Kunpeng Kang, Shuaimin Li, Kaiyuan Zhang, Luyang Zhang, Jiasheng Si, Bing Xu, Kehai Chen, Muyun Yang, Wenpeng Lu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Yeqing Teng, Jiasheng Si, Shuxia Lin, Linhai Zhang, Weiyu Zhang, Wenpeng Lu, Deyu Zhou, Xiaoming Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large language models (LLMs) have demonstrated strong reasoning capabilities, and as existing approaches for enhancing LLM reasoning continue to mature, increasing attention has shifted toward meta-reasoning as a promising direction for further improvement.However, most existing meta-reasoning methods remain episodic: they focus on executing complex meta-reasoning routines within individual instances, but ignore the accumulation of reusable meta-reasoning skills across instances, leading to recurring failure modes and repeatedly high metacognitive effort.In this paper, we introduce Metacognitive Consolidation, a novel framework in which a model consolidates metacognitive experience from past reasoning episodes into reusable knowledge that improves future meta-reasoning.We instantiate this framework by structuring instance-level problem solving into distinct roles for reasoning, monitoring, and control to generate rich, attributable meta-level traces.These traces are then consolidated through a hierarchical, multi-timescale update mechanism that gradually forms evolving metaknowledge.Experimental results demonstrate consistent performance gains across benchmarks and backbone models, and show that performance improves as metacognitive experience accumulates over time.

Scientific fact-checking has largely focused on textual and tabular sources, neglecting scientific charts-a primary medium for conveying quantitative evidence and supporting statistical reasoning in research communication.We introduce CLIMATEVIZ, the first largescale benchmark for scientific fact-checking grounded in real-world, expert-curated scientific charts.CLIMATEVIZ comprises 49,862 claims paired with 2,896 visualizations, each labeled as support, refute, or not enough information.To enable interpretable verification, each instance includes structured knowledge graph explanations that capture statistical patterns, temporal trends, spatial comparisons, and causal relations.We conduct a comprehensive evaluation of state-of-the-art multimodal large language models, including proprietary and open-source systems, under zero-shot and few-shot settings.Our results show that current models struggle to perform fact-checking when statistical reasoning over charts is required: even the best-performing systems, such as Gemini 2.5 and InternVL 2.5, achieve only 76.2-77.8%accuracy in label-only output settings, which is far below human performance (89.3% and 92.7%).While few-shot prompting yields limited improvements, explanationaugmented outputs significantly enhance performance in some closed-source models, notably o3 and Gemini 2.5.We released our dataset and code alongside the paper. 1(c) Subgraph of Relevant Facts Caption: Cumulative mass loss of the Greenland Ice Sheet from 1972 to 2022, showing accelerating ice loss and corresponding sea level rise.

Argumentative essay generation (AEG) is a complex task that requires advanced semantic understanding, logical reasoning, and organized integration of perspectives.Despite showing a promising performance, current efforts often overlook the dynamical and hierarchical nature of structural argumentative planning, and struggle with flexible rhetorical expression, leading to limited argument divergence and rhetorical optimization.Inspired by human debate behavior and Bitzer's rhetorical situation theory, we propose a debate-driven rhetorical framework for argumentative writing.The uniqueness lies in three aspects: (1) it dynamically assesses the divergence of viewpoints and progressively reveals the hierarchical outline of arguments based on a depththen-breadth paradigm, improving the perspective divergence within argumentation; (2) simulates human debate through iterative defenderattacker interactions, improving the logical coherence of arguments; (3) incorporates Bitzer's rhetorical situation theory to flexibly select appropriate rhetorical techniques, enabling the rhetorical expression.Experiments on four benchmarks validate that our approach significantly improves logical depth, argumentative diversity, and rhetorical persuasiveness over existing state-of-the-art models 1 .* Corresponding authors. 1 Code and data are available at https://github.com/ zxg-x/DARE Social media affects attention and distracts people.Social media affects attention and distracts people.Social media affects attention and distracts people.Social media is a waste of time.Social media is a waste of time.Social media is a waste of time.

Slot filling and intent detection are two highly correlated tasks in spoken language understanding (SLU). Recent SLU research attempts to explore zero-shot prompting techniques in large language models to alleviate the data scarcity problem. Nevertheless, the existing prompting work ignores the cross-task interaction information for SLU, which leads to sub-optimal performance. To solve this problem, we present the pioneering work of Cross-task Interactive Prompting (CroPrompt) for SLU, which enables the model to interactively leverage the information exchange across the correlated tasks in SLU. Additionally, we further introduce a multi-task self-consistency mechanism to mitigate the error propagation caused by the intent information injection. We conduct extensive experiments on the standard SLU benchmark and the results reveal that CroPrompt consistently outperforms the existing prompting approaches. In addition, the multi-task self-consistency mechanism can effectively ease the error propagation issue, thereby enhancing the performance. We hope this work can inspire more research on cross-task prompting for SLU.

The remarkable prompting capability of large language models (LLMs) offers substantial convenience to users across diverse backgrounds. Nevertheless, as the sensitive information within prompts is inevitably exposed to LLMs, caution must be exercised to preserve privacy. Among various studies, text anonymization is considered an effective approach to preventing privacy leakage in prompts through text substitution. However, existing works overemphasize privacy while overlooks preserving contextual integrity, degrading semantic consistency. To address these concerns, this paper introduces a context-sensitive prompt privacy-preserving framework, namely Adaptive Linguistic Sanitization and Anonymization (ALSA). In specific, ALSA incorporates a three-dimensional scoring mechanism to dynamically quantify the substitutability of each word within a prompt by integrating the Privacy Leakage Risk Score (PLRS), the Contextual Information Importance Score (CIIS), and the Task Relevance Score (TRS). Subsequently, a clustering technique is adopted to dynamically determine the threshold for assigning an anonymization action (i.e., Retain, Replace, Encrypt, or Delete) by balancing privacy, semantics, and task relevance. Extensive experiments on five benchmark datasets validate the superiority of ALSA over state-of-the-art baselines in terms of accuracy, privacy preservation, and semantic integrity.

Medication recommendation aims to suggest personalized drug combinations to patients based on their longitudinal medical histories stored in electronic health record (EHR) datasets. Patients' Dynamic Treatment Regimes (DTRs) determine how patients' drug combinations change along with the evolution of disease treatment. DTRs are effective for comprehending disease-treatment dynamics and for recommending a timely and personalized combination of medications for patients. However, existing medication recommender systems (MRSs) overlook the multiple treatment pathways generated by the intervention of DTRs and can only recommend a single treatment paradigm, ignoring the fact that patients may be at different treatment stages and thus require different treatment regime. Such disregard leads to a significant limitation in recommending personalized medication combinations tailored to different treatment stages, yielding greatly compromised accuracy and applicability of MRSs. Moreover, existing methods often overlook the time interval information over patients' successive visits, which is critical to indicate patients' treatment evolution. To address these significant gaps, we propose a Time-aware Medication Recommendation Framework via Intervention of Dynamic Treatment Regimes, called MR-DTR. To explicitly illustrate the intervention processes of DTRs on similar patients, we employ a co-guided graph to connect various patient sequences. In addition, to fully utilize the time interval information, we design a time-aware guidance mechanism dedicated to the co-guided graph to efficiently learn medication representation using the patient's guidance information. We also introduce relative time intervals in the encoder to act as positional information. Extensive experiments on two real-world datasets demonstrate that MR-DTR surpasses state-of-the-art models in terms of recommendation performance. Our code is available at: https://github.com/liyifo/MR-DTR.

Multi-modal hashing achieves low storage costs and high retrieval speeds by using compact hash codes to represent complex and heterogeneous multi-modal data, effectively addressing the inefficiency and resource intensiveness challenges faced by the traditional multi-modal retrieval methods. However, balancing intraclass compactness and interclass separability remains a struggle in existing works due to coarse-grained feature limitations, simplified fusion strategies that overlook semantic complementarity, and neglect of the structural information within the multi-modal data. To address these limitations comprehensively, we propose a Proto-centric Multi-modal Hashing with Pronounced Category Differences (PMH-PCD) model. Specifically, PMH-PCD first learns modality-specific prototypes by deeply exploring within-modality class information, ensuring effective fusion of each modality's unique characteristics. Furthermore, it learns multi-modal integrated class prototypes that seamlessly incorporate semantic information across modalities to effectively capture and represent the intricate relationships and complementary semantic content embedded within the multi-modal data. Additionally, to generate more discriminative and representative binary hash codes, PMH-PCD integrates multifaceted semantic information, encompassing both low-level pairwise relations and high-level structural patterns, holistically capturing intricate data details and leveraging underlying structures. The experimental results demonstrate that, compared with existing advanced methods, PMH-PCD achieves superior and consistent performances in multi-modal retrieval tasks. To promote further research and reproducibility, we have publicly released the source code of PMH-PCD at https://github.com/vindahi/PMH-PCD.

Multi-modal sarcasm detection involves determining whether a given multi-modal input conveys sarcastic intent by analyzing the underlying sentiment. Recently, vision large language models have shown remarkable success on various of multi-modal tasks. Inspired by this, we systematically investigate the impact of vision large language models in zero-shot multi-modal sarcasm detection task. Furthermore, to capture different perspectives of sarcastic expressions, we propose a multi-view agent framework, S 3 Agent, designed to enhance zero-shot multi-modal sarcasm detection by leveraging three critical perspectives: superficial expression , semantic information , and sentiment expression . Our experiments on the MMSD2.0 dataset, which involves six models and four prompting strategies, demonstrate that our approach achieves state-of-the-art performance. Our method achieves an average improvement of 13.2% in accuracy. Moreover, we evaluate our method on the text-only sarcasm detection task, where it also surpasses baseline approaches.

With the growing complexity of fact verification tasks, the concern with "thoughtful" reasoning capabilities is increasing.However, recent fact verification benchmarks mainly focus on checking a narrow scope of semantic factoids within claims and lack an explicit logical reasoning process.In this paper, we introduce CHECKWHY, a challenging dataset tailored to a novel causal fact verification task: checking the truthfulness of the causal relation within claims through rigorous reasoning steps.CHECKWHY consists of over 19K "why" claimevidence-argument structure triplets with supports, refutes, and not enough info labels.Each argument structure is composed of connected evidence, representing the reasoning process that begins with foundational evidence and progresses toward claim establishment.Through extensive experiments on state-of-the-art models, we validate the importance of incorporating the argument structure for causal fact verification.Moreover, the automated and human evaluation of argument structure generation reveals the difficulty in producing satisfying argument structure by fine-tuned models or Chainof-Thought prompted LLMs, leaving considerable room for future improvements 1 .

Electronic health records (EHRs) play a crucial role in the development of personalized treatment plans for patients. However, EHRs are often highly incomplete, posing significant challenges for predictive modeling. While existing deep learning models employ various imputation techniques to reconstruct missing values, they often fail to represent missing data in a personalized manner and do not learn from the missingness patterns in EHRs data. This limitation reduces their effectiveness in practical personalized healthcare applications. To address this issue, we propose SPIME, a self-supervised model that generates personalized imputations in EHRs data for patient health status prediction. We introduce a personalized missing mask (PMM) based on the frequency of feature measurements. Additionally, we incorporate a masked imputation task (MIT) loss that minimizes the loss of artificially introduced missing values, thereby enhancing the model’s capability to handle missing data. SPIME adopts self-supervised pretraining to learn representations from personalized missing patterns and reconstructs missing data in the latent space. To further enhance representation learning, two independent attention mechanisms are applied separately across the feature and temporal dimensions. Experimental results on two real-world EHRs datasets show that SPIME outperforms existing state-of-the-art methods in predicting in-hospital mortality and decompensation, demonstrating its effectiveness in reconstructing missing data and predicting patient health status. The code will be published at https://github.com/cling6666/SPIME.

The inherent complexity of real-world time series data, combined with the cost and infeasibility of manual labeling, presents considerable challenges to time series representation learning. Most existing studies tend to utilize data augmentation techniques to construct positive and negative samples and leverage a comparative learning framework to generate time series representations. However, they typically employ simple data augmentation techniques, such as jitter and cropping, to construct positive samples while randomly selecting irrelevant samples as negative ones, which are easily distinguished and unable to guide comparative learning to capture subtle discriminative features. Furthermore, they usually employ only a single positive sample for comparative learning, which is insufficient to model the diversity and hurts the robustness. To address these issues, this paper proposes a Time Series representation learning framework via Dual Reference Contrasting (TS-DRC). Specifically, we first utilize Markov transition field or Gramian angular field to transform the anchor sample of time series into image representations, which are adopted as positive samples. Then, we incorporate two positive samples (dual references) and one negative sample into the comparative learning framework, and devise a novel optimization objective to guide the model to capture more discriminate features, mitigate overfitting, and enhance the robustness. Extensive experiments conducted on four public real-world datasets demonstrate that our TS-DRC outperforms other state-of-the-art baselines.Our code is available at: https://github.com/yurui12138/TS-DRC.

The opaqueness of the multi-hop fact verification model imposes imperative requirements for explainability. One feasible way is to extract rationales, a subset of inputs, where the performance of prediction drops dramatically when being removed. Though being explainable, most rationale extraction methods for multi-hop fact verification explore the semantic information within each piece of evidence individually, while ignoring the topological information interaction among different pieces of evidence. Intuitively, a faithful rationale bears complementary information being able to extract other rationales through the multi-hop reasoning process. To tackle such disadvantages, we cast explainable multi-hop fact verification as subgraph extraction, which can be solved based on graph convolutional network (GCN) with salience-aware graph learning. In specific, GCN is utilized to incorporate the topological interaction information among multiple pieces of evidence for learning evidence representation. Meanwhile, to alleviate the influence of noisy evidence, the salience-aware graph perturbation is induced into the message passing of GCN. Moreover, the multi-task model with three diagnostic properties of rationale is elaborately designed to improve the quality of an explanation without any explicit annotations. Experimental results on the FEVEROUS benchmark show significant gains over previous state-of-the-art methods for both rationale extraction and fact verification.

Automatic multi-hop fact verification task has gained significant attention in recent years.Despite impressive results, these well-designed models perform poorly on out-of-domain data.One possible solution is to augment the training data with counterfactuals, which are generated by minimally altering the causal features of the original data.However, current counterfactual data augmentation techniques fail to handle multi-hop fact verification due to their incapability to preserve the complex logical relationships within multiple correlated texts.In this paper, we overcome this limitation by developing a rationale-sensitive method to generate linguistically diverse and label-flipping counterfactuals while preserving logical relationships.In specific, the diverse and fluent counterfactuals are generated via an Explain-Edit-Generate architecture.Moreover, the checking and filtering modules are proposed to regularize the counterfactual data with logical relations and flipped labels.Experimental results show that the proposed approach outperforms the SOTA baselines and can generate linguistically diverse counterfactual data without disrupting their logical relationships 1 .

Biomedical argument mining aims to automatically identify and extract the argumentative structure in biomedical text. It helps to determine not only what positions people adopt, but also why they hold such opinions, which provides valuable insights into medical decision making. Generally, biomedical argument mining consists of three subtasks: argument component identification, argument component classification and relation identification. Current approaches employ conventional multi-task learning framework for jointly addressing the latter two subtasks, and achieve some success. However, explicit sequential dependency between these two subtasks is ignored, which is crucial for accurate biomedical argument mining. Moreover, relation identification is conducted solely based on the argument component pair without considering its potentially valuable context. Therefore, in this paper, a novel sequential multi-task learning approach is proposed for biomedical argument mining. Specifically, to model explicit sequential dependency between argument component classification and relation identification, an information transfer strategy is employed to capture the information of argument component type that is transferred to relation identification. Furthermore, graph convolutional network is employed to model dependency relation among the related argument component pairs. The proposed method has been evaluated on a benchmark dataset and the experimental results show that the proposed method outperforms the state-of-the-art methods.

Jiasheng Si, Deyu Zhou, Tongzhe Li, Xingyu Shi, Yulan He. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Storyline extraction aims to generate concise summaries of related events unfolding over time from a collection of temporally-ordered news articles. Some existing approaches to storyline extraction are typically built on probabilistic graphical models that jointly model the extraction of events and the storylines from news published in different periods. However, their parameter inference procedures are often complex and require a long time to converge, which hinders their use in practical applications. More recently, a neural network-based approach has been proposed to tackle such limitations. However, event representations of documents, which are important for the quality of the generated storylines, are not learned. In this paper, we propose a novel unsupervised neural network-based approach to extract latent events and link patterns of storylines jointly from documents over time. Specifically, event representations are learned by a stacked autoencoder and clustered for event extraction, then a fusion component is incorporated to link the related events across consecutive periods for storyline extraction. The proposed model has been evaluated on three news corpora and the experimental results show that it outperforms state-of-the-art approaches with significant improvements.