrices (Wq, Wk, and Wv). The second phenomenon, “Attention Matrices with Customized Learning Rate Lead to Better Convergence”, emphasizes the importance of assigning distinct learning rates to these matrices. Specifically, a higher learning rate for the Wv matrix compared to Wq and Wk accelerates convergence and improves performance. Building on these insights, we propose a new strategy that improves fine-tuning efficiency in terms of both storage and time. Experimental results on benchmark datasets validate the effectiveness of this approach, supporting our theoretical findings. Our analysis lays the theoretical groundwork for configuring and improving algorithms in LLMs fine-tuning.

Knowledge-intensive tasks often require complex reasoning and contextual understanding over long contexts. However, the learning and deployment of long-LLMs remains a challenging problem despite recent progresses. In this work, we propose that the short LLMs have great potentiality for solving knowledge-intensive tasks that have long context, i.e. they can be solved by purely working with oracle short-contexts within the input long-context. On top of this argument, we propose a framework called DCISO DynamiC knowledge-Intensive task S>Olver), which enables a short-LLM to address the knowledge-intensive tasks with long context via dynamic context browsing. In our framework, the short-LLM prompts itself to reason for two critical decisions: 1) how to access to the appropriate part of context within the input, 2) how to make effective use of the accessed context. By adaptively accessing and utilizing the context based on the presented tasks, DCISO can serve as a general framework to handle diversified knowledge-intensive long-context problems. We comprehensively evaluate different types of tasks from popular long-context benchmarks, where DCISO is able to achieve a substantially improved performance. Our codes will be released at this repository.

Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.

Augmenting large language models (LLMs) with external retrieval has become a standard method to address their inherent knowledge cutoff limitations. However, traditional retrieval-augmented generation methods employ static, pre-inference retrieval strategies, making them inadequate for complex tasks involving ambiguous, multi-step, or evolving information needs. Recent advances in test-time scaling techniques have demonstrated significant potential in enabling LLMs to dynamically interact with external tools, motivating the shift toward adaptive inference-time retrieval. Inspired by Information Foraging Theory (IFT), we propose InForage, a reinforcement learning framework that formalizes retrieval-augmented reasoning as a dynamic information-seeking process. Unlike existing approaches, InForage explicitly rewards intermediate retrieval quality, encouraging LLMs to iteratively gather and integrate information through adaptive search behaviors. To facilitate training, we construct a human-guided dataset capturing iterative search and reasoning trajectories for complex, real-world web tasks. Extensive evaluations across general question answering, multi-hop reasoning tasks, and a newly developed real-time web QA dataset demonstrate InForage's superior performance over baseline methods. These results highlight InForage's effectiveness in building robust, adaptive, and efficient reasoning agents. We provide all codes and datasets in the supplementary materials.

In real-world information-seeking scenarios, users have dynamic and diverse needs, requiring RAG systems to demonstrate adaptable resilience. To comprehensively evaluate the resilience of current RAG methods, we introduce HawkBench, a human-labeled, multi-domain benchmark designed to rigorously assess RAG performance across categorized task types. By stratifying tasks based on information-seeking behaviors, HawkBench provides a systematic evaluation of how well RAG systems adapt to diverse user needs.Unlike existing benchmarks, which focus primarily on specific task types (mostly factoid queries) and rely on varying knowledge bases, HawkBench offers: (1) systematic task stratification to cover a broad range of query types, including both factoid and rationale queries, (2) integration of multi-domain corpora across all task types to mitigate corpus bias, and (3) rigorous annotation for high-quality evaluation.HawkBench includes 1,600 high-quality test samples, evenly distributed across domains and task types. Using this benchmark, we evaluate representative RAG methods, analyzing their performance in terms of answer quality and response latency. Our findings highlight the need for dynamic task strategies that integrate decision-making, query interpretation, and global knowledge understanding to improve RAG generalizability. We believe HawkBench serves as a pivotal benchmark for advancing the resilience of RAG methods and their ability to achieve general-purpose information seeking.

Processing long contexts presents a significant challenge for large language models (LLMs). While recent advancements allow LLMs to handle much longer contexts than before (e.g., 32K or 128K tokens), it is computationally expensive and can still be insufficient for many applications. Retrieval-Augmented Generation (RAG) is considered a promising strategy to address this problem. However, conventional RAG methods face inherent limitations because of two underlying requirements: 1) explicitly stated queries, and 2) well-structured knowledge. These conditions, however, do not hold in general long-context processing tasks.

As a cornerstone of modern information access, search engines have become indispensable in everyday life. With the rapid advancements in AI and natural language processing (NLP) technologies, particularly large language models (LLMs), search engines have evolved to support more intuitive and intelligent interactions between users and systems. Conversational search, an emerging paradigm for next-generation search engines, leverages natural language dialogue to facilitate complex and precise information retrieval, thus attracting significant attention. Unlike traditional keyword-based search engines, conversational search systems enhance user experience by supporting intricate queries, maintaining context over multi-turn interactions, and providing robust information integration and processing capabilities. Key components such as query reformulation, search clarification, conversational retrieval, and response generation work in unison to enable these sophisticated interactions. In this survey, we explore the recent advancements and potential future directions in conversational search, examining the critical modules that constitute a conversational search system. We highlight the integration of LLMs in enhancing these systems and discuss the challenges and opportunities that lie ahead in this dynamic field. Additionally, we provide insights into real-world applications and robust evaluations of current conversational search systems, aiming to guide future research and development in conversational search.

This paper presents a novel Chunking-Free In-Context (CFIC) retrieval approach, specifically tailored for Retrieval-Augmented Generation (RAG) systems.Traditional RAG systems often struggle with grounding responses using precise evidence text due to the challenges of processing lengthy documents and filtering out irrelevant content.Commonly employed solutions, such as document chunking and adapting language models to handle longer contexts, have their limitations.These methods either disrupt the semantic coherence of the text or fail to effectively address the issues of noise and inaccuracy in evidence retrieval.CFIC addresses these challenges by circumventing the conventional chunking process.It utilizes the encoded hidden states of documents for in-context retrieval, employing autoaggressive decoding to accurately identify the specific evidence text required for user queries, eliminating the need for chunking.CFIC is further enhanced by incorporating two decoding strategies, namely Constrained Sentence Prefix Decoding and Skip Decoding.These strategies not only improve the efficiency of the retrieval process but also ensure that the fidelity of the generated grounding text evidence is maintained.Our evaluations of CFIC on a range of open QA datasets demonstrate its superiority in retrieving relevant and accurate evidence, offering a significant improvement over traditional methods.By doing away with the need for document chunking, CFIC presents a more streamlined, effective, and efficient retrieval solution, making it a valuable advancement in the field of RAG systems.The codes will be released in this repository.

Conversational search supports multi-turn user-system interactions to solve complex information needs. Compared with the traditional single-turn ad-hoc search, conversational search faces a more complex search intent understanding problem because a conversational search session is much longer and contains many noisy tokens. However, existing conversational dense retrieval solutions simply fine-tune the pre-trained ad-hoc query encoder on limited conversational search data, which are hard to achieve satisfactory performance in such a complex conversational search scenario. Meanwhile, the learned latent representation also lacks interpretability that people cannot perceive how the model understands the session. To tackle the above drawbacks, we propose a sparse Lexical-based Conversational REtriever (LeCoRE), which extends the SPLADE model with two well-matched multi-level denoising methods uniformly based on knowledge distillation and external query rewrites to generate denoised and interpretable lexical session representation. Extensive experiments on four public conversational search datasets in both normal and zero-shot evaluation settings demonstrate the strong performance of LeCoRE towards more effective and interpretable conversational search.

Conversational search provides users with natural and convenient new search experience.Recently, conversational dense retrieval has shown to be a promising technique to realize conversational search.However, as conversational search systems have not been widely deployed, it is hard to get large-scale real conversational search sessions and relevance labels to support the training of conversational dense retrieval.To tackle this data scarcity problem, previous methods focus on developing better few-shot learning approaches or generating pseudo relevance labels, but the data they use for training still heavily rely on manual generation.In this paper, we present ConvTrans, a data augmentation method that can automatically transform easily-accessible web search sessions into conversational search sessions to fundamentally alleviate the data scarcity problem for conversational dense retrieval.ConvTrans eliminates the gaps between these two types of sessions in terms of session quality and query form to achieve effective session transformation.Extensive evaluations on two widely used conversational search benchmarks, i.e., CAsT-19 and CAsT-20, demonstrate that the same model trained on the data generated by Con-vTrans can achieve comparable retrieval performance as it trained on high-quality but expensive artificial conversational search data.

In a conversational search scenario, a query might be context-dependent because some words are referred to previous expressions or omitted.Previous works tackle the issue by either reformulating the query into a selfcontained query (query rewriting) or learning a contextualized query embedding from the query context (context modelling).In this paper, we propose a model CRDR that can perform query rewriting and context modelling in a unified framework in which the query rewriting's supervision signals further enhance the context modelling.Instead of generating a new query, CRDR only performs necessary modifications on the original query, which improves both accuracy and efficiency of query rewriting.In the meantime, the query rewriting benefits the context modelling by explicitly highlighting relevant terms in the query context, which improves the quality of the learned contextualized query embedding.To verify the effectiveness of CRDR, we perform comprehensive experiments on TREC CAsT-19 and TREC CAsT-20 datasets, and the results show that our method outperforms all baseline models in terms of both quality of query rewriting and quality of context-aware ranking.

Hanxun Zhong, Zhicheng Dou, Yutao Zhu, Hongjin Qian, Ji-Rong Wen. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Pre-trained language models (PLMs) have achieved great success in the area of Information Retrieval. Studies show that applying these models to ad-hoc document ranking can achieve better retrieval effectiveness. However, on the Web, most information is organized in the form of HTML web pages. In addition to the pure text content, the structure of the content organized by HTML tags is also an important part of the information delivered on a web page. Currently, such structured information is totally ignored by pre-trained models which are trained solely based on text content. In this paper, we propose to leverage large-scale web pages and their DOM (Document Object Model) tree structures to pre-train models for information retrieval. We argue that using the hierarchical structure contained in web pages, we can get richer contextual information for training better language models. To exploit this kind of information, we devise four pre-training objectives based on the structure of web pages, then pre-train a Transformer model towards these tasks jointly with traditional masked language model objective. Experimental results on two authoritative ad-hoc retrieval datasets prove that our model can significantly improve ranking performance compared to existing pre-trained models.

Conversational search is a crucial and promising branch in information retrieval. In this paper, we reveal that not all historical conversational turns are necessary for understanding the intent of the current query. The redundant noisy turns in the context largely hinder the improvement of search performance. However, enhancing the context denoising ability for conversational search is quite challenging due to data scarcity and the steep difficulty for simultaneously learning conversational query encoding and context denoising. To address these issues, in this paper, we present a novel Curriculum cOntrastive conTExt Denoising framework, COTED, towards few-shot conversational dense retrieval. Under a curriculum training order, we progressively endow the model with the capability of context denoising via contrastive learning between noised samples and denoised samples generated by a new conversation data augmentation strategy. Three curriculums tailored to conversational search are exploited in our framework. Extensive experiments on two few-shot conversational search datasets, i.e., CAsT-19 and CAsT-20, validate the effectiveness and superiority of our method compared with the state-of-the-art baselines.

In this paper, we explore the problem of developing personalized chatbots. A personalized chatbot is designed as a digital chatting assistant for a user. The key characteristic of a personalized chatbot is that it should have a consistent personality with the corresponding user. It can talk the same way as the user when it is delegated to respond to others' messages. Many methods have been proposed to assign a personality to dialogue chatbots, but most of them utilize explicit user profiles, including several persona descriptions or key-value-based personal information. In a practical scenario, however, users might be reluctant to write detailed persona descriptions, and obtaining a large number of explicit user profiles requires tremendous manual labour. To tackle the problem, we present a retrieval-based personalized chatbot model, namely IMPChat, to learn an implicit user profile from the user's dialogue history. We argue that the implicit user profile is superior to the explicit user profile regarding accessibility and flexibility. IMPChat aims to learn an implicit user profile through modeling user's personalized language style and personalized preferences separately. To learn a user's personalized language style, we elaborately build language models from shallow to deep using the user's historical responses; To model a user's personalized preferences, we explore the conditional relations underneath each post-response pair of the user. The personalized preferences are dynamic and context-aware: we assign higher weights to those historical pairs that are topically related to the current query when aggregating the personalized preferences. We match each response candidate with the personalized language style and personalized preference, respectively, and fuse the two matching signals to determine the final ranking score. We conduct comprehensive experiments on two large datasets, and the results show that our method outperforms all baseline models.

atural language dialogue systems raise great attention recently. As many dialogue models are data-driven, high-quality datasets are essential to these systems. In this paper, we introduce Pchatbot, a large-scale dialogue dataset that contains two subsets collected from Weibo and Judicial forums respectively. To adapt the raw dataset to dialogue systems, we elaborately normalize the raw dataset via processes such as anonymization, deduplication, segmentation, and filtering. The scale of Pchatbot is significantly larger than existing Chinese datasets, which might benefit the data-driven models. Besides, current dialogue datasets for personalized chatbot usually contain several persona sentences or attributes. Different from existing datasets, Pchatbot provides anonymized user IDs and timestamps for both posts and responses. This enables the development of personalized dialogue models that directly learn implicit user personality from the user's dialogue history. Our preliminary experimental study benchmarks several state-of-the-art dialogue models to provide a comparison for future work. The dataset can be publicly accessed at Github: https://github.com/qhjqhj00/Pchatbot.