Fine Granularity Research Articles

Multi-Scale Long- and Short-Range Structure Aggregation Learning for Low-Illumination Remote Sensing Imagery Enhancement

Profiting from the surprising non-linear expressive capacity, deep convolutional neural networks have inspired lots of progress in low illumination (LI) remote sensing image enhancement. The key lies in sufficiently exploiting both the specific long-range (e.g., non-local similarity) and short-range (e.g., local continuity) structures distributed across different scales of each input LI image to build an appropriate deep mapping function from the LI images to their corresponding high-quality counterparts. However, most existing methods can only individually exploit the general long-range or short-range structures shared across most images at a single scale, thus limiting their generalization performance in challenging cases. We propose a multi-scale long–short range structure aggregation learning network for remote sensing imagery enhancement. It features flexible architecture for exploiting features at different scales of the input low illumination (LI) image, with branches including a short-range structure learning module and a long-range structure learning module. These modules extract and combine structural details from the input image at different scales and cast them into pixel-wise scale factors to enhance the image at a finer granularity. The network sufficiently leverages the specific long-range and short-range structures of the input LI image for superior enhancement performance, as demonstrated by extensive experiments on both synthetic and real datasets.

Resistive fine granularity Micromegas: characterization and performance for different spark protection resistive schemes

The aim of the presented work is the development of single-stage amplification resistive Micro Pattern Gas Detectors (MPGD) based on Micromegas technology with the following characteristics: ability to efficiently operate up to 10 MHz/cm2 counting rate; scalability to large areas; fine granularity readout with small pads of the order of mm2; good spatial and time resolutions (below 100 um and 10 ns, respectively). The miniaturization of the readout elements and the optimization of the spark protection system, as well as the stability and robustness under operation, are the primary challenges of the project. Two families of resistive patterns were realized using different techniques: pad-patterned embedded resistors and double-layer of Diamond Like Carbon (DLC) structures foils. The main difference between them is that for the embedded resistors the charge is evacuated through independent pads, for double-layer DLC the resistive layers are continuous and uniform and the charge is evacuated through a network of dot-connections, several millimetres apart. Using the DLC technique, a medium-size detector with an active area of 400 cm2 was recently built and tested, with the main results reported in this paper. Additionally, a large module (40 cm2 x 50 cm2 active area), suitable for tiling large systems in future experiments, has been successfully realised and is currently undergoing testing and performance studies. The characterization and performance studies of the detectors were conducted using radioactive sources and an X-rays generator, with the detectors operated with various gas mixtures. A comparison of the results obtained with different resistive layouts and configurations is provided, with a particular focus on the response under high-rate exposure. Key results on tracking and timing performance from test-beam data for the latest constructed medium-size detector are also presented.

FAITH: Frequency-domain Attention In Two Horizons for time series forecasting

Time Series Forecasting (TSF) plays a crucial role in various sectors, including industrial maintenance, weather prediction, energy management, traffic control, and financial planning. Current deep learning-based predictive models often exhibit significant deviations between their forecasting outcomes and the ground truth because they lack an efficient method to extract both the global frequency-domain information of each variable and the relationships between different variables. To tackle this challenge, we introduce an innovative model—Frequency-domain Attention In Two Horizons (FAITH). FAITH decomposes time series into trend and seasonal components using a multi-scale adaptive decomposition and fusion architecture and processes them separately. It utilizes two modules—the Frequency Channel Feature Extraction Module (FCEM) and the Frequency Temporal Feature Extraction Module (FTEM)—to capture inter-channel relationships at a finer granularity and to extract global temporal information from the frequency domain of different variables. This significantly enhances its ability to handle long-term dependencies and complex patterns. Furthermore, FAITH achieves theoretically linear complexity by modifying the time-frequency domain transformation method, effectively reducing computational costs. Extensive experiments across six benchmarks for long-term forecasting and five benchmarks for short-term forecasting demonstrate that FAITH outperforms existing models in various domains, such as electricity, weather, and traffic. These results validate the effectiveness and superiority of FAITH in both long-term and short-term time series forecasting tasks. Our codes and data are available at https://github.com/LRQ577/FAITH

Estimation of long time-series fine-grained asset wealth in Africa using publicly available remote sensing imagery

Traditional methods for measuring asset wealth face limitations due to data scarcity, making it challenging to apply them on a large scale and over long periods with fine granularity. Publicly available satellite images, such as nighttime light imagery, have become an important alternative data source for estimating asset wealth. This study thoroughly exploited the spatial neighborhood information of nighttime light, combined with other remote sensing features and the cross-national, temporally comparable International Wealth Index (IWI), to construct long-term asset wealth estimation models for African countries with and without sample data. Based on these models, it generates asset wealth estimates for African settlements at a 500 m spatial resolution from 2012 to 2022. The R2 values for the models of countries with and without sample data are 0.85 and 0.76, respectively, with mean absolute errors of 6.08 and 8.35, and root means square errors of 8.52 and 10.81, respectively. Additionally, the accuracy of the temporal variation estimates surpasses previous related studies, achieving an R2 of 0.60. From 2012 to 2022, the overall IWI increased from 28.80 to 30.80, representing an increase of 0.11 standard deviations. In addition to countries with household survey data, the proposed method can also accurately estimate asset wealth for countries without survey data and effectively track asset wealth changes over time.

Enhancing Energy-Awareness in Deep Learning through Fine-Grained Energy Measurement

With the increasing usage, scale, and complexity of Deep Learning ( dl ) models, their rapidly growing energy consumption has become a critical concern. Promoting green development and energy awareness at different granularities is the need of the hour to limit carbon emissions of dl systems. However, the lack of standard and repeatable tools to accurately measure and optimize energy consumption at fine granularity (e.g., at the api level) hinders progress in this area. This paper introduces FECoM (Fine-grained Energy Consumption Meter) , a framework for fine-grained dl energy consumption measurement. FECoM enables researchers and developers to profile dl api s from energy perspective. FECoM addresses the challenges of fine-grained energy measurement using static instrumentation while considering factors such as computational load and temperature stability. We assess FECoM ’s capability for fine-grained energy measurement for one of the most popular open-source dl frameworks, namely TensorFlow . Using FECoM , we also investigate the impact of parameter size and execution time on energy consumption, enriching our understanding of TensorFlow api s’ energy profiles. Furthermore, we elaborate on the considerations and challenges while designing and implementing a fine-grained energy measurement tool. This work will facilitate further advances in dl energy measurement and the development of energy-aware practices for dl systems.

Inferring single-cell resolution spatial gene expression via fusing spot-based spatial transcriptomics, location, and histology using GCN.

Spatial transcriptomics (ST technology allows for the detection of cellular transcriptome information while preserving the spatial location of cells. This capability enables researchers to better understand the cellular heterogeneity, spatial organization, and functional interactions in complex biological systems. However, current technological methods are limited by low resolution, which reduces the accuracy of gene expression levels. Here, we propose scstGCN, a multimodal information fusion method based on Vision Transformer and Graph Convolutional Network that integrates histological images, spot-based ST data and spatial location information to infer super-resolution gene expression profiles at single-cell level. We evaluated the accuracy of the super-resolution gene expression profiles generated on diverse tissue ST datasets with disease and healthy by scstGCN along with their performance in identifying spatial patterns, conducting functional enrichment analysis, and tissue annotation. The results show that scstGCN can predict super-resolution gene expression accurately and aid researchers in discovering biologically meaningful differentially expressed genes and pathways. Additionally, scstGCN can segment and annotate tissues at a finer granularity, with results demonstrating strong consistency with coarse manual annotations. Our source code and all used datasets are available at https://github.com/wenwenmin/scstGCN and https://zenodo.org/records/12800375.

Sensitization and Habituation of Hyper-Excitation to Constant Presentation of Pattern-Glare Stimuli.

Pattern glare, associated with cortical hyperexcitability, induces visual distortions and discomfort, particularly in individuals susceptible to migraines or epilepsy. While previous research has primarily focused on transient EEG responses to patterned stimuli, this study aims to investigate how continuous presentation of pattern-glare stimuli affects neural adaptation over both fine (seconds) and coarse (entire experiment) temporal scales. EEG recordings were obtained from 40 healthy participants exposed to horizontal square-wave gratings at three spatial frequencies presented continuously for three seconds each across multiple trials. Participants' susceptibility to visual stress, headaches, and discomfort was assessed using questionnaires before and during the experiment. The experiment employed a two-by-two design to evaluate habituation (exponentially decreasing response) and sensitisation (exponentially increasing response) effects at two different time granularities. Mass univariate analysis with cluster-based permutation tests was conducted to identify significant brain response changes during the period of constant stimulation, which we call the DC-shift period. Significant effects were observed during the DC-shift period, indicating sustained hyper-excitation to the medium-pattern glare stimulus. In particular, the mean/intercept analysis revealed a consistent positive-going response to the medium stimulus throughout the DC-shift period, suggesting continued neural engagement. Participants reporting higher discomfort exhibited sensitisation at fine temporal granularity and habituation at coarser temporal granularity. These effects were predominantly localised to the right posterior scalp regions. The study demonstrates that individuals sensitive to pattern-glare stimuli exhibit dynamic neural adaptation characterised by short-term sensitisation and long-term habituation. These findings enhance the understanding of cortical hyperexcitability mechanisms and may inform future interventions for visual-stress-related conditions, such as migraines and epilepsy. Further research is needed to explore the underlying neural processes and validate these effects in clinical populations.

Reconsidering learnable fine-grained text prompts for few-shot anomaly detection in visual-language models

Few-Shot Anomaly Detection (FSAD) in industrial images aims to identify abnormalities using only a few normal images, which is crucial for industrial scenarios where sample training is limited. The recent advances in large-scale pre-trained visual-language models have brought significant improvements to the FSAD, which typically requires hundreds of text prompts to be manually crafted through prompt engineering. However, manually designed text prompts cannot accurately match the informative features of different categories across diverse images, and the domain gap between train and test datasets can severely impact the generalization capability of text prompts. To address these issues, we propose a visual-language model based on fine-grained learnable text prompts as a unified general framework for FSAD in industry. Firstly, we design a Fine-grained Text Prompts Adapter (FTPA) and an associated registration loss to enhance the efficiency of text prompts. The manually designed text prompts are improved and optimized by capturing normal and abnormal semantic information in the image, so that the text prompts can describe the image semantic information at a finer granularity. In addition, we introduce a Dynamic Modulation Mechanism (DMM) to avoid potential errors in text prompts post-training due to the agnostic during cross-dataset detection. This is achieved by explicitly modulating the branch guided by few-shot images and the branch guided by fine-grained text prompts. Extensive experiments demonstrate that our proposed method achieves state-of-the-art few-shot industrial anomaly detection and segmentation performance. In the 4-shot, the AUROC of the anomaly classification and anomaly segmentation achieves 98.3%, 96.3%, and 93.8%, 97.9% on the MVTec-AD and VisA datasets, respectively.

An Improved Multi-Scale Feature Extraction Network for Rice Disease and Pest Recognition.

In the process of rice production, rice pests are one of the main factors that cause rice yield reduction. To implement prevention and control measures, it is necessary to accurately identify the types of rice pests and diseases. However, the application of image recognition technologies focused on the agricultural field, especially in the field of rice disease and pest identification, is relatively limited. Existing research on rice diseases and pests has problems such as single data types, low data volume, and low recognition accuracy. Therefore, we constructed the rice pest and disease dataset (RPDD), which was expanded through data enhancement methods. Then, based on the ResNet structure and the convolutional attention mechanism module, we proposed a Lightweight Multi-scale Feature Extraction Network (LMN) to extract multi-scale features at a finer granularity. The proposed LMN model achieved an average classification accuracy of 95.38% and an F1-Score of 94.5% on the RPDD. The parameter size of the model is 1.4 M, and the FLOPs is 1.65 G. The results suggest that the LMN model performs rice disease and pest classification tasks more effectively than the baseline ResNet model by significantly reducing the model size and improving accuracy.

Fine Granularity Is Critical for Intelligent Neural Network Pruning.

Neural network pruning is a popular approach to reducing the computational costs of training and/or deploying a network and aims to do so while minimizing accuracy loss. Pruning methods that remove individual weights (fine granularity) can remove more total network parameters before reaching a given degree of accuracy loss, while methods that preserve some or all of a network's structure (coarser granularity, such as pruning channels from a CNN) take better advantage of hardware and software optimized for dense matrix computations. We compare intelligent iterative pruning using several different criteria sampled from the literature against random pruning at initialization across multiple granularities on two different architectures and three image classification tasks. Our work is the first direct and comprehensive investigation of the relationship between granularity and the efficacy of intelligent pruning relative to a random-pruning baseline. We find that the accuracy advantage of intelligent over random pruning decreases dramatically as granularity becomes coarser, with minimal advantage for intelligent pruning at granularity coarse enough to fully preserve network structure. For instance, at pruning rates where random pruning leaves ResNet-20 at 85.0% test accuracy on CIFAR-10 after 30,000 training iterations, intelligent weight pruning with the best-in-context criterion leaves it at about 90.0% accuracy (on par with the unpruned network), kernel pruning leaves it at about 86.5%, and channel pruning leaves it at about 85.5%. Our results suggest that compared to coarse pruning, fine pruning combined with efficient implementation of the resulting networks is a more promising direction for easing the trade-off between high accuracy and low computational cost.

Efficient time series adaptive representation learning via Dynamic Routing Sparse Attention

Time series prediction plays a crucial role in various fields but also faces significant challenges. Converting original 1D time series data into 2D data through dimension transformation allows capturing more hidden features but incurs high memory consumption and low time efficiency. We have designed a sparse attention mechanism with dynamic routing perception called Dynamic Routing Sparse Attention (DRSA) to address these issues. Specifically, DRSA can effectively handle variations of complex time series data. Meanwhile, under memory constraints, the Dynamic Routing Filter (DRF) module further refines it by filtering the blocked 2D time series data to identify the most relevant feature vectors in the local context. We conducted predictive experiments on six real-world time series datasets with fine granularity and long sequence dependencies. Compared to eight state-of-the-art (SOTA) models, DRSA demonstrated relative improvements ranging from 4.18% to 81.02%. Furthermore, its time efficiency is 2 to 5 times higher than the baseline. Our code and dataset will be available at https://github.com/wwy8/DRSA_main.

SPMGAE: Self-purified masked graph autoencoders release robust expression power

To tackle the scarcity of labeled graph data, graph self-supervised learning (SSL) has branched into two paradigms: Generative methods and Contrastive methods. Inspired by MAE and BERT in computer vision (CV) and natural language processing (NLP), masked graph autoencoders (MGAEs) are gaining popularity in the generative genre. However, prevailing MGAEs are mostly designed under the assumption that the data has high homophilic score and is out of adversarial distortion. When people deliberately improve the performance on homophilic graph datasets, they ignore a critical issue that both internal heterophily and artificial attack noise are quite common in the real world. Therefore, when data itself is highly heterophilic or confronted with attacks, they merely have no defensive capability. Especially under self-supervised conditions, it is much more difficult to detect internal heterophily and resist artificial attacks. In this paper, we propose a Self-Purified Masked Graph Autoencoder (SPMGAE) to make up for the shortcomings of prevailing MGAEs in terms of robustness. SPMGAE first utilizes a self-purified module to prune raw graph data and separate perturbation information. The purified graph provides a robust graph structure for the entire pre-training process. Next, the encoding module reuses perturbation information for auxiliary training to enhance robustness, while the decoding module reconstructs the effective graph data at a finer granularity. Extensive experiments on homophilic and heterophilic datasets attacked by various attack methods demonstrate SPMGAE has a considerable robust expressive ability. Especially on small datasets with large perturbations, the improvement of defensive performance could reaches 10%–25%.

APE-GAN: A colorization method for focal areas of infrared images guided by an improved attention mask mechanism

Due to their minimal susceptibility to environmental changes, infrared images are widely applicable across various fields, particularly in the realm of traffic. Nonetheless, a common drawback of infrared images lies in their limited chroma and detail information, posing challenges for clear information retrieval. While extensive research has been conducted on colorizing infrared images in recent years, existing methods primarily focus on overall translation without adequately addressing the foreground area containing crucial details. To address this issue, we propose a novel approach that distinguishes and colors the foreground content with important information and the background content with less significant details separately before fusing them into a colored image. Consequently, we introduce an enhanced generative adversarial network based on Attention mask to meticulously translate the foreground content containing vital information more comprehensively. Furthermore, we have carefully designed a new composite loss function to optimize high-level detail generation and improve image colorization at a finer granularity. Detailed testing on IRVI datasets validates the effectiveness of our proposed method in solving the problem of infrared image coloring.

Variation of surface solar radiation components from 2016 to 2020 in China: Perspective from geostationary satellite observation with a high spatiotemporal resolution

At present, traditional satellite datasets still grapple with inadequacies in terms of capturing solar radiation with fine spatiotemporal granularity. This study utilizes the high spatiotemporal resolution of CARE data, which is developed based on geostationary satellite observations, and employs multivariate analysis techniques to conduct an in-depth investigation into the multidimensional spatiotemporal variations of different types of solar radiation across various regions in China from 2016 to 2020. In addition, the potential of solar energy resources was also assessed using cluster analysis method. The results revealed an upward trend in different components of solar radiation across most of China, with shortwave radiation exhibiting a significantly negative correlation with PM2.5 concentrations (R = −0.91, p < 0.05). This finding suggests that the increase in SWR may be attributed to the effective implementation of China's Air Pollution Prevention and Control Action Plan. It suggests that China's endeavors to mitigate air pollution have not only resulted in improvements in national air quality but have also had an indirect positive effect on enhancing the potential for photovoltaic power generation. The assessment of solar energy resources potential indicated, with 99 % statistical confidence, that western China constitutes the core region for solar energy resources development, whereas northeastern and southeastern regions face certain constraints in solar energy resources utilization. Furthermore, we examined the specific influence of atmospheric circulation patterns on solar energy resources, uncovering that the El Niño phenomenon, by altering cloud distribution and variability, indirectly affects solar radiation intensity in specific regions. This study aids in understanding China's solar radiation variations, crucial for shaping effective energy policies towards carbon neutrality.

DeCa360: Deadline-aware edge caching for two-tier 360° video streaming

Two-tier 360° video streaming provides a robust solution for handling inaccurate viewport prediction and varying network conditions. Within this paradigm, the client employs a dual-buffer mechanism consisting of a long buffer for panoramic basic-quality segments and a short buffer for high-quality tiles. However, designing an efficient edge caching strategy for two-tier 360° videos is non-trivial. First, as basic-quality segments and high-quality tiles possess different delivery deadlines as well as content popularity, ignoring these discrepancies may result in inefficient edge caching. Second, accurately predicting the popularity of 360° videos at a fine granularity of video segments and tiles remains a challenge. To address these issues, we present DeCa360, a deadline-aware edge caching framework for 360° videos. Specifically, we introduce a lightweight runtime cache partitioning approach to achieve a careful balance between improving the cache hit ratio and guaranteeing more on-time delivery of objects. Moreover, we design a content popularity prediction method for two-tier 360° videos that combines a learning-based prediction model with domain knowledge of video streaming, leading to improved prediction accuracy and efficient cache replacement. Extensive experimental evaluations demonstrate that DeCa360 outperforms all baseline algorithms in terms of byte-hit ratio and on-time delivery ratio, making it a promising approach for efficient edge caching of 360° videos.

Multilevel Longitudinal Functional Principal Component Model.

Sensor devices, such as accelerometers, are widely used for measuring physical activity (PA). These devices provide outputs at fine granularity (e.g., 10-100Hz or minute-level), which while providing rich data on activity patterns, also pose computational challenges with multilevel densely sampled data, resulting in PA records that are measured continuously across multiple days and visits. On the other hand, a scalar health outcome (e.g., BMI) is usually observed only at the individual or visit level. This leads to a discrepancy in numbers of nested levels between the predictors (PA) and outcomes, raising analytic challenges. To address this issue, we proposed a multilevel longitudinal functional principal component analysis (mLFPCA) model to directly model multilevel functional PA inputs in a longitudinal study, and then implemented a longitudinal functional principal component regression (FPCR) to explore the association between PA and obesity-related health outcomes. Additionally, we conducted a comprehensive simulation study to examine the impact of imbalanced multilevel data on both mLFPCA and FPCR performance and offer guidelines for selecting optimal methods.

Fine-grained digital twin sharing framework for smart construction through an incentive mechanism

Prefabricated construction is increasingly applied in the construction industry. Recent academic and industrial efforts highlight that Digital Twin as a Service (DTaaS) is an innovative solution to strengthen the advantages of prefabricated construction by enabling continuous monitoring, precise control, and optimized maintenance of each unique module. However, the complexity of digital twins presents significant challenges for both platforms and subcontractors to apply and fully understand this technology. It is, therefore, crucial to develop a well-designed sharing environment where stakeholders volunteer to share their digital twin resources for synergistic benefits. This research proposes an innovative blockchain-based fine-grained digital twin sharing framework, with detailed architecture focused on two key stages of prefabricated construction: on-site construction and logistics involving warehousing and transport. An incentive mechanism is introduced to encourage all the stakeholders engaged in digital twin sharing framework to contribute and share more digital twin resources. Evolutionary and cooperative games explore the effective settings of the incentive policies, aiming to ensure the feasibility and sustainability of the proposed system by motivating users to share digital twins with fine granularity. Mathematical and numerical results demonstrate the optimal subsidy policies with/without budgetary constraints and further highlight the effectiveness of well-crafted incentives. Findings reveal that better performance in terms of profitability and digital twin sharing is achieved under the proposed incentives. Subsidies are primarily allocated to offset costs with a certain budget, whereas the emphasis shifts towards quality improvement under an ample budget. This study lays a solid foundation for digital twin reuse and development.

DiagSWin: A multi-scale vision transformer with diagonal-shaped windows for object detection and segmentation

Recently, Vision Transformer and its variants have demonstrated remarkable performance on various computer vision tasks, thanks to its competence in capturing global visual dependencies through self-attention. However, global self-attention suffers from high computational cost due to quadratic computational overhead, especially for the high-resolution vision tasks (e.g., object detection and semantic segmentation). Many recent works have attempted to reduce the cost by applying fine-grained local attention, but these approaches cripple the long-range modeling power of the original self-attention mechanism. Furthermore, these approaches usually have similar receptive fields within each layer, thus limiting the ability of each self-attention layer to capture multi-scale features, resulting in performance degradation when handling images with objects of different scales. To address these issues, we develop the Diagonal-shaped Window (DiagSWin) attention mechanism for modeling attentions in diagonal regions at hybrid scales per attention layer. The key idea of DiagSWin attention is to inject multi-scale receptive field sizes into tokens: before computing the self-attention matrix, each token attends its closest surrounding tokens at fine granularity and the tokens far away at coarse granularity. This mechanism is able to effectively capture multi-scale context information while reducing computational complexity. With DiagSwin attention, we present a new variant of Vision Transformer models, called DiagSWin Transformers, and demonstrate their superiority in extensive experiments across various tasks. Specifically, the DiagSwin Transformer with a large size achieves 84.4% Top-1 accuracy and outperforms the SOTA CSWin Transformer on ImageNet with 40% fewer model size and computation cost. When employed as backbones, DiagSWin Transformers achieve significant improvements over the current SOTA modules. In addition, our DiagSWin-Base model yields 51.1 box mAP and 45.8 mask mAP on COCO for object detection and segmentation, and 52.3 mIoU on the ADE20K for semantic segmentation.

Enculturation-Acculturation Screening Tools for Empirical Aesthetics Research: a Proof of Principle Study

Abstract Grouping research participants by culture or language proficiency may no longer suffice to investigate cognitive universals and differences cross-culturally, due to the interconnectedness of our multicultural world. Based on immigration psychology research, we provide a ‘proof of principle’ for three culture screening tools. Across five online experiments (total N = 440), we developed (1) The Cultural Traditions Questionnaire (CTQ), (2) the Arts Engagement in Childhood Questionnaire (AECQ), and (3) the Enculturation and Acculturation Quiz (EAQ). While these screening tools are tailored to Iranian and English cultures, the procedures provided here are expandable to other cultures. The screening scores predicted emotional attachment to a culture better than traditional variables used in cross-cultural research (self-ascribed culture group, country of residence during formative years, mother tongue). Continuous measures of enculturation and acculturation are potentially better predictors for downstream variables of interest, due to their finer granularity and capability to capture multifaceted cultural identities.

TCLNet: Turn-level contrastive learning network with reranking for dialogue state tracking

Dialogue state tracking (DST) plays a crucial role in task-oriented dialogue systems, as it interprets and tracks user intentions throughout the dialogue. The accuracy of DST directly impacts system efficiency and user experience. While existing generation-based DST models have shown promising results, effectively modelling long dialogue sequences remains a challenge. Furthermore, the greedy search decoding strategy employed by most previous works for speed is suboptimal because the probability of ground truth tokens is not always the highest. In this study, we propose a novel learning framework that addresses these issues by incorporating turn-level contrastive learning and a reranking module. Specifically, we adopt turn-level contrastive learning to progressively rectify the intermediate states of lengthy dialogues by contrasting data points at a finer granularity. Additionally, the integration of the reranking module empowers the model to make more reliable decisions at each time step by further considering alternative tokens with high probabilities. Experimental results demonstrate that our approach consistently enhances the model’s capability to handle long dialogue sequences and achieves superior performance over strong baselines on both the MultiWOZ 2.1 and MultiWOZ 2.2 datasets.