Xueqing Peng, Lingfei Qian, Yan Wang, Ruoyu Xiang, Yueru He, Yang Ren, Mingyang Jiang, Vincent Jim Zhang, Yuqing Guo, Jeff Zhao, Huan He, Yi Han, Yun Feng, Yuechen Jiang, Yupeng Cao, Haohang Li, Yangyang Yu, Xiaoyu Wang, Penglei Gao, Shengyuan Lin, Keyi Wang, Shanshan Yang, Yilun Zhao, Zhiwei Liu, Peng Lu, Jerry Huang, Suyuchen Wang, Triantafillos Papadopoulos, Polydoros Giannouris, Efstathia Soufleri, Nuo Chen, Zhiyang Deng, Heming Fu, Yijia Zhao, Mingquan Lin, Meikang Qiu, Kaleb E Smith, Arman Cohan, Xiao-Yang Liu, Jimin Huang, Guojun Xiong, Alejandro Lopez-Lira, Xi Chen, Junichi Tsujii, Jian-Yun Nie, Sophia Ananiadou, Qianqian Xie. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Haohang Li, Yupeng Cao, Yangyang Yu, Shashidhar Reddy Javaji, Zhiyang Deng, Yueru He, Yuechen Jiang, Zining Zhu, K.p. Subbalakshmi, Jimin Huang, Lingfei Qian, Xueqing Peng, Jordan W. Suchow, Qianqian Xie. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The attention mechanism has been successfully used in a sequence consisted of a series of word embeddings to improve the representation of the sequence. Inspired by this, we leverage the attention mechanism on a set of tasks to implement a multi-task transfer learning method called AMP (Attentions between Multiple Prompts). First, we encode a task into a prompt as task representation called task embedding. Second, we learn an attention component on all task embeddings to generate a combined prompt for each task, which is an attention-weighted sum of task embeddings. Each combined prompt incorporates the knowledge of all tasks. The word embedding is a vector, but the task embedding is a 2D matrix. The attention mechanism can be exploited on a set of vectors rather than on a set of matrices. The prior methods employ pooling or flattened method to transform the matrix to the vector for computing the attentions between matrices. We propose a method called DAM (Direct Attention Mechanism) which can compute attentions between matrices directly without transforming. DAM method can more exactly compute the attentions between matrices. Wide experiments demonstrate that AMP outperforms prompt-tuning method and prior prompt transfer methods.

Automatic classification and survival prediction of pancreatic cancer based on computed tomography (CT) images have become a key research direction in the field of intelligent healthcare. Various deep learning-based diagnostic models have been developed for classifying pancreatic cancer from CT images. However, CT images often contain noise and a significant amount of irrelevant information unrelated to pancreatic cancer diagnosis. This irrelevant information weakens the discriminative power of the extracted features, leading to the problem of information redundancy. In addition, existing models generally suffer from poor feature generalization. To address these challenges, we propose a Deep Reconstruction Bottleneck Representation Network (DRBRN) for pancreatic cancer diagnosis, based on Internet of Medical Things (IoMT) and CT imaging. In essence, the DRBRN leverages the information bottleneck principle and a reconstruction task to learn more compact and generalizable features from CT images. It effectively filters out redundant information and noise, preserving only the critical features relevant to the diagnosis of pancreatic cancer. The experimental results on real-world pancreatic cancer CT dataset demonstrate that our proposed DRBRN model achieved the best performance across all three metrics, with a Precision of 0.996, Accuracy of 0.995, and F1-Score of 0.995.

We introduceFinMem, a novel Large Language Models (LLM)-based agent framework for financial trading, designed to address the need for automated systems that can transform real-time data into executable decisions.FinMemcomprises three core modules: Profile for customizing agent characteristics, Memory for hierarchical financial data assimilation, and Decision-making for converting insights into investment choices. The Memory module, which mimics human traders' cognitive structure, offers interpretability and real-time tuning while handling the critical timing of various information types. It employs a layered approach to process and prioritize data based on its timeliness and relevance, ensuring that the most recent and impactful information is given appropriate weight in decision-making.FinMem's adjustable cognitive span allows retention of critical information beyond human limits, enabling it to balance historical patterns with current market dynamics. This framework facilitates self-evolution of professional knowledge, agile reactions to investment cues, and continuous refinement of trading decisions in financial environments. When compared against advanced algorithmic agents using a large-scale real-world financial dataset,FinMemdemonstrates superior performance across classic metrics like Cumulative Return and Sharpe ratio. Further tuning of the agent's perceptual span and character setting enhances its trading performance, positioningFinMemas a cutting-edge solution for automated trading.

Cross-scene classification of remote sensing (RS) imagery faces significant challenges due to large domain discrepancies and diverse imaging conditions, which hinder the acquisition of generalizable multi-modal semantic features. Unsupervised domain adaptation (UDA) has emerged as a promising solution for knowledge transfer across domains. Although prompt learning and vision-language models (VLMs) have recently been employed to enrich cross-modal information and improve the effectiveness of UDA, most existing methods either design prompts solely for the textual modality or employ unidirectional mapping from textual to visual prompts, resulting in insufficient cross-modal fusion. To overcome these limitations, we propose Mutual Cascade Prompting for Probability-Aligned Unsupervised Domain Adaptation in cross-scene classification (MCPPA). Within this framework, the progressive attention interactive prompt (PAIP) module implements a cascade prompt mechanism where, at each layer, visual and textual prompts from the previous stage are mutually updated through a cross-attention module. This mutual learning allows both branches to progressively integrate and reinforce multi-modal information in a deep and hierarchical fashion. Additionally, the category probability discrepancy measure (CPDM) module aligns the probability matrices of source and target domains using the nuclear norm within an adversarial training paradigm, while the feature stability constraint (FSC) module enforces consistency between learnable prompt features and those from the pre-trained model, enhancing overall robustness. Extensive evaluations on 30 transfer tasks across 10 benchmark RS datasets confirm that MCPPA consistently outperforms existing advanced methods, highlighting its effectiveness and broad application potential.

With the rapid advancement of remote sensing (RS) technologies, the role of automated cross-scene recognition in environmental monitoring and resource management has become increasingly prominent. Facing the challenge of scarce annotated data in RS, unsupervised domain adaptation (UDA) technology stands out for its ability to transfer knowledge across domains. Most RS scene recognition methods based on prompt learning only optimize the textual component. They cannot flexibly and dynamically adjust textual and visual representations, which may lead to poor performance when facing complex UDA tasks. In order to solve this problem, we propose the prompt-integrated adversarial UDA for scene recognition (PADA-Net) framework, introducing a prompt-integrated method in the domain adaptation task of RS scene classification for the first time to enhance the semantic association between images and text and better align visual-linguistic representations. PADA-Net fosters cross-modal information exchange via an interactive bridging mechanism (IBM) and combines dual Meta-nets to reinforce feature discriminability. The collaborative operation of these two components constitutes a novel system for feature discrimination. Additionally, we incorporate optimal transport theory to provide meaningful gradients and geometric guidance for training, and we use game-theoretic strategies to further enhance the efficient alignment of feature distributions, thereby addressing the domain shift problem. Finally, we carry out 24 different scene recognition tasks on multiple RS benchmark datasets, such as AID and WHU-RS19, and several experimental findings verify the excellence of our suggested approach.

News sentiment is different from the true investor sentiment, and there is a conductive process of information flow from news sentiment to the latent investor sentiment and vice versa. This study aims to develop a methodology to estimate the latent effect between the investor sentiment jumps and the market return jumps using a multivariate Hawkes process along with a deep reinforcement learning algorithm. We achieve this goal through a three-step process: (i) identify the baseline intensity among the events of news sentiment and market return by a multivariate Hawkes process; (ii) estimate the hidden effect that drives the movement of events of news sentiment and market return from the baseline intensity via deep reinforcement learning; (iii) reveal the interaction mechanism among the true investor sentiment and the market return that is responsible to the latent investor sentiment. This approach can be broadly applied to analyzing many phenomena in finance and economics where latent events are non-stationary and can not be observed directly.

LLMs have transformed NLP and shown promise in various fields, yet their potential in finance is underexplored due to a lack of comprehensive benchmarks, the rapid development of LLMs, and the complexity of financial tasks. In this paper, we introduce FinBen, the first extensive open-source evaluation benchmark, including 42 datasets spanning 24 financial tasks, covering eight critical aspects: information extraction (IE), textual analysis, question answering (QA), text generation, risk management, forecasting, decision-making, and bilingual (English and Spanish). FinBen offers several key innovations: a broader range of tasks and datasets, the first evaluation of stock trading, novel agent and Retrieval-Augmented Generation (RAG) evaluation, and two novel datasets for regulations and stock trading. Our evaluation of 21 representative LLMs, including GPT-4, ChatGPT, and the latest Gemini, reveals several key findings: While LLMs excel in IE and textual analysis, they struggle with advanced reasoning and complex tasks like text generation and forecasting. GPT-4 excels in IE and stock trading, while Gemini is better at text generation and forecasting. Instruction-tuned LLMs improve textual analysis but offer limited benefits for complex tasks such as QA. FinBen has been used to host the first financial LLMs shared task at the FinNLP-AgentScen workshop during IJCAI-2024, attracting 12 teams. Their novel solutions outperformed GPT-4, showcasing FinBen's potential to drive innovations in financial LLMs. All datasets and code are publicly available for the research community, with results shared and updated regularly on the Open Financial LLM Leaderboard.

Large language models (LLMs) have demonstrated notable potential in conducting complex tasks and are increasingly utilized in various financial applications. However, high-quality sequential financial investment decision-making remains challenging. These tasks require multiple interactions with a volatile environment for every decision, demanding sufficient intelligence to maximize returns and manage risks. Although LLMs have been used to develop agent systems that surpass human teams and yield impressive investment returns, opportunities to enhance multi-source information synthesis and optimize decision-making outcomes through timely experience refinement remain unexplored. Here, we introduce FinCon, an LLM-based multi-agent framework tailored for diverse financial tasks. Inspired by effective real-world investment firm organizational structures, FinCon utilizes a manager-analyst communication hierarchy. This structure allows for synchronized cross-functional agent collaboration towards unified goals through natural language interactions and equips each agent with greater memory capacity than humans. Additionally, a risk-control component in FinCon enhances decision quality by episodically initiating a self-critiquing mechanism to update systematic investment beliefs. The conceptualized beliefs serve as verbal reinforcement for the future agent’s behavior and can be selectively propagated to the appropriate node that requires knowledge updates. This feature significantly improves performance while reducing unnecessary peer-to-peer communication costs. Moreover, FinCon demonstrates strong generalization capabilities in various financial tasks, including stock trading and portfolio management.

Benign epilepsy with centrotemporal spikes (BECTS) is a common childhood epilepsy syndrome associated with abnormalities in neurocognitive domains, particularly during interictal epileptiform discharges (IEDs). Here, we investigated the effects of IEDs on brain's intrinsic connectivity networks in 43 BECTS patients and 28 matched healthy controls (HCs). Patients were further divided into IED and non-IED subgroups based on simultaneous EEG-fMRI recordings. Functional connectivity within and between five networks, corresponding to seizure origination and cognitive processes, were analyzed to measure IED effects. We found that patients exhibited increased connectivity within the auditory network (AN) and the somato-motor network (SMN), and decreased connectivity within the basal ganglia network and the dorsal attention network, suggesting that both transient and chronic seizure activity may disturb normal network organization. The IED group showed decreased functional connectivity within the default mode network (DMN) compared with the non-IED group and HCs, implying that the DMN was selectively impaired during epileptiform discharges associated with altered self-referential cognitive functions. Moreover, the IED group exhibited increased positive correlations between the AN and the SMN, which suggests a possible excessive influence of centrotemporal spiking on information processing in the auditory system. The association between epileptic activity and network dysfunctions highlights their importance in investigating the pathological mechanism underlying BECTS.

The software fault failure rate bound is discussed and generalized for different reliability growth models. The fault introduction during testing and the fault removal efficiency are modeled to relax the two common assumptions made in software reliability models. Three approaches are introduced for the fault content estimation, and thus they are applied to software coverage estimation. A three-state nonhomogenous Markov model is constructed for software safety assessment. The two most important metrics for safety assessment, Steady State Safety and MTTUF, are estimated using the three-state Markov model. A case study is conducted to verify the theory proposed in the paper.