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.

Retrieval-Augmented Generation (RAG) enriches LLMs by dynamically retrieving external knowledge, reducing hallucinations and satisfying real-time information needs. While existing research mainly targets RAG's performance and efficiency, emerging studies highlight critical security concerns. Yet, current adversarial approaches remain limited, mostly addressing white-box scenarios or heuristic black-box attacks without fully investigating vulnerabilities in the retrieval phase. Additionally, prior works mainly focus on factoid Q&A tasks, their attacks lack complexity and can be easily corrected by advanced LLMs. In this paper, we investigate a more realistic and critical threat scenario: adversarial attacks intended for opinion manipulation against black-box RAG models, particularly on controversial topics. Specifically, we propose FlippedRAG, a transfer-based adversarial attack against black-box RAG-like systems. We first demonstrate that the underlying retriever of a black-box RAG can be reverse-engineered and approximated by enumerating critical queries, candidates, and answers, enabling us to train a surrogate retriever. Leveraging the surrogate retriever, we further craft target poisoning triggers, altering vary few documents to effectively manipulate both retrieval and subsequent generation, transferring the attack to the original black-box RAG model. Extensive empirical results show that FlippedRAG substantially outperforms baseline methods, improving the average attack success rate by 16.7%. Across four diverse domains, FlippedRAG achieves on average a 50% directional shift in the opinion polarity of RAG-generated responses, ultimately causing a notable 20% shift in user cognition. Furthermore, we actively evaluate the performance of several potential defensive measures, concluding that existing mitigation strategies remain insufficient against such sophisticated manipulation attacks. These results highlight an urgent need for developing innovative defensive solutions to ensure the security and trustworthiness of RAG systems.

Abstract Each section header of an article has its distinct communicative function. Citations from distinct sections may be different regarding citing motivation. In this paper, we grouped section headers with similar functions as a structural function and defined the distribution of citations from structural functions for a paper as its citation structure. We aim to explore the relationship between citation structure and the future impact of a publication and disclose the relative importance among citations from different structural functions. Specifically, we proposed two citation counting methods and a citation life cycle identification method, by which the regression data were built. Subsequently, we employed a ridge regression model to predict the future impact of the paper and analyzed the relative weights of regressors. Based on documents collected from the Association for Computational Linguistics Anthology website, our empirical experiments disclosed that functional structure features improve the prediction accuracy of citation count prediction and that there exist differences among citations from different structural functions. Specifically, at the early stage of citation lifetime, citations from Introduction and Method are particularly important for perceiving future impact of papers, and citations from Result and Conclusion are also vital. However, early accumulation of citations from the Background seems less important.

BACKGROUND: Biomedical sciences, with their focus on human health and disease, have attracted unprecedented attention in the 21st century. The proliferation of biomedical sciences has also led to a large number of scientific articles being produced, which makes it difficult for biomedical researchers to find relevant articles and hinders the dissemination of valuable discoveries. To bridge this gap, the research community has initiated the article recommendation task, with the aim of recommending articles to biomedical researchers automatically based on their research interests. Over the past two decades, many recommendation methods have been developed. However, an algorithm-level comparison and rigorous evaluation of the most important methods on a shared dataset is still lacking. METHOD: In this study, we first investigate 15 methods for automated article recommendation in the biomedical domain. We then conduct an empirical evaluation of the 15 methods, including six term-based methods, two word embedding methods, three sentence embedding methods, two document embedding methods, and two BERT-based methods. These methods are evaluated in two scenarios: article-oriented recommenders and user-oriented recommenders, with two publicly available datasets: TREC 2005 Genomics and RELISH, respectively. RESULTS: Our experimental results show that the text representation models BERT and BioSenVec outperform many existing recommendation methods (e.g., BM25, PMRA, XPRC) and web-based recommendation systems (e.g., MScanner, MedlineRanker, BioReader) on both datasets regarding most of the evaluation metrics, and fine-tuning can improve the performance of the BERT-based methods. CONCLUSIONS: Our comparison study is useful for researchers and practitioners in selecting the best modeling strategies for building article recommendation systems in the biomedical domain. The code and datasets are publicly available.

TextRank is a variant of PageRank typically used in graphs that represent documents, and where vertices denote terms and edges denote relations between terms. Quite often the relation between terms is simple term co-occurrence within a fixed window of k terms. The output of TextRank when applied iteratively is a score for each vertex, i.e. a term weight, that can be used for information retrieval (IR) just like conventional term frequency based term weights.