Retrieval-Augmented Generation (RAG) systems are widely deployed and increasingly influential, but their reliance on external corpora exposes new security risks from poisoned retrieval content. Existing RAG attacks are largely focusing on individual queries or narrow topic-local query sets, which limits their practical reach and offers limited camouflage in real-world settings. In this paper, we introduce discourse-level opinion manipulation, a new threat model in which coordinated influence across a semantic query network induces opinion shifts over a holistic, multi-topic query space. We formalize this threat in a black-box setting and propose DiscourseFlip, an agentic, graph-guided attack that dynamically allocates a limited poisoning budget to maximize discourse-level opinion deviation. Extensive experiments demonstrate that DiscourseFlip consistently induces targeted opinion shifts across the contextualized query network and significantly outperforms existing baselines in terms of coverage and effectiveness. User studies further confirm that DiscourseFlip is effective while remaining well camouflaged from user detection. Moreover, systematic analyses show that existing mitigation strategies are ineffective against discourse-level manipulation, underscoring the urgent need for more robust and adaptive defenses to address discourse-level vulnerabilities.

Large language models (LLMs) are increasingly employed in scientometric studies for scientific text analysis and knowledge discovery. A key factor influencing their in-context learning (ICL) performance is the organization of demonstrations, which guides how LLMs interpret complex scientific texts. However, existing organization strategies often involve high computational costs and lack interpretability, limiting their effectiveness and transferability across models. We propose In-Context Curriculum Learning (ICCL), a lightweight and explainable framework that organizes demonstrations in ascending order of complexity to enhance LLMs’ deep semantic understanding of scientific text. ICCL comprises three stages: demonstration retrieval, label recall, and curriculum-guided ordering based on a novel Instruction Alignment Score (IAS) to quantify contextual difficulty. We evaluate ICCL on three representative scientific text mining benchmarks, SciCite, SciNLI, and SciERC, covering citation intent classification, scientific language inference, and entity extraction tasks. Relative to few-shot prompting, ICCL yields consistent average F1-score gains of 4.65 https://github.com/61peng/sci-iccl .

Scientific knowledge flows enable cumulative progress by connecting researchers across disciplines, institutions, and countries. Yet it remains unclear how geography and national structures continue to shape these exchanges in an increasingly connected world. Using a large-scale bibliometric dataset from OpenAlex, which covers 39.35 million publications across 95 countries and 3,794 cities between 2000 and 2022, we examine global knowledge diffusion through two complementary channels: co-authorship and citation. We find that the constraining effect of geographic distance on collaboration has not diminished over time but has instead intensified, suggesting persistent structural or institutional barriers. Citation flows, by contrast, are less sensitive to spatial proximity, indicating that intellectual influence may diffuse more freely across borders. At the country level, research networks exhibit strong domestic preferences and a shared citation orientation toward the United States. China, while increasingly favored as a collaboration partner by other countries, continues to be systematically undercited within global citation flows. International mobility increases researchers' collaboration with scholars in their host country but has limited effects on citation flows. These results highlight the structural persistence of spatial and country biases in global science, with implications for equitable participation and recognition across regions.

The advancement of Document Intelligence (DI) demands large-scale, high-quality training data, yet manual annotation remains a critical bottleneck. While data generation methods are evolving rapidly, existing surveys are constrained by fragmented focuses on single modalities or specific tasks, lacking a unified perspective aligned with real-world workflows. To fill this gap, this survey establishes the first comprehensive technical map for data generation in DI. Data generation is redefined as supervisory signal production, and a novel taxonomy is introduced based on the "availability of data and labels." This framework organizes methodologies into four resource-centric paradigms: Data Augmentation, Data Generation from Scratch, Automated Data Annotation, and Self-Supervised Signal Construction. Furthermore, a multi-level evaluation framework is established to integrate intrinsic quality and extrinsic utility, compiling performance gains across diverse DI benchmarks. Guided by this unified structure, the methodological landscape is dissected to reveal critical challenges such as fidelity gaps and frontiers including co-evolutionary ecosystems. Ultimately, by systematizing this fragmented field, data generation is positioned as the central engine for next-generation DI.

Topic evolution analysis is essential for uncovering, comprehending, and leveraging the underlying mechanisms affecting the development of topics. However, previous studies mainly examine topic evolution in isolation, neglecting topic interactions and the functional roles of topics, thus failing to unveil and explain the interactive relationship driving their evolution. In this study, we define and identify the function of topics via a large language model (LLM) and disclose a competitive-cooperative mechanism (CCM) by scrutinizing the spatial autocorrelation of topics. Specifically, we collected about 23 million papers in the field of computer science and utilized zero-shot learning to identify the four functions (i.e., Problem, Method, Problem Method, and Other) of topics. Subsequently, we used Word2Vec to generate topic embeddings, then investigated CCM among topics with the same and/or different functions by examining the global and local spatial autocorrelation of topic frequency within their semantic space. Our findings reveal that Problem topics cooperate when the solution of one problem facilitates the solving of another one, yet compete for researchers’ attention. Method topics cooperate through complementarity but compete as researchers’ preference for more efficient methods. Among Problem and Method topics, cooperation emerges from the use of diverse methods to tackle problems. With the innovative insight into topic evolution, this study not only develops a new method for topic term identification but also reveals global and local CCM across topics of different functions.

The expeditious identification of papers with potentially disruptive and developmental contributions is a meaningful issue in science of science research. To achieve this aim, we proposed an effective method to automatically build an abstract format-balanced dataset based on the basic idea of disruption index. Subsequently, we respectively utilized a range of machine learning models (MLs), deep learning models (DLs), and large language models (LLMs) to build a content-based identification model based on the dataset. The optimal model, Mistral-7B fine-tuned via QLoRA, significantly outperforms MLs and DLs, and achieved an F1 score of 0.7735 on this extremely challenging task. Hence, the model can promptly and effectively distinguish between disruptive, developmental, and general research based purely on research content. Further, we developed three evaluation datasets to scrutinize the effectiveness of the model, demonstrating its remarkable ability to distinguish Nobel Prize papers from randomly paired counterparts, as well as to differentiate between randomly sampled papers in journals of varying impact factors (IF). Additionally, the papers’ scores derived from our model’s prediction exhibit a strong positive correlation with their future scientific impact. We also reveal that evaluating papers based solely on IF is insufficient, and papers boasting high view counts are more likely to represent potentially disruptive or developmental research. Thus, our work not only provides an effective method for timely identifying high-quality research, but also provides guidance for promoting fair scientific evaluation.

Large language models (LLMs) are currently applied to scientific paper evaluation by assigning an absolute score to each paper independently. However, since score scales vary across conferences, time periods, and evaluation criteria, models trained on absolute scores are prone to fitting narrow, context-specific rules rather than developing robust scholarly judgment. To overcome this limitation, we propose shifting paper evaluation from isolated scoring to collaborative ranking. In particular, we design \textbf{C}omparison-\textbf{N}ative framework for \textbf{P}aper \textbf{E}valuation (\textbf{CNPE}), integrating comparison into both data construction and model learning. We first propose a graph-based similarity ranking algorithm to facilitate the sampling of more informative and discriminative paper pairs from a collection. We then enhance relative quality judgment through supervised fine-tuning and reinforcement learning with comparison-based rewards. At inference, the model performs pairwise comparisons over sampled paper pairs and aggregates these preference signals into a global relative quality ranking. Experimental results demonstrate that our framework achieves an average relative improvement of \textbf{21.8\%} over the strong baseline DeepReview-14B, while exhibiting robust generalization to five previously unseen datasets. \href{https://github.com/ECNU-Text-Computing/ComparisonReview}{Code}.

Understanding scientific figures is challenging due to their complexity and the need for domain expertise. Existing studies, limited by data and model capabilities, primarily focus on captioning and classifying simpler visualizations, such as bar and pie charts, while neglecting the fine-grained interpretation of knowledge-rich diagrams like frameworks and flowcharts. To address this gap, we introduce a novel task, Figure Integrity Verification, to assess models’ ability to align textual knowledge with detailed visual elements and identify visual elements that lack textual explanations. To support this task, we construct a large-scale text-figure alignment dataset, Figure-seg, containing 15,761 instances. Building upon this dataset, we propose an innovative framework, Every Part Matters (EPM), which employs Multimodal Large Language Models (MLLMs) and analogical reasoning to verify and enhance figure integrity. Experimental results demonstrate that the proposed framework improves the IoU metric for text-figure alignment by 33.83 https://github.com/shixiang1a/figure_understanding ).

Multimodal fusion is susceptible to modality imbalance, where dominant modalities overshadow weak ones, easily leading to biased learning and suboptimal fusion, especially for incomplete modality conditions. To address this problem, we introduce an Equilibrium Deviation Metric (EDM) to quantify this imbalance and verify, in both theoretical and empirical terms, that the optimization order of modalities plays a critical role in approaching equilibrium. In particular, we demonstrate that an EDM-ranked weak-to-strong schedule achieves the tightest convergence bound among all possible ordering strategies. Leveraging these insights, we design an alternating strategy that dynamically prioritises under-optimised modalities, plus a modality-mapping layer for feature alignment and a memory module for information filtering and inheritance. Our framework is compatible with both conventional and MLLM-based backbones. It achieves new state-of-the-art (SOTA) on four benchmarks (e.g., +3.36% on CREMA-D, +3.51% on Kinetics-400), and remains robust under missing-modality conditions. These findings highlight the value of modality scheduling, offering a principled alternative to conventional joint training.

As online resources expand, users increasingly engage in complex search tasks to acquire knowledge or explore unfamiliar topics. Supporting dynamic and interactive search behaviors remains challenging for current search engines, but conversational search offers a promising solution. In conversational search, users express intent naturally through spoken language and engage in multi-turn interactions to progressively satisfy their information needs. Previous studies on traditional web search have shown that task complexity and domain expertise significantly influence user behavior and performance. However, it remains unclear whether these factors exert different effects on user behavior and performance in conversational search, particularly in exploratory tasks. In this study, we conducted a carefully designed user experiment with 32 participants and collected a comprehensive dataset of search behaviors, including querying, clicking, and eye fixation sequences, along with explicit feedback and questionnaire data. Using this dataset, we investigated the influence of task complexity and domain expertise on search performance and user behavior in conversational exploratory search, focusing on three interactive aspects: user effort, user queries, and user clicks. Our findings provide valuable insights into how task complexity and expertise influence user behavior and search performance, with implications for optimizing conversational systems to better support complex exploratory tasks.

Understanding how processing, structure, properties, and performance interact is essential for guiding materials design and discovery. Yet, causal mechanisms linking these elements are typically scattered across text, figures, and references in the literature, and efforts to systematically mine and organize such knowledge remain limited. In this work, we leverage an LLM-based mechanism deduction framework to construct a dataset of 207,200 fine-grained mechanisms with 1,113,940 multimodal evidences from 61,766 materials science research articles. Each mechanism is linked to a specific causal relation among the tetrahedral elements and is supported by evidence from experiment information, characterization results, and external knowledge, with its accuracy verified by materials science researchers. This dataset provides a large-scale, cross-validated collection of multimodal mechanism knowledge in materials science, serving as a resource for data-driven research and intelligent analysis.

Recent studies have demonstrated that incorporating Chain-of-Thought (CoT) reasoning into the detection process can enhance a model's ability to detect synthetic images. However, excessively lengthy reasoning incurs substantial resource overhead, including token consumption and latency, which is particularly redundant when handling obviously generated forgeries. To address this issue, we propose Fake-HR1, a large-scale hybrid-reasoning model that, to the best of our knowledge, is the first to adaptively determine whether reasoning is necessary based on the characteristics of the generative detection task. To achieve this, we design a two-stage training framework: we first perform Hybrid Fine-Tuning (HFT) for cold-start initialization, followed by online reinforcement learning with Hybrid-Reasoning Grouped Policy Optimization (HGRPO) to implicitly learn when to select an appropriate reasoning mode. Experimental results show that Fake-HR1 adaptively performs reasoning across different types of queries, surpassing existing LLMs in both reasoning ability and generative detection performance, while significantly improving response efficiency.

PurposeThis study aims to enhance the performance of fine-tuned small named entity recognition (NER) models by leveraging the language generation capability of large language models (LLMs) to rewrite entity contexts, addressing the limited performance gains of state-of-the-art (SoTA) NER methods in recent years.Design/methodology/approachThe authors propose a novel paradigm that uses LLMs to modify the surrounding context of entities in the input text, transforming hard cases into easier ones for the NER model to recognize. Experiments are conducted on multiple widely used NER data sets to evaluate the effectiveness of the proposed method. No additional retraining or architectural modifications are applied to the vanilla NER models.FindingsExperimental results demonstrate that the method consistently improves the performance of existing NER models across all tested data sets, achieving SoTA results. The approach effectively mitigates challenges associated with out-of-vocabulary entities, syntactically complex sentences and linguistic distractors while maintaining relatively low computational cost.Originality/valueThis work introduces a new paradigm for NER that integrates LLM text rewriting with pretrained small NER models. Unlike previous approaches relying on data augmentation or NER model retraining, the authors' method achieves performance gains purely through LLM-based context rewriting, offering a cost-efficient and scalable solution for real-world applications.

Neural ranking models have achieved remarkable progress and are now widely deployed in real-world applications such as Retrieval-Augmented Generation (RAG). However, like other neural architectures, they remain vulnerable to adversarial manipulations: subtle character-, word-, or phrase-level perturbations can poison retrieval results and artificially promote targeted candidates, undermining the integrity of search engines and downstream systems. Existing defenses either rely on heuristics with poor generalization or on certified methods that assume overly strong adversarial knowledge, limiting their practical use. To address these challenges, we propose RobustMask, a novel defense that combines the context-prediction capability of pretrained language models with a randomized masking-based smoothing mechanism. Our approach strengthens neural ranking models against adversarial perturbations at the character, word, and phrase levels. Leveraging both the pairwise comparison ability of ranking models and probabilistic statistical analysis, we provide a theoretical proof of RobustMask's certified top-K robustness. Extensive experiments further demonstrate that RobustMask successfully certifies over 20

As Large Language Models (LLMs) like ChatGPT and GPT-4 continue to evolve, their application in facilitating the academic writing process is increasing. Writing the "Related Work" section in a scientific manuscript requires scholars to integrate knowledge from multiple sources, making it labor-intensive and time-consuming. Current predominant methods typically format the "Related Work" section Generation (RWG) in isolated environments with limited knowledge, which is quite different from how humans write. Therefore, our research pivots RWG toward a more author-centered approach that simulates the human academic writing procedure. We introduce a novel methodology that synergizes the strengths of both small language models (SLMs) and large language models. Our method involves generating author-oriented cues via SLMs, drafting initial versions with LLMs, and designing a delta feedback mechanism for cue refinement, which reveals the information gap between drafts. In the generation process, our approach assembles multiple sources of academic knowledge. This includes the LLMs' tacit knowledge, the explicit knowledge of domain-specific academic publications, and the context derived from scholarly graphs. Comparative evaluations, both LLM-based and human-based, demonstrate that our method significantly surpasses existing baseline models like original GPT-4 and CoT-based GPT-4. Specifically, the LLM-based evaluation indicates that, when guided by enhanced cues, our model can produce high-quality content with an average of only 3.46 revision edits, compared to approximately 4 revision edits required by the baseline model. Additionally, the human evaluation demonstrates that our model obtained a higher score on the content, structure, and argumentation dimensions. In conclusion, our model has enhanced capability in synthesizing input papers comprehensively and aligning more closely with the authors' focal points.

The advent of Large Language Models (LLMs) has shown the potential to improve relevance and provide direct answers in web searches. However, challenges arise in validating the reliability of generated results and the credibility of contributing sources, due to the limitations of traditional information retrieval algorithms and the LLM hallucination problem. Aiming to create a "PageRank" for the LLM era, we strive to transform LLM into a relevant, responsible, and trustworthy searcher. We propose a novel generative retrieval framework leveraging the knowledge of LLMs to foster a direct link between queries and online sources. This framework consists of three core modules: Generator, Validator, and Optimizer, each focusing on generating trustworthy online sources, verifying source reliability, and refining unreliable sources, respectively. Extensive experiments and evaluations highlight our method's superior relevance, responsibility, and trustfulness against various SOTA methods.

Topic analysis aims to study topic evolution and trends in order to help researchers understand the process of knowledge evolution and creation. This paper develops a novel topic evolution analysis framework, which we use to demonstrate, forecast, and explain topic evolution from the perspective of the geometrical motion of topic embeddings generated by pretrained language models. Our data set comprises approximately 15 million papers in the computer science field, with 7,000 "fields of study" to represent the topics. First, we demonstrate that over 80% of topics have undergone obvious motion in the semantic vector space, based on the hyperplane and its normal vector generated by a support vector machine. Subsequently, we verified the predictability of the motion based on three vector regression models by predicting topic embeddings. Finally, we employed a decoder to explain the predicted motion, whose forecast embeddings can capture about 50% of unseen topics. Our research framework shows that topic evolution can be analyzed via the geometrical motion of topic embeddings, and the semantic motion of old topics nurtures new topics. The current study opens new research pathways in topic analysis and sheds light on the topic evolution mechanism from a novel geometric perspective.

PurposeEvaluating the innovation of scientific publications remains a critical challenge, as traditional citation-based metrics often fail to capture both the novelty of research content and the dynamics of knowledge diffusion. This study aims to propose a dual-perspective framework that jointly measures how research draws upon prior knowledge and contributes to subsequent scientific development.Design/methodology/approachUsing a data set of 193,193 papers in computer science and library and information science, this paper examines how research builds on existing knowledge and influences future work. A dual-perspective framework is proposed, incorporating four interpretable indicators: knowledge adoption novelty, knowledge diffusion breakthrough, knowledge duality index and knowledge innovation contribution ratio. Based on these indicators, a quadrant-based innovation typology is developed to classify research into four types of scientific innovation.FindingsThe analysis shows that innovation indicators vary across fields, reflecting differences in how knowledge is produced and diffused. Based on these indicators, research can be grouped into four types with distinct innovation profiles. While citation and usage patterns differ across types, some contrasts are modest and field-dependent.Research limitations/implicationsThe paper observed temporal trends in indicator distributions, whereas the study did not apply temporal normalization to the innovation indicators.Practical implicationsThis study offers a structured and interpretable framework for evaluating scholarly innovation.Social implicationsBy addressing both how knowledge is built upon and how it is diffused, the proposed indicators provide a comprehensive perspective on research contribution.Originality/valueThis study proposes a dual-perspective framework that jointly captures the novelty and diffusion of scientific publications through four interpretable indicators.

The immediate and efficient evaluation of scientific papers is crucial for advancing scientific progress. However, traditional peer review faces numerous challenges, including reviewer bias, limited expertise, and an overwhelming volume of publications. Recent advancements in large language models (LLMs) suggest their potential as promising evaluators, capable of approximating human cognition and understanding both ordinary and scientific language. In this study, we propose a novel AI-empowered paper evaluation method, PaperEval (PE), which utilizes a multi-agent system powered by LLMs to design evaluation criteria, assess paper quality along different dimensions, and generate explainable scores. We also introduce two variants of PE, Multi-round PaperEval (MPE) and Self-correcting PaperEval (SPE), which produce comparable scores and iteratively refine the evaluation criteria, respectively. To test our methods, we conducted a comprehensive analysis of three curated datasets, encompassing about 66,000 target papers of varying quality across the fields of mathematics, physics, chemistry, and medicine. The results show that our methods can effectively discern between high- and low-quality papers based on scores derived in four dimensions: Question, Method, Result, and Conclusion. Moreover, the results highlight the evaluation’s stability over time, the impact of comparative papers, the advantages of the multi-round evaluation strategy, and the varying correlation between AI ratings and scientific impact across different disciplines. Our method can seamlessly integrate into the existing scientific evaluation system, offering valuable insights for the development of AI-driven scientific evaluation.

Timely identification of potentially disruptive research is a significant research issue, since disruptive innovation in science transforms the existing paradigm and/or opens a new paradigm. This study proposes a comparative power-based large model that can promptly and accurately identify potentially disruptive research via comparative analysis of semantically-related papers. To this end, a self-constructed dataset was built by treating accumulated disruptive and consolidating citations as crowdsourced annotation data. We employed a range of machine learning models (MLs), deep learning models (DLs), and large language models (LLMs) to build classifiers. Our optimal model, Mistral-7B+*, attains an impressive F1 score of 0.8210 and outperforms the best-performing ML and DL models by approximately 27.05 % and 14.03 %, respectively. Testing on 275 recently published biomedical papers further verifies its effectiveness. Additionally, we conduct comprehensive experiments to scrutinize the comparative power of the large model as well as the impact of the number and quality of comparative papers and distinct functional paragraphs within abstracts on identification performance. Our findings show that an appropriate number and quality of comparative papers can promote identification performance. Moreover, result-based paragraphs are the most important for identifying disruptive research, while method-based paragraphs are least important.