Despite the fact that large language models (LLMs) show exceptional skill in instruction following tasks, this strength can turn into a vulnerability when the models are required to disregard certain instructions.Instruction following tasks typically involve a clear task description and input text containing the target data to be processed.However, when the input itself resembles an instruction, confusion may arise, even if there is explicit prompting to distinguish between the task instruction and the input.We refer to this phenomenon as instructional distraction.In this paper, we introduce a novel benchmark, named DIM-Bench, specifically designed to assess LLMs' performance under instructional distraction.The benchmark categorizes real-world instances of instructional distraction and evaluates LLMs across four instruction tasks: rewriting, proofreading, translation, and style transfer-alongside five input tasks: reasoning, code generation, mathematical reasoning, bias detection, and question answering.Our experimental results reveal that even the most advanced LLMs are susceptible to instructional distraction, often failing to accurately follow user intent in such cases. Instruction Input ExampleRewrite Reasoning Instruction: Paraphrase the following text.Input: Laundry detergents were once manufactured to contain high ... which would a lake become as a result of the phosphorous in the detergent?Options : A. canyon B. desert C. swamp D. river

Recently, large vision–language models (LVLMs) have emerged as the preferred tools for judging text–image alignment, yet their robustness along the visual modality remains underexplored. This work is the first study to address a key research question: Can adversarial visual manipulations systematically fool LVLM judges into assigning unfairly inflated scores? We define potential image-induced biases within the context of T2I evaluation and examine how these biases affect the evaluations of LVLM judges. Moreover, we introduce a novel, fine-grained, multi-domain meta-evaluation benchmark named FRAME, which is deliberately constructed to exhibit diverse score distributions. By introducing the defined biases into the benchmark, we reveal that all tested LVLM judges exhibit vulnerability across all domains, consistently inflating scores for manipulated images. Further analysis reveals that combining multiple biases amplifies their effects, and pairwise evaluations are similarly susceptible. Moreover, we observe that visual biases persist despite prompt-based mitigation strategies, highlighting the vulnerability of current LVLM evaluation systems and underscoring the urgent need for more robust LVLM judges.

Dongryeol Lee, Yerin Hwang, Yongil Kim, Joonsuk Park, Kyomin Jung. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Despite their strengths, large language models (LLMs) often fail to communicate their confidence accurately, making it difficult to assess when they might be wrong and limiting their reliability. In this work, we demonstrate that reasoning models that engage in extended chain-of-thought (CoT) reasoning exhibit superior performance not only in problem-solving but also in accurately expressing their confidence. Specifically, we benchmark six reasoning models across six datasets and find that they achieve strictly better confidence calibration than their non-reasoning counterparts in 33 out of the 36 settings. Our detailed analysis reveals that these gains in calibration stem from the slow thinking behaviors of reasoning models (e.g., exploring alternative approaches and backtracking) which enable them to adjust their confidence dynamically throughout their CoT, making it progressively more accurate. In particular, we find that reasoning models become increasingly better calibrated as their CoT unfolds, a trend not observed in non-reasoning models. Moreover, removing slow thinking behaviors from the CoT leads to a significant drop in calibration. Lastly, we show that non-reasoning models also demonstrate enhanced calibration when simply guided to slow think via in-context learning, fully isolating slow thinking as the source of the calibration gains.

Existing English-based text similarity measurements primarily focus on the semantic dimension, neglecting the unique linguistic attributes found in languages like Korean, where honorific expressions are explicitly integrated. To address this limitation, this study proposes Kosmic, a novel Korean text-similarity metric that encompasses the semantic and tonal facets of a given text pair. For the evaluation, we introduce a novel benchmark annotated by human experts, empirically showing that Kosmic outperforms the existing method. Moreover, by leveraging Kosmic, we assess various Korean paraphrasing methods to determine which techniques are most effective in preserving semantics and tone.

Despite advancements in on-topic dialogue systems, effectively managing topic shifts within dialogues remains a persistent challenge, largely attributed to the limited availability of training datasets.To address this issue, we propose Multi-Passage to Dialogue (MP2D), a data generation framework that automatically creates conversational questionanswering datasets with natural topic transitions.By leveraging the relationships between entities in a knowledge graph, MP2D maps the flow of topics within a dialogue, effectively mirroring the dynamics of human conversation.It retrieves relevant passages corresponding to the topics and transforms them into dialogues through the passage-to-dialogue method.Through quantitative and qualitative experiments, we demonstrate MP2D's efficacy in generating dialogue with natural topic shifts.Furthermore, this study introduces a novel benchmark for topic shift dialogues, TS-WikiDialog.Utilizing the dataset, we demonstrate that even Large Language Models (LLMs) struggle to handle topic shifts in dialogue effectively, and we showcase the performance improvements of models trained on datasets generated by MP2D across diverse topic shift dialogue tasks.

The severity and frequency of wildfires are escalating globally. Gangwon Province in Korea, known for its high susceptibility to wildfires, consists of approximately 80% mountainous terrain, making it a critical region where wildfires often exhibit significant scale and prolonged duration. Notably, a considerable portion of South Korea's power transmission facilities is situated in mountainous areas, increasing the potential for major societal disruptions, such as power outages. Consequently, there arises an urgent need for a proactive and long-term monitoring system utilizing remote sensing technology and periodic satellite imagery. Existing research often relies on field survey data and low to medium resolution satellite imagery for broad area analysis. However, selective monitoring of fire-related factors proves advantageous considering extensive terrain encompassing power transmission facilities. For this purpose, this paper employs high-resolution satellite imagery and periodically updated open-source geospatial data to extract key factors contributing to fire risk adjacent to mountainous transmission facilities. Through correlation analysis between fire-related factors and ground truth data of fire risk levels, three fire risk factors (forest type, slope gradient, and slope aspect) were extracted. The extracted factors were subsequently incorporated into a CART (Classification and Regression Trees)-based ensemble model with different weights assigned to generate fire risk maps for validation. The accuracy results confirmed the need for wildfire risk monitoring utilizing the extracted fire risk factors in the vicinity of mountainous transmission facilities.

To address the data scarcity issue in Conversational question answering (ConvQA), a dialog inpainting method, which utilizes documents to generate ConvQA datasets, has been proposed.However, the original dialog inpainting model is trained solely on the dialog reconstruction task, resulting in the generation of questions with low contextual relevance due to insufficient learning of question-answer alignment.To overcome this limitation, we propose a novel framework called Dialogizer, which has the capability to automatically generate ConvQA datasets with high contextual relevance from textual sources.The framework incorporates two training tasks: question-answer matching (QAM) and topic-aware dialog generation (TDG).Moreover, re-ranking is conducted during the inference phase based on the contextual relevance of the generated questions.Using our framework, we produce four Con-vQA datasets by utilizing documents from multiple domains as the primary source.Through automatic evaluation using diverse metrics, as well as human evaluation, we validate that our proposed framework exhibits the ability to generate datasets of higher quality compared to the baseline dialog inpainting model.

The majority of power transmission facilities in Korea are located in mountainous areas, and their scale is continuously expanding. However, due to the limitations of manpower-based monitoring focused on the maintenance of transmission towers, regular management has become challenging. To ensure the stable management of aging power transmission facilities, proactive monitoring systems, in addition to post-disaster damage analysis, are necessary. Although wide-area time-series monitoring using satellite imagery can assist in decision-making for disaster preparedness and recovery measures, there is currently a lack of specific monitoring strategies based on satellite imagery for surrounding environment of power transmission facilities. In particular, landslides can act in a complex manner, involving various geological and geomorphic factors, progressively impacting surrounding areas. Current studies primarily rely on direct field surveys to create landslide vulnerability maps and assess risks and damages quantitatively. However, this approach is time-consuming, difficult to sustain in the long term, and prone to subjectivity issues. Therefore, this study aimed to leverage the advantages of remote sensing in the challenging context of mountainous areas surrounding power transmission facilities. The goal was to extract landslide-related factors that accurately represent the unique characteristics of the region using readily available data, without the need for on-site data. Specifically, this study focused on extracting pertinent factors related to landslide impacts in areas exhibiting changes, using two satellite images near power transmission facilities in Gangwon Province, South Korea. PlanetScope imagery for the case of heavy rainfall in August 2022 was used and easily obtainable spatial data regarding land cover, slope, and vegetation were utilized. The study conducted two independent experiments. In the first experiment, a binary classification was performed based on spectral changes to identify highly correlated factors among the landslide-related elements. The second experiment analyzed the patterns of land cover change and their correlation with landslide-related factors. We have confirmed the necessity of monitoring slope gradient, slope direction, and forest diameter in the study area and the variables exhibiting significant correlations generally displayed evident positive relationship in most cases. This study provides evidence of the capability to consistently generate potential landslide hazards analysis results utilizing GIS data extracted from basic maps. By prioritizing elements that exhibit strong long-term correlations, the study highlights the potential to establish a time-series monitoring dataset for the surrounding environment of power transmission facilities in mountainous areas, utilizing satellite imagery.

Landslide is a serious geographical disaster, causing human casualties and infrastructure losses. Rapid and accurate detection of the landslide-damaged area is very important for effective disaster response and recovery. For this purpose, various landslide detection approaches using remote sensing images have been studied. Especially, recent deep neural networks have been applied to remote sensing image analysis including landslide detection, which tries to overcome the conventional hand-crafted features. However, these deep learning-based methods usually require lots of labeled datasets which are extremely obtained or made in case of landslide detection. To overcome this problem, this study proposed an unsupervised landslide detection method to learn the pixel-lever deep feature representation. Nanosatellites, which have been launched in large numbers recently due to their low cost and accessibility, are useful for disaster management that requires rapid disaster response and damage analysis. For this purpose, the proposed method in this study was applied to the bi-temporal nanosatellite images, Planetscope. The experimental results showed enhanced detection performance than the conventional approaches using hand-crafted features.

Satellite-based moving vehicle (MV) detection plays a crucial role in traffic monitoring and surveillance. However, it presents challenges due to the large scale of scenes and the small size of MV features in satellite images. This study presents an optical flow-based approach for MV detection using single-pass WorldView-3 (WV-3) images. Representative images from two different multispectral sensors in WV-3 were created, and the Gunnar Farnebäck optical flow algorithm was employed to estimate MV motions. The estimated motion fields were then transformed into the HSV color space and segmented into eight ranges representing different moving direction angles. To refine the MV shapes and eliminate noise, opening operations and the DBSCAN clustering algorithm were applied. Finally, the MV detection results were obtained by combining eight images with bounding boxes. Although the algorithm exhibits high precision, challenges remain in accurately extracting MVs from the motion field, handling adjacent MVs and avoiding misinterpretations of non-moving edge features.

Video-and-Language learning, such as video question answering or video captioning, is the next challenge in the deep learning society, as it pursues the way how human intelligence perceives everyday life. These tasks require the ability of multi-modal reasoning which is to handle both visual information and text information simultaneously across time. In this point of view, a cross-modality attention module that fuses video representation and text representation takes a critical role in most recent approaches. However, existing Video-and-Language models merely compute the attention weights without considering the different characteristics of video modality and text modality. Such na ̈ıve attention module hinders the current models to fully enjoy the strength of cross-modality. In this paper, we propose a novel Modality Alignment method that benefits the cross-modality attention module by guiding it to easily amalgamate multiple modalities. Specifically, we exploit Centered Kernel Alignment (CKA) which was originally proposed to measure the similarity between two deep representations. Our method directly optimizes CKA to make an alignment between video and text embedding representations, hence it aids the cross-modality attention module to combine information over different modalities. Experiments on real-world Video QA tasks demonstrate that our method outperforms conventional multi-modal methods significantly with +3.57% accuracy increment compared to the baseline in a popular benchmark dataset. Additionally, in a synthetic data environment, we show that learning the alignment with our method boosts the performance of the cross-modality attention.

For the application of agricultural area, remote sensing techniques were studied and applied for its advantages for continuous and quantitative monitoring. Especially, hyperspectral images have been studied for the precise agriculture since they provide chemical and physical information of vegetation. In this study, we analyzed crop types using hyperspectral image data collected by a ground scanner. Spectral mixture analysis, which is widely used for processing hyperspectral images, was adopted for the crop discrimination. Endmember extraction algorithms used in this study were N-FINDR, Vertex Component Analysis (VCA), and Simplex Identification via variable Splitting and Augmented Lagrangian (SISAL), and classification was processed using fully constrained linear spectral unmixing (FCLSU). This study presents the application of spectral mixture analysis for hyperspectral scanner data at canopy level and optimal endmember extraction algorithms for different crop types for precise agriculture.

Iterative error analysis (IEA) is one of popular, sequential and linear constrained endmember extraction algorithm that uses spectral angle mapping (SAM) to calculate angles between spectral vectors. However, IEA has a limit that discriminating similar spectral vector is difficult because SAM does not consider positive and negative correlations. Since vegetation has similar spectral properties, it is difficult to classify different vegetation types. To improve IEA for various applications, such as crop classification and change detection, spectral similarity measures other than SAM have been applied to IEA. Many spectral similarity measures have been developed to calculate the similarities among spectral signatures and these are divided into the original methods and the newly developed hybrid algorithms. In this study, the original methods used were SAM, SCA, and SID, while the hybrid methods included SAMSID, SCASID, Jeffries-matusita measures-SAM (JMSAM), and normalized spectral similarity score (NS3). A Compact airborne spectrographic imager image including three crops and road was used and similarity values of four endmembers extracted by modified IEA were calculated. The CASI image was classified using endmembers and minimum distance classifier. The classification accuracy of the modified IEA with SMA, SCA, SID, SAMSID, SCASID, JMSAM, and NS3 were 84.45%, 85.56%, 61.47%, 65.83%, 62.11%, 93.47%, 90.29%. SID based algorithm has lower accuracy because SID tends to make two similar spectral signatures more similar. The results showed that JASAM was most effective to classify different vegetation types. The modified IEA with JMSAM could classify vegetation more effectively than the original IEA.

Most spaceborne sensors have an inevitable tradeoff between spatial and spectral resolutions. This is a typical ill-posed inverse problem in the field of image fusion. To solve this problem, this letter proposes an image fusion method using spatial principal component analysis and modulation transfer function-based filters. The key behind the proposed fusion method is to efficiently estimate the missing spatial details by considering the spatial structures of the low-resolution multispectral (MS) imagery. Also, this letter proposes a newly developed injection gain model to resolve the local and global dissimilarity between panchromatic and MS imageries, which could prevent over- and under-injections. Finally, spatial details, optimized to be injected into the MS images, were constructed and paired with the developed injection gain model to produce high-resolution MS images. Two data sets acquired by WorldView-2 are employed for validation. The experimental results demonstrate that the proposed fusion method generates high-quality imagery in terms of both qualitative and quantitative standards.

Irrigated rice fields in Asia have a distinct topography and scattering mechanisms. Because of these factors, many studies using radar backscattering information have been conducted to improve the mapping accuracy; however, relatively little attention has been paid to the topographic features (TF) provided by polarimetric synthetic aperture Radar (PolSAR) data. To address this issue, this letter presents a simple rice field mapping method in its late-vegetative stage using both the TF and scattering features (SF) of the PolSAR data. First, two TF, the polarization orientation angle and the dominant beta angle from the PolSAR data, were analyzed with respect to the slope from a digital elevation model. The results indicated that this feature pair had a moderately high correlation. As a result, flat areas were easily extracted using these two TF. Second, to use the SF of rice in the late-vegetative stage, the copolarized ratio and the copolarized phase difference were simply combined to map only paddy fields, of which scattering was dominated by the double-bounce scattering caused by the ground and canopy reflection. The results using C-band Radarsat-2 data were compared with a support vector machine classifier and demonstrated that the proposed mapping method provided an overall better performance. The proposed method might also be efficiently applied, with modification, for the aforementioned purpose in inaccessible areas.

In many agricultural applications, PolSAR data are widely used because they can be decomposed into various scattering components, which can be of help in observing the characteristics of agricultural areas. Recently, studies have been conducted to find suitable polarimetric parameters for specific applications. This paper tried to find appropriate polarimetric parameters for line extraction from agricultural areas as the line features are among the basic features of the surface. Towards this end, various polarimetric parameters were produced using polarimetric decomposition methods. LSD was used to extract lines from each polarimetric parameter image over agricultural areas, without any threshold value selection. The comparison of the line extraction result from each polarimetric parameter with the others was conducted through quantitative evaluation. Through this process, three parameters of the Pauli decomposition is found to be the suitable polarimetric parameters for line extraction from agricultural areas.

The crop boundary data provide the basic information for agricultural management. In this study, line-based boundary enhancement and extraction methods are proposed to delineate detailed paddy boundaries. Line extraction is performed based on the boundary enhancement result and Hough line extraction using RapidEye satellite image. The proposed line-based method is efficient in detecting the paddy boundaries while preserving linearity. In addition, the proposed method adopts an automated boundary detection process except for line editing. Therefore, the proposed method economically provides information on the paddy boundaries that can be utilized as the basic spatial unit for yield analysis, ownership management, and precision farming.

This letter proposes a method based on the fusion of high-resolution satellite images and airborne light detection and ranging (LiDAR) data for improving classification accuracy. Based on output-level fusion during classification, the proposed method utilizes a three-step process to minimize the misclassification of buildings and road objects. First, elevated road areas are detected in ground points, which are extracted for the generation of a digital terrain model based on statistical values. Second, building information is extracted from a satellite image through the output-level fusion of various data results. Third, supervised classification is conducted using a support vector machine for areas that lack elevated roads and buildings. We evaluated the proposed method by comparing it with a pixel-based method and analyzing experimental WorldView-2 images and airborne LiDAR data. We conducted a visual interpretation and quantitative accuracy assessment. The overall accuracy and kappa coefficient of the proposed method were 90.91% and 0.892, respectively. These results demonstrated an improvement in the overall accuracy and kappa coefficient by 11.27 percentage points and 0.135, respectively, compared with the pixel-based method. The results confirmed that our proposed method has significant potential for classifying urban environments using high-resolution satellite imagery and airborne LiDAR data.

We propose a building boundary modeling method using rectangles to overcome the zigzag boundary shapes of rooftop fenced buildings that are inherent in light detection and ranging (LiDAR) data. The method first finds 3-D patches through building detection using a morphological opening filter and histogram mode analysis. Then, rectangles that encase the traced boundaries of the 3-D patches are adjusted adaptively based on traced boundary coordinates and perpendicular constraints among the sides of the rectangles. The boundary of each 3-D patch is remodeled by subtracting void region rectangles from the parent rectangle. The superstructures of rooftops are also modeled with perpendicularity and adaptive constraints.

The objective of this paper is to extract a suitable number of evenly distributed matched points, given the characteristics of the site and the sensors involved. The intent is to increase the accuracy of automatic image-to-image registration for high-resolution multisensor data. The initial set of matching points is extracted using a scale-invariant feature transform (SIFT)-based method, which is further used to evaluate the initial geometric relationship between the features of the reference and sensed images. The precise matching points are extracted considering location differences and local properties of features. The values of the parameters used in the precise matching are optimized using an objective function that considers both the distribution of the matching points and the reliability of the transformation model. In case studies, the proposed algorithm extracts an appropriate number of well-distributed matching points and achieves a higher correct-match rate than the SIFT method. The registration results for all sensors are acceptably accurate, with a root-mean-square error of less than 1.5 m.

Most pansharpened images from existing algorithms are apt to present a tradeoff relationship between the spectral preservation and the spatial enhancement. In this letter, we developed a hybrid pansharpening algorithm based on primary and secondary high-frequency information injection to efficiently improve the spatial quality of the pansharpened image. The injected high-frequency information in our algorithm is composed of two types of data, i.e., the difference between panchromatic and intensity images, and the Laplacian filtered image of high-frequency information. The extracted high frequencies are injected by the multispectral image using the local adaptive fusion parameter and postprocessing of the fusion parameter. In the experiments using various satellite images, our results show better spatial quality than those of other fusion algorithms while maintaining as much spectral information as possible.

Precise image-to-image registration is required to use multi-sensor data implementing a diversity of applications related with remote sensing. The purpose of this paper is to develop an automatic algorithm that co-registers high-resolution optical and SAR images based on an integrated intensity-and feature-based approach. As a pre-registration step, initial differences between the translation of the x and y directions between images were estimated with the Simulated Annealing optimization method using Mutual Information as an objective function. After the pre-registration, the line features were extracted to design a cost function that finds matching features based on the similarities of their locations and gradient orientations. Only one feature at each regular grid region having a minimum value of cost function was selected as a final matching point to extract the large number of well-distributed points. The final points were then used to construct a transformation combining the piecewise linear function with the affine transformation to increase the accuracy of the geometric correction.

It is generally difficult to classify an object type having different colors into the same class using only optical data such as a satellite or aerial image. This paper proposes a method that solves this problem by combining LiDAR data and an aerial image. The method extracts building pixels from LiDAR data and then identifies building objects on the aerial image by overlaying the LiDAR result to a segmented aerial image through the definite rule. This process plays a role in transforming building objects of LiDAR data to ones of the aerial image.

Effective image-fusion methods inject the necessary geometric information and preserve the radiometric information. To preserve the radiometric information, the injected high frequency of a panchromatic (pan) image must follow the frequency of the multispectral (MS) image. In this letter, an improved additive-wavelet (AW) fusion method is presented using the à trous algorithm. The proposed method does not decompose the MS image; thus, it preserves the radiometric information of the MS image and can inject high frequency following the frequency of the MS image using a low-resolution pan image. Experimental results obtained using IKONOS data indicate that the proposed method produces superior-quality images compared with the AW luminance proportional method in a quantitative analysis.

Preservation of spectral information and enhancement of spatial resolution are regarded as important issues in remote sensing satellite image fusion. In previous research, various algorithms have been proposed. Although they have been successful, there are still some margins of spatial and spectral quality that can be improved. In addition, a new method that can be used for various types of sensors is required. In this paper, a new adaptive fusion method based on component substitution is proposed to merge a high-spatial-resolution panchromatic (PAN) image with a multispectral image. This method generates high-/low-resolution synthetic component images by partial replacement and uses statistical ratio-based high-frequency injection. Various remote sensing satellite images, such as IKONOS-2, QuickBird, LANDSAT ETM+, and SPOT-5, were employed in the evaluation. Experiments showed that this approach can resolve spectral distortion problems and successfully conserve the spatial information of a PAN image. Thus, the fused image obtained from the proposed method gave higher fusion quality than the images from some other methods. In addition, the proposed method worked efficiently with the different sensors considered in the evaluation.

Despite the recent developments in light detection and ranging systems, discrepancies between strips on overlapping areas persist due to the systematic errors. This letter presents an algorithm that can be used to detect and adjust such discrepancies. To achieve this, extracting conjugate features from the strips is a prerequisite step. In this letter, linear features are chosen as conjugate features because they can be accurately extracted from man-made structures in urban area and more easily extracted than the point features. Based on such a selection strategy, a simple and robust algorithm is proposed that is generally applicable for extracting such features. The algorithm includes methods that can be used to establish observation equations from similarity measurements of the extracted features. Then, several transformations are selected and used to adjust the strips. Following the transformation, the fitness of linear features is tested to determine whether the discrepancies have been resolved; the results are then evaluated statistically. The results demonstrate that the algorithm is effective in reducing the discrepancies between the strips.