Language models excel in various tasks by making complex decisions, yet understanding the rationale behind these decisions remains a challenge.This paper investigates data-centric interpretability in language models, focusing on the next-word prediction task.Using representer theorem, we identify two types of support samples-those that either promote or deter specific predictions.Our findings reveal that being a support sample is an intrinsic property, predictable even before training begins.Additionally, while non-support samples are less influential in direct predictions, they play a critical role in preventing overfitting and shaping generalization and representation learning.Notably, the importance of non-support samples increases in deeper layers, suggesting their significant role in intermediate representation formation.These insights shed light on the interplay between data and model decisions, offering a new dimension to understanding language model behavior and interpretability. 1 * Equal contribution. 1 Our source code is publicly available at https://github.com/liyuqian44/ On-Support-Samples-of-Next-Word-Prediction.

Yufang Liu, Yao Du, Tao Ji, Jianing Wang, Yang Liu, Yuanbin Wu, Aimin Zhou, Mengdi Zhang, Xunliang Cai. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Daiye Miao, Yufang Liu, Jie Wang, Changzhi Sun, Yunke Zhang, Demei Yan, Shaokang Dong, Qi Zhang, Yuanbin Wu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

In this paper, we introduce BMMR, a large-scale bilingual, multimodal, multi-disciplinary reasoning dataset for the community to develop and evaluate large multimodal models (LMMs). BMMR comprises 100k university-level questions drawn from 300 UNESCO-defined subjects, spanning diverse formats—multiple-choice, fill-in-the-blank, and open-ended QA—and sourced from both print and digital media such as books, exams, and quizzes. All data are curated and filtered via a human-in-the-loop, automated, and scalable framework, and each instance is paired with a high-quality reasoning path. The dataset is organized into two parts: BMMR-Eval that comprises 20k high-quality instances to comprehensively assess LMMs’ knowledge and reasoning across multiple disciplines in both Chinese and English; and BMMR-Train that contains 80k instances to support further research and development, extending the current focus on mathematical reasoning to diverse disciplines and domains. In addition, we propose the process-based multi-discipline BMMR-Verifier for accurate and fine-grained evaluation of LMMs’ reasoning. Extensive experiments reveal that (i) even SOTA models leave substantial headroom on BMMR-Eval; (ii) reasoning models exhibit discipline bias and outperform LMMs only on specific subjects; (iii) open-source models still trail their proprietary counterparts; and (iv) fine-tuning on BMMR-Train narrows this gap. Additionally, we conduct reasoning-chain analyses using BMMR-Verifier and other in-depth studies, uncovering the challenges LMMs currently face in multidisciplinary reasoning. We will release the data and models, and we believe our work can offers valuable insights and contributions to the community.

Large Vision-Language Models (LVLMs) have achieved impressive performance, yet research has pointed out a serious issue with object hallucinations within these models.However, there is no clear conclusion as to which part of the model these hallucinations originate from.In this paper, we present an in-depth investigation into the object hallucination problem specifically within the CLIP model, which serves as the backbone for many state-of-the-art visionlanguage systems.We unveil that even in isolation, the CLIP model is prone to object hallucinations, suggesting that the hallucination problem is not solely due to the interaction between vision and language modalities.To address this, we propose a counterfactual data augmentation method by creating negative samples with a variety of hallucination issues.We demonstrate that our method can effectively mitigate object hallucinations for the CLIP model, and we show that the enhanced model can be employed as a visual encoder, effectively alleviating the object hallucination issue in LVLMs. 1

Multimodal integration combines information from different sources or modalities to gain a more comprehensive understanding of a phenomenon. The challenges in multi-omics data analysis lie in the complexity, high dimensionality, and heterogeneity of the data, which demands sophisticated computational tools and visualization methods for proper interpretation and visualization of multi-omics data. In this paper, we propose a novel method, termed Orthogonal Multimodality Integration and Clustering (OMIC), for analyzing CITE-seq. Our approach enables researchers to integrate multiple sources of information while accounting for the dependence among them. We demonstrate the effectiveness of our approach using CITE-seq data sets for cell clustering. Our results show that our approach outperforms existing methods in terms of accuracy, computational efficiency, and interpretability. We conclude that our proposed OMIC method provides a powerful tool for multimodal data analysis that greatly improves the feasibility and reliability of integrated data.

Abstract Sequential recommendation help users find interesting items by modeling the dynamic user-item interaction sequences. Due to the data sparseness problem, cross-domain sequential recommendation (CDSR) are proposed. CDSR explore rich data from a source domain to improve performance of the target domain. However, most of the existing CDSR methods are difficult to capture the temporal context of sequences and only learn user preference based on interactions of single domain, which leads to suboptimal performance. To address these shortcomings, we propose a channel-enhanced contrastive cross-domain sequential recommendation model (C3DSR). To be specific, (1) we design a feature extractor, which extends attention to the channel dimension, to extract the user’s channel feature and capture the temporal contextual relationships between sequences. Then we calculate the weights of each channel by using three SE-Res2Blocks and multiply it with the channel feature to obtain user preference. (2) We concatenate the user’s single-domain representation, the cross-domain representation, and the user features to make CDSR. Contrastive learning is leveraged to enhance mutual information between two domains. Experimental results show that the proposed model achieves the significant improvement of performance compared with other CDSR models on Amazon and HVIDEO datasets.

Abstract The cold-start problem in recommender systems has been facing a great challenge. Cross-domain recommendation can improve the performance of cold-start user recommendations in the target domain by using the rich information of users in the source domain. In cross-domain cold-start recommendation, users in target domain lack sufficient historical behaviors. Existing meta-learning-based methods depend on the feature distribution of training data and limit the adaptability in new tasks. To address these issues, we propose a meta-adversarial framework for cross-domain cold-start recommendation (MAFCDR) . Specifically, we employ a multi-level feature attention mechanism for independently learning the weights of long-term and short-term features to construct preferences of users in source domain. To migrate user representations, we train a meta-adversarial network that utilizes feature embeddings in the source domain as input and enhances the robustness and stability of the model. Then, the personalized bridge function transfers the user preferences in the source domain to the target domain. We build three cross-domain tasks using Amazon dataset and conduct extensive experiments, which demonstrate the effectiveness of the proposed model in cold-start user recommendation.

Zhicheng Wang, Yufang Liu, Tao Ji, Xiaoling Wang, Yuanbin Wu, Congcong Jiang, Ye Chao, Zhencong Han, Ling Wang, Xu Shao, Wenqiu Zeng. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Existing distantly supervised relation extractors usually rely on noisy data for both model training and evaluation, which may lead to garbage-in-garbage-out systems. To alleviate the problem, we study whether a small clean dataset could help improve the quality of distantly supervised models. We show that besides getting a more convincing evaluation of models, a small clean dataset also helps us to build more robust denoising models. Specifically, we propose a new criterion for clean instance selection based on influence functions. It collects sample-level evidence for recognizing good instances (which is more informative than loss-level evidence). We also propose a teacher-student mechanism for controlling purity of intermediate results when bootstrapping the clean set. The whole approach is model-agnostic and demonstrates strong performances on both denoising real (NYT) and synthetic noisy datasets.

For electronic systems equipped with early warning devices known as canaries, the early warning of failure can provide the opportunity for better spare-parts ordering, timely replacement of parts, and a resulting reduction in system life cycle costs. This paper presents both theoretical and simulation-based models for spare-part ordering and system replacement decisions for systems equipped with canaries. We first develop a theoretical model for the expected total cost per unit time in terms of the order interval, and maximum allowable stock level. This model is then compared with a simulation algorithm to validate the model. On the basis of the simulation algorithm, at each ordering point, we remove the constraint of a maximum allowable stock level, and order a quantity proportional to the predicted replacement quantity based on the expected number of canary warnings over the next order period. We then further expand the model to consider the risk (after a canary warning) with respect to the expected cost for delaying the replacement until the next order arrival point if a spare is not available, and then decide whether to replace the system immediately or wait. Finally, we consider the case where a preventive maintenance plan is in place. The comparison using numerical examples among all developed models and algorithms shows that the expected total operating cost in terms of inventory and maintenance is lowered significantly when the canary is used, and particularly when the risk analysis is implemented.