Recent studies have explored the capabilities of large language models (LLMs) in solving knowledge-intensive mathematical reasoning problems. However, existing benchmarks predominantly involve static theorems that LLMs have encountered during pretraining, failing to assess dynamic knowledge integration. In this work, we introduce TaxReasoning, a novel benchmark designed to evaluate LLMs’ abilities in real-world tax calculation scenarios. These tasks require not only mathematical reasoning and numerical computation, but also the extraction and application of complex, frequently updated tax regulations. Through extensive experiments with state-of-the-art LLMs using diverse prompting strategies and knowledge augmentation techniques, we uncover substantial limitations in their ability to handle dynamic, knowledge-intensive questions—primarily due to missing domain-specific knowledge and ineffective retrieval. Even the best-performing models fall significantly short of human-level performance. Our analysis points to key avenues for improvement, including enhancing LLMs' reasoning capabilities, developing more effective knowledge summarization techniques, and improving retrieval strategies. TaxReasoning offers a critical testbed for advancing LLMs in dynamic knowledge-intensive domains.

Nan Hu, Jiaoyan Chen, Yike Wu, Guilin Qi, Hongru Wang, Sheng Bi, Yongrui Chen, Tongtong Wu, Jeff Z. Pan. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Weixiang Zhao, Yulin Hu, Yang Deng, Tongtong Wu, Wenxuan Zhang, Jiahe Guo, An Zhang, Yanyan Zhao, Bing Qin, Tat-Seng Chua, Ting Liu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Continual Structured Knowledge Reasoning (CSKR) focuses on training models to handle sequential tasks, where each task involves translating natural language questions into structured queries grounded in structured knowledge. Existing general continual learning approaches face significant challenges when applied to this task, including poor generalization to heterogeneous structured knowledge and inefficient reasoning due to parameter growth as tasks increase. To address these limitations, we propose a novel CSKR framework, \textsc{K-DeCore}, which operates with a fixed number of tunable parameters. Unlike prior methods, \textsc{K-DeCore} introduces a knowledge decoupling mechanism that disentangles the reasoning process into task-specific and task-agnostic stages, effectively bridging the gaps across diverse tasks. Building on this foundation, \textsc{K-DeCore} integrates a dual-perspective memory consolidation mechanism for distinct stages and introduces a structure-guided pseudo-data synthesis strategy to further enhance the model's generalization capabilities. Extensive experiments on four benchmark datasets demonstrate the superiority of \textsc{K-DeCore} over existing continual learning methods across multiple metrics, leveraging various backbone large language models.

Multilingual reasoning remains a significant challenge for large language models (LLMs), with performance disproportionately favoring high-resource languages. Drawing inspiration from cognitive neuroscience, which suggests that human reasoning functions largely independently of language processing, we hypothesize that LLMs similarly encode reasoning and language as separable components that can be disentangled to enhance multilingual reasoning. To evaluate this, we perform a causal intervention by ablating language-specific representations at inference time. Experiments on 10 open-weight LLMs spanning 11 typologically diverse languages show that this language-specific ablation consistently boosts multilingual reasoning performance. Layer-wise analyses further confirm that language and reasoning representations can be effectively disentangled throughout the model, yielding improved multilingual reasoning capabilities, while preserving top-layer language features remains essential for maintaining linguistic fidelity. Compared to post-training methods such as supervised fine-tuning or reinforcement learning, our training-free language-reasoning disentanglement achieves comparable or superior results with minimal computational overhead. These findings shed light on the internal mechanisms underlying multilingual reasoning in LLMs and suggest a lightweight and interpretable strategy for improving cross-lingual generalization.

When we integrate factual knowledge from knowledge graphs (KGs) into large language models (LLMs) to enhance their performance, the cost of injection through training increases with the scale of the models. Consequently, there is significant interest in developing prompt strategies that effectively incorporate KG information into LLMs. However, the community has not yet comprehensively understood how LLMs process and interpret KG information in different input formats and organizations within prompts, and researchers often rely on trial and error. To address this gap, we design extensive experiments to empirically study LLMs’ comprehension of different KG prompts. At the literal level, we reveal LLMs’ preferences for various input formats (from linearized triples to fluent natural language text). At the attention distribution level, we discuss the underlying mechanisms driving these preferences. We then investigate how the organization of structured knowledge impacts LLMs and evaluate LLMs’ robustness in processing and utilizing KG information in practical scenarios. Our experiments show that (1) linearized triples are more effective than fluent NL text in helping LLMs understand KG information and answer fact-intensive questions; (2) Different LLMs exhibit varying preferences for different organizational formats of triples; (3) LLMs with larger scales are more susceptible to noisy, incomplete subgraphs.

Conventional text-to-SQL studies are limited to a single task with a fixed-size training and test set. When confronted with a stream of tasks common in real-world applications, existing methods struggle with the problems of insufficient supervised data and high retraining costs. The former tends to cause overfitting on unseen databases for the new task, while the latter makes a full review of instances from past tasks impractical for the model, resulting in forgetting of learned SQL structures and database schemas. To address the problems, this paper proposes integrating semi-supervised learning (SSL) and continual learning (CL) in a stream of text-to-SQL tasks and offers two promising solutions in turn. The first solution Vanilla is to perform self-training, augmenting the supervised training data with predicted pseudo-labeled instances of the current task, while replacing the full volume retraining with episodic memory replay to balance the training efficiency with the performance of previous tasks. The improved solution SFNet takes advantage of the intrinsic connection between CL and SSL. It uses in-memory past information to help current SSL, while adding high-quality pseudo instances in memory to improve future replay. The experiments on two datasets shows that SFNet outperforms the widely-used SSL-only and CL-only baselines on multiple metrics.

Federated learning (FL) is a scheme that enables multiple participants to cooperate to train a high-performance machine learning model in a way that data cannot be exported. FL effectively protects the data privacy of all participants and reduces communication costs. However, a key challenge for federated learning is the data heterogeneity across clients. In addition, in real FL applications, the class distribution of data is usually unbalanced. Although many researches have been conducted to solve the problem of data heterogeneity, class imbalance problem usually arises along with the heterogeneity data, resulting in the poor performance of the global model. In this paper, a novel FL method (we call it FedEF) is designed for heterogeneous data and local class imbalance problem via optimize feature extractors and classifiers. FedEF optimizes the local feature extractor representation of individual clients through contrastive learning to maximize the consistency of the feature extractor representation trained by the local client and the central server to handle heterogeneous data. Meanwhile, we modified the cross entropy loss in the model, assigned different loss weights to different classes of data, paid more attention to the class with fewer samples in the training process, and corrected the biased classifier to alleviate the problem of class imbalance, thus can improve the performance of the global model. Experiments show that FedEF is an effective solution to FL model obtained under heterogeneous and local class imbalance.

Abstract Event extraction is an important, but challenging task. Many existing techniques decompose it into event and argument detection/classification subtasks, which are complex structured prediction problems. Generation-based extraction techniques lessen the complexity of the problem formulation and are able to leverage the reasoning capabilities of large pretrained language models. However, they still suffer from poor zero-shot generalizability and are ineffective in handling long contexts such as documents. We propose a generative event extraction model, KC-GEE, that addresses these limitations. A key contribution of KC-GEE is a novel knowledge-based conditioning technique that injects the schema of candidate event types as the prefix into each layer of an encoder-decoder language model. This enables effective zero-shot learning and improves supervised learning. Our experiments on two benchmark datasets demonstrate the strong performance of our KC-GEE model. It achieves particularly strong results in the challenging document-level extraction task and in the zero-shot learning setting, outperforming state-of-the-art models by up to 5.4 absolute F1 points.

This paper studies event causality identification, which aims at predicting the causality relation for a pair of events in a sentence. Regarding event causality identification as a supervised classification task, most existing methods suffer from the problem of insufficient annotated data. In this paper, we propose a new derivative prompt joint learning model for event causality identification, which leverages potential causal knowledge in the pre-trained language model to tackle the data scarcity problem. Specifically, rather than external data or knowledge augmentation, we derive two relevant prompt tasks from event causality identification to enhance the model’s ability to identify explicit and implicit causality. We evaluate our model on two benchmark datasets and the results show that our model has great advantages over previous methods.

In this paper, we propose a variational autoencoder with disentanglement priors, VAE-DPRIOR, for task-specific natural language generation with none or a handful of taskspecific labeled examples.In order to tackle compositional generalization across tasks, our model performs disentangled representation learning by introducing a conditional prior for the latent content space and another conditional prior for the latent label space.Both types of priors satisfy a novel property called ϵ-disentangled.We show both empirically and theoretically that the novel priors can disentangle representations even without specific regularizations as in the prior work.The content prior enables directly sampling diverse content representations from the content space learned from the seen tasks, and fuse them with the representations of novel tasks for generating semantically diverse texts in the low-resource settings.Our extensive experiments demonstrate the superior performance of our model over competitive baselines in terms of i) data augmentation in continuous zero/few-shot learning, and ii) text style transfer in the few-shot setting.The code is available at https://github. com/zhuang-li/VAE-DPrior.

Relation extraction has focused on extracting semantic relationships between entities from the unstructured written textual data.However, with the vast and rapidly increasing amounts of spoken data, relation extraction from speech is an important but under-explored problem. In this paper, we propose a new information extraction task, speech relation extraction (SpeechRE).To facilitate further research, we construct the first synthetic training datasets, as well as the first human-spoken test set with native English speakers.We establish strong baseline performance for SpeechRE via two approaches.The pipeline approach connects a pretrained ASR module with a text-based relation extraction module.The end-to-end approach employs a cross-modal encoder-decoder architecture.Our comprehensive experiments reveal the relative strengths and weaknesses of these approaches, and shed light on important future directions in SpeechRE research.We share the source code and datasets on https://github.com/ wutong8023/SpeechRE.* denotes the equal contribution.

Continual relation extraction is an important task that focuses on extracting new facts incrementally from unstructured text. Given the sequential arrival order of the relations, this task is prone to two serious challenges, namely catastrophic forgetting and order-sensitivity. We propose a novel curriculum-meta learning method to tackle the above two challenges in continual relation extraction. We combine meta learning and curriculum learning to quickly adapt model parameters to a new task and to reduce interference of previously seen tasks on the current task. We design a novel relation representation learning method through the distribution of domain and range types of relations. Such representations are utilized to quantify the difficulty of tasks for the construction of curricula. Moreover, we also present novel difficulty-based metrics to quantitatively measure the extent of order-sensitivity of a given model, suggesting new ways to evaluate model robustness. Our comprehensive experiments on three benchmark datasets show that our proposed method outperforms the state-of-the-art techniques. The code is available at the anonymous GitHub repository: https://github.com/wutong8023/AAAI_CML.

Knowledge graph representation learning, also called knowledge graph embedding, is the task of mapping entities and relations into a low-dimensional, continuous vector space, and, as a result, can support various machine learning models to perform knowledge completion tasks with good performance and robustness. However, most of existing embedding models focus on improving the link prediction accuracy while ignoring the time-efficiency in search-intensive applications over large-scale knowledge graphs. To tackle this problem, in this paper, we encode knowledge graph into Hamming space and introduce a novel HAsh Learning Framework (HALF) for search-oriented knowledge graph embedding. The proposed method can be applied to recent various knowledge graph embedding models for accelerating the computation of searching embeddings by utilizing the bitwise operations (XNOR and Bitcount). Experimental results on benchmark datasets demonstrate the effectiveness of our proposed method, which gets a bonus of speed-up in the searching embeddings while the accuracy and scalability of the original model are basically maintained.

Complex question-answering (CQA) involves answering complex natural-language questions on a knowledge base (KB). However, the conventional neural program induction (NPI) approach exhibits uneven performance when the questions have different types, harboring inherently different characteristics, e.g., difficulty level. This paper proposes a meta-reinforcement learning approach to program induction in CQA to tackle the potential distributional bias in questions. Our method quickly and effectively adapts the meta-learned programmer to new questions based on the most similar questions retrieved from the training data. The meta-learned policy is then used to learn a good programming policy, utilizing the trial trajectories and their rewards for similar questions in the support set. Our method achieves state-of-the-art performance on the CQA dataset (Saha et al., 2018) while using only five trial trajectories for the top-5 retrieved questions in each support set, and meta-training on tasks constructed from only 1% of the training set. We have released our code at https://github.com/DevinJake/MRL-CQA.

BACKGROUND: Survival time is an important clinical trait for many disease studies. Previous works have shown certain relationship between patients' gene expression profiles and survival time. However, due to the censoring effects of survival time and the high dimensionality of gene expression data, effective and unbiased selection of a gene expression signature to predict survival probabilities requires further study. METHOD: We propose a method for an integrated study of survival time and gene expression. This method can be summarized as a two-step procedure: in the first step, a moderate number of genes are pre-selected using correlation or liquid association (LA). Imputation and transformation methods are employed for the correlation/LA calculation. In the second step, the dimension of the predictors is further reduced using the modified sliced inverse regression for censored data (censorSIR). RESULTS: The new method is tested via both simulated and real data. For the real data application, we employed a set of 295 breast cancer patients and found a linear combination of 22 gene expression profiles that are significantly correlated with patients' survival rate. CONCLUSION: By an appropriate combination of feature selection and dimension reduction, we find a method of identifying gene expression signatures which is effective for survival prediction.