n the agricultural domain, the deployment of large language models (LLMs) is hindered by the lack of training data and evaluation benchmarks. To mitigate this issue, we propose AgriEval, the first comprehensive Chinese agricultural benchmark with three main characteristics: (1) Comprehensive Capability Evaluation. AgriEval covers six major agriculture categories and 29 subcategories within agriculture, addressing four core cognitive scenarios—memorization, understanding, inference, and generation. (2) High-Quality Data. The dataset is curated from university-level examinations and assignments, providing a natural and robust benchmark for assessing the capacity of LLMs to apply knowledge and make expert-like decisions. (3) Diverse Formats and Extensive Scale. AgriEval comprises 14,697 multiple-choice questions and 2,167 open-ended question-and-answer questions, establishing it as the most extensive agricultural benchmark available to date. We also present comprehensive experimental results over 51 open-source and commercial LLMs. The experimental results reveal that most existing LLMs struggle to achieve 60 percent accuracy, underscoring the developmental potential in agricultural LLMs. Additionally, we conduct extensive experiments to investigate factors influencing model performance and propose strategies for enhancement.

Lifelong model editing (LME) aims to sequentially rectify outdated or inaccurate knowledge in deployed LLMs while minimizing side effects on unrelated inputs.However, existing approaches typically apply parameter perturbations to a static and dense set of LLM layers for all editing instances.This practice is counter-intuitive, as we hypothesize that different pieces of knowledge are stored in distinct layers of the model.Neglecting this layer-wise specificity can impede adaptability in integrating new knowledge and result in catastrophic forgetting for both general and previously edited knowledge.To address this, we propose HiEdit, a hierarchical reinforcement learning framework that adaptively identifies the most knowledge-relevant layers for each editing instance.By enabling dynamic, instance-aware layer selection and incorporating an intrinsic reward for sparsity, HiEdit achieves precise, localized updates.Experiments on various LLMs show that HiEdit boosts the performance of the competitive RLEdit by an average of 8.48% with perturbing only half of the layers per edit.Our code is available at: https://github.com/yangfanww/hiedit.

Reasoning segmentation (RS) interprets implicit textual instructions to accurately segment target regions. This reasoning capability transforms ambiguous non-expert queries into precise pixel-level masks, thereby enabling downstream tasks like area measurement and density analysis with a level of precision unattainable by detection methods. However, existing RS models are not tailored for agriculture and lack domain-specific knowledge, which poses challenges in handling similar pest appearances and small target scales. To bridge this gap, we introduce a fine-grained pest RS task with two subtasks: Pest Discriminative Referring Expression Segmentation (PDRES) and Pest Exclusion Reasoning Segmentation (PERS). Based on this, we propose PestScope, which integrates vision, language, and reasoning for fine-grained pest segmentation. To tackle the exclusion of small non-target pests, we introduce a dedicated [NON] token alongside the standard [SEG] token for target pests. This guides the model to prioritize small target pests and suppress non-target background regions. To further address pest similarity, we propose an Exclusivity Suppression Loss, applying differentiated supervision to [SEG] and [NON] tokens to better separate target and non-target pests. Additionally, we develop an automated dataset construction pipeline to address the scarcity of fine-grained, difficulty-controllable pest RS datasets. It produces 45k and 27.6k image-text-mask samples for the PDRES and PERS tasks, respectively, covering 18 pest categories. Experiments show that in small and similar pest scenarios, integrating PestScope into mainstream models improves average gIoU by 4.28% on PDRES and 6.49% on PERS. For unseen pest categories, gIoU increases by 21.72% and 8.66%, respectively, demonstrating strong generalization. Code and datasets will be available at: https://github.com/aluodaydayup/PestScope.

Breast cancer is the leading cause of death among women worldwide. Ultrasound imaging is an important means for the early detection of breast cancer to improve the survival rate. Due to the shortage of experienced sonographers, computer-aided systems for breast cancer recognition become particularly important. Some recent studies analyze tumor types in lesion regions but rely on predefined ROIs. Some other studies recognize cancer in the whole ultrasound image, but always suffer from the extremely variable proportion, location and quantity of the tumor lesions. In this paper, we propose a Regional Prior Distillation (RPD) framework for breast cancer diagnosis in ultrasound images. To enhance the analysis of the tumor region, we propose an Image-Cross Attention (ICA) to fuse the predefined ROI prior information with ultrasound images by training a prior-fused model. To remove the constraint of predefined ROIs, we propose a Distribution Distillation Learning (DDL) to distill the prior-fused sample distribution from the prior-fused model into a diagnostic model, which analyzes the disease from only ultrasound images, based on the knowledge distillation paradigm of the teacher-student framework. Comprehensive experiments are conducted on multi-institutional datasets to validate the proposed RPD framework. The results demonstrate the following points. The ICA fuses regional prior information adequately, leading to a high-performance prior-fused model. The DDL distills the prior information effectively, enhancing the diagnostic model to focus on the tumor lesions. The performance of the diagnostic model surpasses that of current SOTA methods. In addition, the diagnostic model is robust to slight perturbations and achieves good generalization performance.

Multi-modal Large Language Models (MLLMs) integrating images, text, and speech can provide farmers with accurate diagnoses and treatment of pests and diseases, enhancing agricultural efficiency and sustainability. However, existing benchmarks lack comprehensive evaluations, particularly in multi-level reasoning, making it challenging to identify model limitations. To address this issue, we introduce Agri-CM^3, an expert-validated benchmark assessing MLLMs’ understanding and reasoning in agricultural management. It includes 3,939 images and 15,901 multi-level multiple-choice questions with detailed explanations. Evaluations of 45 MLLMs reveal significant gaps. Even GPT-4o achieves only 63.64% accuracy, falling short in fine-grained reasoning tasks. Analysis across three reasoning levels and seven compositional abilities highlights key challenges in accuracy and cognitive understanding. Our study provides insights for advancing MLLMs in agricultural management, driving their development and application. Code and data are available at https://github.com/HIT-Kwoo/Agri-CM3.

The management of Type 1 Diabetes (T1D) involves comprehensive scenarios, including blood glucose prediction, risk assessment, and insulin dosing control. However, the heterogeneity of information requirements, task objectives, and behavioral logic poses challenges in constructing unified T1D management systems with conventional deep learning models, mainly attributed to insufficient capability of feature alignment and lack of high-quality multi-scenario T1D data. In this paper, we propose T1D-MLLM, the first multimodal large language model designed for unified multi-scenario T1D management, as well as construct LCT1D, a large-scale and cross-scenario T1D dataset. Specifically, T1D-MLLM integrates time-series physiological data with natural language descriptions to capture longterm dependencies across multiple management scenarios while also enhancing fine-grained perception of time-series. Meanwhile, to overcome data scarcity, we proposed a multimodal data generation paradigm based on expert strategies. By constructing task templates and applying a rule-driven alignment mechanism, we generated 150,000 high-quality expert samples with individualized physiological parameters, which provide rich and diverse training samples, significantly improving the T1D-MLLM's capabilities in heterogeneous feature alignment and cross-scenario inference. These experiments demonstrate the effectiveness of the T1D-MLLM in multi-scenarios of various tasks as a unified system, with an excellent performance that surpasses both opensource and proprietary models.

Blood glucose (BG) forecasting aims to help people with Type-1 diabetes (T1D) avoid hyperglycemia or hypoglycemia, which plays a crucial role in medical monitoring. Despite the advancements in deep learning methods for BG forecasting, their ability to predict long-term time series remains limited, and they cannot fully meet the demand for BG forecasting. This limitation stems from the failure to account for both intra- and inter-variable variations simultaneously. To address this challenge, we introduce the$\text{Fi}^{2}$VBlock, which exploits the frequency perspective to fuse intra- and intervariable variations. After transforming to the frequency domain using the Frequency Transform Module, the Frequency Cross Attention between the real and imaginary parts is designed to obtain enhanced frequency representations and capture intravariable variations. In addition, inception blocks are employed to integrate information, thus capturing correlations across different variables. Our backbone network,$\text{Fi}^{2} \mathrm{V}$, employs a residual architecture by concatenating multiple$\text{Fi}^{2}$VBlocks, thereby avoiding degradation problems. Experimental evaluations reveal that$\text{Fi}^{2} \mathrm{V}$outperforms other baselines on the T1DMS and Dnurse datasets and demonstrates zero-shot generalization across patients.

Multi-hop question answering faces substantial challenges due to data sparsity, which increases the likelihood of language models learning spurious patterns.To address this issue, prior research has focused on diversifying question generation through content planning and varied expression.However, these approaches often emphasize generating simple questions and neglect the integration of essential knowledge, such as relevant sentences within documents.This paper introduces the Knowledge Composition Sampling (KCS), an innovative framework designed to expand the diversity of generated multi-hop questions by sampling varied knowledge compositions within a given context.KCS models the knowledge composition selection as a sentence-level conditional prediction task and utilizes a probabilistic contrastive loss to predict the next most relevant piece of knowledge.During inference, we employ a stochastic decoding strategy to effectively balance accuracy and diversity.Compared to competitive baselines, our KCS improves the overall accuracy of knowledge composition selection by 3.9%, and its application for data augmentation yields improvements on HotpotQA and 2Wiki-MultihopQA datasets.

It is a long-standing question to discover causal relations from observed variables in many empirical sciences. However, current causal discovery methods are inefficient when dealing with large-scale observed variables due to challenges in conditional independence (CI) tests or complex computations of acyclicity, and may even fail altogether. To address the efficiency issue in causal discovery from large-scale observed variables, we propose a Hierarchical Causal Discovery (HCD) framework with a bilevel policy that handles this issue by boosting existing models. Specifically, the high-level policy first finds a causal cut set to partition observed variables into several causal clusters and releases the clusters to the low-level policy. The low-level policy applies any causal discovery method to process these causal clusters in parallel and obtain intra-cluster structures for subsequently inter-cluster structure merging in the high-level policy. To avoid missing inter-cluster edges, we theoretically demonstrate the feasibility of causal cluster cut and inter-cluster structure merging. We also prove the completeness and correctness of HCD for causal discovery. Experiments on both synthetic and real-world datasets demonstrate that HCD consistently and significantly enhances the efficiency and effectiveness of existing advanced methods.

The classification of whole slide images (WSIs) is crucial in computational pathology and has significant clinical implications. Due to the extremely high resolution of WSIs and the lack of detailed lesion annotations, Multiple instance learning (MIL) has recently shown great promise for WSI classification by modeling WSIs as "bags" and treating cropped patches as "instances". However, using pre-trained feature extractors often leads to biased instance representations as the data used to pre-train these models differ significantly from histopathology data. Furthermore, since focusing on only certain instances may lead to overlooking important details, it is crucial to comprehensively assess the relevance of all instances for positive instance selection. In this paper, we propose a weakly supervised method to enhance WSI classification using adaptive representation and an instance relevance scoring strategy. To address the issue of biased data representation, we introduce an adaptive representation designed to enhance features relevant to lesion regions. This involves an adaptive block that transforms input features to better represent these critical characteristics, while simultaneously applying an attention-based probability distribution to maintain consistency between the transformed features. Additionally, we propose an instance relevance scoring strategy that assigns importance scores to each instance based on its contribution to the classification. Two publicly available datasets, CAMELYON-16 and TCGA-NSCLC, are used to validate the proposed method. The experimental results show that our proposed method outperforms existing state-of-the-art approaches in WSI classification.

Influenza seriously endangers human health and even causes a large number of deaths every year. Transformers for Influenza-like illness (ILI) forecasting have recently been proven effective. However, these end-to-end deep models are mathematically almost black-box, hindering us from inferring the specific roles and functionalities of each layer within the models, which constitutes a common key challenge in deep neural networks. At the same time, explainability helps trust and use AI systems effectively. In this paper, we propose an efficient ILI forecasting framework incorporating a patching design and variable-channel pairs, which can accommodate any white-box transformer, thereby endowing ILI forecasting with both interpretability and analytical accuracy. Through extensive experimental validation, leveraging white-box transformers such as CRATE, our White-box Time Series Transformer (WhiteTST) framework achieves the state-of-the-art accuracy on ILI datasets. We visually present the self-attention maps within WhiteTST to indicate further explainability. Our results suggest a pathway for designing white-box foundational models for ILI forecasting that concurrently exhibit high accuracy and interpretability. The code is available online in https://github.com/HITshenrj/WhiteTST.

Dialogue System for Medical Diagnosis (DSMD) based on reinforcement learning (RL) can simulate patient-doctor interactions, playing a crucial role in clinical diagnosis. However, due to the complexity of disease etiology, DSMD faces the challenges of low efficiency in diagnostic evidence search. Moreover, solely RL-based DSMS, without the constraints of professional medical knowledge, often generates irrational, meaningless, or even erroneous symptom inquiries, leading to poor interpretability of diagnostic path and high misdiagnosis rates. To address these issues, we propose anEvidence-based dialogue system with highlyInterpretableReasoning path forAutomaticDiagnosis (EIRAD) grounded in medical knowledge graph (MKG). Specifically, our automated diagnostic model captures key symptoms for suspected diseases by explicitly leveraging the topology of MKG, enhancing the interpretability and accuracy of diagnosis. To expedite the retrieval of factual evidence, we develop two mechanisms: 1) Mapping mechanism between the entity set of MKG and DSMD's diagnostic evidence and diseases. According to the patient's symptoms, EIRAD prunes irrelevant disease and symptom nodes from the MKG, which can truncate the invalid action of RL-based DSMD. 2) Reward Mechanism of integrating the effectiveness of symptom inquiry and the accuracy of disease diagnosis. The comprehensive reward system is suitable for intelligent consultation, which can effectively drive DSMD to accelerate evidence collection. Experimental results demonstrate that our model significantly outperforms competitive benchmark methods in symptom inquiry efficiency and diagnostic accuracy.

With the rise of intelligent medical assistance, the Dialogue System for Medical Diagnosis(DSMD) guided by reinforcement learning(RL) has gained much attention. However, currently available medical dialogue datasets suffer from insufficient diagnostic evidence caused by sparse symptoms, making it difficult to reproduce the evidence-based process of doctors in differential diagnosis and disease confirmation. Moreover, purely data-driven RL often involves extensive and blind trial-and-error, leading to inquiries about irrelevant symptoms to the patient’s chief complaints in limited dialogue turns, further exacerbating the issue of inadequate diagnostic evidence. To enhance the quantity and effectiveness of potential symptom collection in DSMD, we first construct a more comprehensive medical dialogue dataset CMD based on electronic medical records. The diversity of diseases and symptoms mentioned in the dialogue context of CMD surpasses that of existing public datasets. Furthermore, to enhance the efficiency of diagnostic evidence collection in DSMD, inspired by the logic of symptom inquiries in doctor-patient interactions, we combine experiential diagnostic knowledge with a specialized medical knowledge graph to constrain the inquiry of symptoms via RL, eliminating the introduction of symptoms unrelated to the patient. Experimental results demonstrate that our model significantly outperforms competitive benchmark methods in terms of diagnostic accuracy and the efficiency of symptom inquiries. Our codes and the CMD dataset are available at https://github.com/YanPioneer/EBAD.

Severity grading plays a vitally important role in the diagnosis and treatment of acne. However, due to its special characteristics such as similar samples, unclear class boundaries and imbalanced categories, the diagnosis is extremely easy to be confused. In this paper, we propose a novel, simple and intuitive loss function, namely Aggregation Separation Loss (ASLoss), as an adjunct for classification loss to clarify the common easily-confused cases. The ASLoss mines the commonalities of the same severity and the gaps among different severities in deep feature spaces. To demonstrate the generality of the proposed ASLoss, we also validate ASLoss on another common easily-confused task of expression recognition. The experimental results show that representations extracted by ASLoss are sufficiently clear and distinguishable, the performance of various popular methods can be improved significantly by ASLoss, the optimal network reaches the state-of-the-art and diagnostic level of dermatologists and the ASLoss can be generalized to improve the performance on other easily-confused tasks.

Acne seriously affects people’s daily lives. Several studies of automated acne diagnosis either lack reasonable interpretation to support the diagnosis or ignore evidence in the diagnosis process. In this paper, we propose an interpretable diagnosis framework for face acne. This framework uses complementation learning of evidence localization and severity level grading to boost both streams by sharing the features that support each other. Evidence localization learns to identify lesion areas for supporting the diagnosis stream as well as providing interpretation. Severity level grading learns to recognize the diagnosis result and also provides reference and rectification for evidence localization. Experimental results show that complementation learning improves both evidence localization and severity level grading, the lesion areas from evidence localization can support the diagnosis and provide interpretations, and the diagnosis framework reaches the state-of-the-art level and the diagnostic performance of dermatologists.

Biomedical Named Entity Recognition (BioNER) is a critical component of biomedical information extraction. NER is more challenging in the biomedical domain because of fine-grained entity types and more common nested and discontinuous entity forms. However, none of the BioNER datasets contains a large amount of all three entity forms, including flat, nested, and discontinuous. Not to mention that there is a unified BioNER model for dealing with the above three entity forms simultaneously. Methods in the public domain only focus on identifying text spans and ignore distinguishing fine-grained entity types. To address these issues, we propose a unified framework based on our own BioNER dataset CCNER, which innovatively models the BioNER task as a combination of boundary recognition and sequence generation. CCNER is a comprehensive and fine-grained BioNER dataset, where the proportion of discontinuous, nested, and flat entities in the dataset is 8.9%, 52.6%, and 38.5%, respectively. Meanwhile, it includes five fine-grained entity types. Our proposed framework includes two modules which are boundary detection and entity generation. In the boundary detection module, we propose a sample-based span representation method to determine fine-grained entity boundaries better. Finally, we conduct experiments on four datasets and achieve competitive results1.1Code is available at https://github.com/lx-hit/BioNER.

Acne Vulgaris seriously affects people’s daily life. In this paper, we propose a face acne grading framework which is a new paradigm to solve the image classification problem where the number and type of small objects are the evidence. This framework includes two components: prior knowledge extraction and prior knowledge guided network. The prior knowledge extraction uses an excellent segmentation method to predict the lesion areas as prior knowledge. The prior knowledge guided network fuses the prior knowledge and its corresponding image to grade the severity. The experiment results demonstrate that our framework achieves the state-of-the-art and diagnosis level of dermatologists.

The collection and publication of medical images on the face are quite difficult because of the invasion of privacy. Meanwhile, it takes a major expenditure of time and effort to manually label large-scale face images covered with s o many fine skin lesions. In this work, a multi-class object large-scale image inpainting model Class-Guided PG-GAN (CGPG-GAN) is proposed and its application in boosting downstream model performances is explored. This model is applied on face acne lesion inpainting where the image size is very large and missing areas are different types of lesions. The experiment results show that our method is superior to some existing methods and can improve the performance of downstream diagnosis remarkably.

The auxiliary disease diagnosis based on electronic medical records is of great significance, providing doctors with diagnostic advice and avoiding misdiagnosis. Existing work on disease diagnosis mainly utilizes deep learning models to extract sequence information in electronic medical records, ignoring the interpretability of results and the structural knowledge, especially causal knowledge. In our work, we propose a multiscale label attention network based on abductive causal graph (MSLAN-ACG) to improve model accuracy and interpretability of results. First, we construct multiple encoders in the multiscale label attention network, which can extract n-gram segment information of different lengths for each disease. Meanwhile, to enhance the interpretability of results, we visualize the weight score of different segments for disease results. Second, we propose a disease representation method by defining an abductive causal graph and then using graph convolutional network for knowledge fusion on this graph. The information propagation based on abductive causal graph is consistent with the actual abductive reasoning process from symptoms to diseases, making the model more reasonable. The effectiveness of our model is demonstrated by achieving state-of-the-art results on MIMICIII-50 and ChineseEMR datasets.

Complex systems are ubiquitous in the real world and tend to have complicated and poorly understood dynamics. For their control issues, the challenge is to guarantee accuracy, robustness, and generalization in such bloated and troubled environments. Fortunately, a complex system can be divided into multiple modular structures that human cognition appears to exploit. Inspired by this cognition, a novel control method, Causal Coupled Mechanisms (CCMs), is proposed that explores the cooperation in division and competition in combination. Our method employs the theory of hierarchical reinforcement learning (HRL), in which 1) the high-level policy with competitive awareness divides the whole complex system into multiple functional mechanisms, and 2) the low-level policy finishes the control task of each mechanism. Specifically for cooperation, a cascade control module helps the series operation of CCMs, and a forward coupled reasoning module is used to recover the coupling information lost in the division process. On both synthetic systems and a real-world biological regulatory system, the CCM method achieves robust and state-of-the-art control results even with unpredictable random noise. Moreover, generalization results show that reusing prepared specialized CCMs helps to perform well in environments with different confounders and dynamics.

Severity level grading is a vitally important step to make correct diagnoses and personalized treatment schemes for acne, which is mainly carried out in two ways: criterion-based lesion counting and experience-based global estimation. In this paper, the global estimation of acne severity grading is studied by Convolutional Neural Networks (CNNs) and a unified acne grading framework that can diagnose referring to different grading criteria is proposed. Firstly, an adaptive image preprocessing method that can efficiently reduce the background noise and emphasize the skin information is proposed. Next, an innovative CNN structure SFNet, which fuses local skin features with global features to effectively enhance the perception of color gaps between skin and lesion, is presented. The proposed framework is verified on two datasets with different acne grading criteria. Experimental results show that the accuracy of the proposed framework reaches 84.52% exceeding the state-of-the-art method by 1.7% and reaches the diagnostic level of a professional dermatologist.

Retweet has become one of the most prominent feature on social networks and an important mean for secondary content promotion. Most existing investigations of retweet behaviors on social networks are conducted based on empirical studies or information diffusion models (such as stochastic process or cascading model). To the best of our knowledge, such a retweet process has not been investigated as a chaotic process. In this paper, we have first examined that retweet time series by 0-1 test where the results provide identification of chaotic behaviors. Furthermore, taking into account of the proven chaotic characteristic, chaos LS-SVM prediction method is applied to form predictions using only a small fraction of the retweet time series. Our evaluation on Sina Weibo dataset and comparisons with a bayesian model and strawman modal show that this nonlinear prediction method can translate to good step ahead forecasts and perform high accuracy in retweet prediction.

Although cascading failures have occurred on many real-world networks, to our best knowledge, no one has clearly identified this issue on a social network. In this paper, we identify this potential issue on social networks, and develop a theoretical model to analyze related issues. Note that highly-influential “super” users play critical roles on a social network. When they are suddenly unavailable, a large portion of the social network may be seriously disrupted. The proposed model captures this dynamic process and helps us better understand related issues. Furthermore, we evaluate the proposed model under four attack strategies based on real social network datasets collected on Twitter and Sina Weibo. We also analyze the connectivity, the persistent time, and the cascade effect of a social network under these attacks. Our results show that social network service providers have to pay closer attention to super users to avoid dramatic failures.