-
Taneth RUANGRAJITPAKORN, Wasan NA CHAI, Thepchai SUPNITHI
原稿種別: PAPER
論文ID: 2026DKP0011
発行日: 2026年
[早期公開] 公開日: 2026/11/12
ジャーナル
フリー
早期公開
Large language models (LLMs) have shown potential for evaluating textual information against predefined criteria, but criteria-based assessment often requires more than semantic interpretation. Evidence must be identified and judged against explicit conditions, while complex criteria may require sequential scoring and detailed procedural rules. This study presents an LLM-based approach for criteria-based assessment and feedback generation that separates evidence interpretation from procedural assessment. The approach uses instructional prompts and a procedural decision graph to represent assessment knowledge and applies LLMs to interpret organisational reports, performcriterion-level assessment, and generate reasoning and improvement feedback. PMQA 4.0 is used as a real-world case study involving heterogeneous organisational reports, multiple assessment criteria, and detailed binary and sequential scoring procedures. Evaluation on real assessment data showed that the system produced results generally consistent with human assessment across both binary and sequential scoring. Expert evaluation further indicated high correctness and relevance of the generated reasoning and feedback, with high inter-rater agreement among the assessors. Analysis of disagreements showed complementary knowledge boundaries between human and automated assessment. Human assessors could incorporate contextual knowledge beyond the submitted reports, while the system provided consistent application of explicitly represented assessment rules and calculations. The results demonstrate the potential of combining LLM-based evidence interpretation with explicit procedural knowledge for supporting complex criteria-based assessment while retaining human judgement for cases requiring contextual interpretation.
抄録全体を表示
-
Toshihide UBUKATA, Zhiyao WANG, Yusei ISHIMIZU, Enhong MU, Jialong LI, ...
原稿種別: PAPER
論文ID: 2026EDP7061
発行日: 2026年
[早期公開] 公開日: 2026/11/05
ジャーナル
フリー
早期公開
On-the-fly Directed Controller Synthesis (OTF-DCS) mitigates state-space explosion by incrementally exploring the system and relies critically on an exploration policy to guide search efficiently. Recent reinforcement learning (RL) approaches learn such policies and achieve promising zero-shot generalization from small training instances to larger unseen ones. However, a fundamental limitation is anisotropic generalization, where an RL policy exhibits strong performance only in a specific region of the domain-parameter space while remaining fragile elsewhere due to training stochasticity and trajectory-dependent bias. To address this, we propose a Soft Mixture-of-Experts framework that combines multiple RL experts via a prior-confidence gating mechanism and treats these anisotropic behaviors as complementary specializations. The evaluation on the Air Traffic benchmark shows that Soft-MoE substantially expands the solvable parameter space and improves robustness compared to any single expert.
抄録全体を表示
-
Fumiyoshi MATANO, Masanori MORISE
原稿種別: PAPER
論文ID: 2026EDP7101
発行日: 2026年
[早期公開] 公開日: 2026/11/05
ジャーナル
フリー
早期公開
Text-to-speech synthesis can improve output audio quality by performing pre-processing to remove silent intervals surrounding speech segments. As a result, some implementations such as ESPnet actually implement silent interval removal. While various methods exist for silent interval detection, including Voice-Activity-Detection techniques, this particular application requires no judgment for silent intervals that occur within speech segments. This study introduces the Voice Start and End Detection (VSED) task, which is formulated by adding assumptions to VAD, such as “at least one valid speech segment must always exist”. We also present an algorithm to solve this task, named SATEN. SATEN employs a signal processing approach based on Root-Mean-Square and Zero-Crossing-Rate metrics. When tested under clean conditions and simulated background noise environments (white, pink, and fan noise), the algorithm demonstrated competitive boundary estimation accuracy in terms of MdAE compared with tested methods under Signal-to-Noise Ratios above 0 dB. Furthermore, SATEN showed robust performance even against pulse noise.
抄録全体を表示
-
Peiling LI, Huawei TAO, Yue XIE
原稿種別: LETTER
論文ID: 2026EDL8063
発行日: 2026年
[早期公開] 公開日: 2026/10/29
ジャーナル
フリー
早期公開
Acoustic scene classification (ASC) aims to automatically identify the acoustic environment of an audio recording. The core challenge of ASC lies in capturing the relationships among salient acoustic events that characterize specific scenes. Existing methods, however, often do not adequately leverage these relationships, thereby limiting overall classification performance. To address this challenge, we propose a novel Frequency-Aware Model (FAM) for ASC. Specifically, we first extract log-mel spectrograms from the raw audio signals as inputs to the model. The FAM architecture integrates a convolutional neural network (CNN), Conformer modules, and a Transformer encoder to effectively and hierarchically model both local and global dependencies of acoustic events. In the first stage, the log-mel spectrogram is evenly divided along the frequency axis into four bands, and each band is processed independently through CNN layers to extract preliminary band-specific features. Subsequently, the features from each frequency band are input to Conformer modules to capture band-specific global correlations and salient frame-level events. Finally, the processed features are concatenated along the frame dimension and input to a Transformer encoder to produce the final scene representation. Experimental evaluations on the DCASE2019 and DCASE2020 datasets demonstrate that the proposed FAM model achieves accuracies of 83.52% and 73.72%, respectively, outperforming the official DCASE baselines by 23.82% and 11.22%.
抄録全体を表示
-
Gengtian SHI, Hiroshi ONODA, Jiang LIU
原稿種別: PAPER
論文ID: 2026EDP7127
発行日: 2026年
[早期公開] 公開日: 2026/10/29
ジャーナル
フリー
早期公開
Diffusion models are a promising paradigm for end-to-end autonomous driving motion planning because they natively represent multimodal futures, support test-time guidance without retraining, and expose interpretable intermediate trajectories during denoising. In practice, however, they face three problems. First, diffusion planners produce trajectories with an average jerk of 3.43 m/s3, above the 2 m/s3 comfort threshold used in vehicle dynamics, and this comfort failure is invisible to the displacement-only metrics (ADE/FDE) the field has standardized on. Second, 50-100 denoising steps strain onboard latency budgets. Third, a single diffusion sample carries no collision guarantee. We make two contributions. (1) Anchor-based initialization: the reverse process starts from one of K=6 smooth K-means maneuver centroids at an intermediate noise level and executes only 15 of 50 steps (3.3× fewer). Controlled ablations isolate the mechanism: reducing steps alone leaves jerk unchanged (3.43 → 3.38 m/s3), whereas centroid initialization reduces jerk by 64% (3.43 → 1.25 m/s3) and ADE from 0.794 to 0.742 m (single training runs), with collision rate within seed variation. A λ ablation shows that an auxiliary centroid-classification loss is unnecessary and degrades all metrics, so the method changes inference only. (2) Best-of-N safety sampling: N stochastic candidates are ranked by a collision-event count computed from ground-truth agent futures (an oracle assumption). Over three seeds the global oracle collision rate does not change significantly (N=1: 0.371%±0.133; N=5: 0.399%±0.041, 95% CI); a conditional analysis shows that only 0.73% of scenes contain any colliding candidate and that, where selection can act, it picks a collision-free candidate 76.3% of the time. The safety result is therefore preliminary. All evaluation is open-loop on nuScenes, so the conclusions do not extend to interactive driving without closed-loop validation.
抄録全体を表示
-
Dong Yun JEONG, Jung Min LIM, Jong Wook KWAK
原稿種別: LETTER
論文ID: 2026EDL8027
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
In this letter, we propose a depth estimation method called Adaptive Confidence Fusion Network (ACFNet) for robust estimation under adverse weather conditions. The main idea of ACFNet is to independently estimate the confidence of camera and radar sensors and adaptively fuse them, thereby improving robustness to degradation in sensor reliability under heavy rain or snow. ACFNet adjusts the fusion weights through learnable confidence scaling, enabling its fusion module to balance sensor contributions according to their reliability. The evaluation results show that ACFNet improves depth accuracy and scale consistency compared with existing camera-only and radar-camera fusion methods on the Boreas dataset, reducing the Mean Absolute Error (MAE) by up to 67.8% and the Absolute Relative Error (AbsRel) by 46.3%, and demonstrating robust and reliable depth estimation for autonomous driving under adverse weather conditions.
抄録全体を表示
-
Zhe WEI, Zhipan DENG, Huitong YOU, Fang WANG
原稿種別: LETTER
論文ID: 2026EDL8036
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
We propose TemporalLSTM-Net (TL-Net), a lightweight dual-stream framework for early hypoxemia warning in elderly PACU patients. TL-Net combines BiLSTM and TS-Mixer encoders with adaptive gating to fuse physiological signals and clinical phenotypes. Validated on 766 patients, TL-Net achieves AUC 0.934, sensitivity 91.50%, precision 78.20%, and supports 4-minute advance warnings in internal patient-level validation. With 226K parameters and sub-5 ms latency, TL-Net reduces parameters by 75.5% versus Transformer alternatives. Ablation confirms that adaptive gating improves sensitivity by 5.05 percentage points over fixed fusion.
抄録全体を表示
-
Weijie DUAN, Jiale WANG, Bo JIANG, Hong LIU, Weifeng PAN
原稿種別: PAPER
論文ID: 2026EDP7066
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
Software reliability is a critical quality attribute, and software defect prediction (SDP) is a key means to improve and maintain reliability. Recent semantic representation-based SDP models typically extract information from abstract syntax trees (ASTs) or software dependency relations, and have achieved better performance over traditional approaches using handcrafted features. However, most existing methods either depend on a single representation or linearize ASTs into sequences, which may lose structural information, constrain defect detection capability, and increase the false-positive rate. To address these issues, we propose DPACD (Defect Prediction with AST, Context and Dependencies), a multi-view defect prediction model that jointly leverages internal semantics, internal structure, and external structure. Specifically, we feed the AST token sequence into a convolutional neural network to capture internal semantics; we represent internal structure by encoding AST path contexts (i.e., source token, connecting path, and target token) into vectors and aggregating multiple contexts to form a structural embedding; and we model the class dependency network with a graph convolutional network to obtain external structural embeddings. We then aggregate internal semantic features, internal structural embeddings, and external structural embeddings into a unified representation, and further concatenate it with traditional hand-crafted features for defect classification. Extensive experiments on 20 open-source Java projects with a total of 34 versions show that DPACD consistently outperforms recent state-of-the-art SDP models in both within-project (including cross-version) and cross-project scenarios, demonstrating the effectiveness of fusing multi-view semantic and structural features for software defect prediction.
抄録全体を表示
-
Lin TENG, Kohei NISHIKAWA, Hayato YAMAKI, Hiroki HONDA, Ryota SHIOYA, ...
原稿種別: PAPER
論文ID: 2026EDP7071
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
High-bandwidth memory (HBM), originally developed for GPUs, is increasingly being adopted as main memory for CPUs. While current HBM-equipped CPUs mainly target high-performance computing (HPC), the slowdown in DDR memory performance scaling suggests that HBM adoption is likely to expand to more diverse computing systems such as cloud platforms. A straightforward way to exploit HBM bandwidth is to increase cache block size; however, server workloads often exhibit low spatial locality, so larger blocks do not necessarily improve performance. In this paper, we propose a new cache architecture for HBM-equipped CPUs to improve server-workload performance. The proposed architecture dynamically resizes cache blocks by dividing large blocks into smaller sub-blocks and adaptively changing the number of sub-blocks stored according to the memory access patterns of running applications. Our ChampSim- and CACTI-based evaluations showthat the proposed architecture improves IPC by 6.57% on average for server workloads, compared with a conventional 64B-block cache architecture, without increasing the area-power product.
抄録全体を表示
-
Songsong FANG, Shicong HAN, Yue CHEN, Xiaoxiao SUN
原稿種別: PAPER
論文ID: 2026EDP7091
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
Business processes are prone to deviate from predesigned reference models in order to adjust to the evolving business environment. Few of the deviations are well-documented while most of them remain unnoticed. Process drift detection helps enterprises to locate and analyze these changes during process execution. Previous researches in this field primarily relied on comparing features that are extracted from two successive sliding windows to detect drifts, which require users to manually select these features or directly explore a vast feature space. Moreover, most of them focus on locating drift points but lack the capability for drift classification. To bridge the gap, in this article, we present a novel process drift detection method based on next activity prediction. The drift points are detected by comparing the performance of the prediction model overtime, while the categories of the drifts (i.e., sudden or gradual drift) are further determined through overlap analysis.A statistical-based evaluation on both synthetic and real-life event logs demonstrate that the proposed method outperforms the evaluated baselines, achieving the highest macro F1-score.
抄録全体を表示
-
Juntao WANG, Tsunenori MINE
原稿種別: PAPER
論文ID: 2026MUP0005
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
Emotion recognition in conversations (ERC) plays a crucial role in enhancing the interaction quality of modern conversational agents, yet remains challenging because the emotion of a target utterance often depends on surrounding dialogue context and speaker-specific affective dynamics. This work investigates three hypotheses: contextual information outside the target utterance is useful for ERC; prosodic variation can provide informative cues for contextual attention allocation; and dialogue-level and speaker-level contexts capture distinct aspects of conversational context that may be jointly useful for ERC. Based on these hypotheses, we propose CAPSNet, a multimodal ERC model that integrates textual and speech modalities and models context at both dialogue and speaker levels. The core component of CAPSNet is Contextual Attention Conditioned on Prosodic States (CAPS), which uses continuous normalized pitch-variation descriptors to guide contextual attention through semantic prosody gating and structural prosody bias. Experiments on IEMOCAP and MultiDialog show that CAPSNet achieves strong performance under the text-audio setting. Ablation and diagnostic analyses further provide supporting evidence for contextual modeling, prosody conditioning, and the use of both dialogue-level and speaker-level tracks.
抄録全体を表示
-
Nopparut LI, Candy Olivia MAWALIM, Gia-Huy BUI, Md. Mostafizur RAHMAN, ...
原稿種別: PAPER
論文ID: 2026MUP0012
発行日: 2026年
[早期公開] 公開日: 2026/10/28
ジャーナル
フリー
早期公開
The rapid advancement of speech synthesis systems has enable the generation of realistic speech signals capable of imitating specific individuals, thereby posing serious security risks such as targeted impersonation attack. While recent studies have focused on developing deepfake detection models, these approaches are inherently reactive and can only be applied after the synthesized speech has been generated. In this study, we propose a signal processing technique based on fundamental frequency (F0) component elimination filtering and benchmark it against conventional signal processing methods as well as AntiFake, a state-of-the-art adversarial example approach. Using speech generated by the Tortoise Text-to-Speech (TTS) synthesis model, we evaluate both their effectiveness in reducing the identifiably of synthesized speech signals as well as their impact on perceptual quality of modified speech signals compare to the original. Experimental results show that our proposed method achieve minimal degradation in quality while offering protection against speech synthesis system, highlighting its potential for voice privacy protection in real-world applications.
抄録全体を表示
-
Dong Kwan KIM
原稿種別: LETTER
論文ID: 2026EDL8056
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
We ask how often a commit that changes production source code also changes a test file—test-file co-delivery—and whether agent-attributed and human-attributed commits differ. In 20,766 source commits from 52 repositories, agent commits co-deliver test files in 40.8% of cases against 25.0% for human commits (odds ratio 2.06). We show that this contrast is largely a composition effect. Leaving general-purpose automation bots in the non-agent group inflates it to 3.83; and because agent commits are recent, most human commits come from an earlier, less-tested period of the same repository. Under stratification by repository and calendar month no statistically significant difference is detected (OR 0.98, 95% CI [0.84, 1.15]), and the same holds when test-file deletions and renames are excluded from the outcome. Raw agent-versus-human comparisons of test-file co-delivery need to control both confounds.
抄録全体を表示
-
Emuun PUREVDAGVA, Masayuki SHIMODA, Satoshi TAYU, Atsushi TAKAHASHI
原稿種別: PAPER
論文ID: 2026EDP7076
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
For efficient manipulation of Micro Electrode Dot Array (MEDA) biochips, droplet routing that completes droplet movements in a minimum number of time steps using a small number of cells is required. In this paper, we consider routing of a single type of droplets, where unit droplets are treated as indistinguishable and the routing specifies their occupied-cell configuration at each time step. In FCDR, droplet routing is represented by a flow in a 3D flow graph. FCDR first finds a feasible routing by gradually increasing the number of layers of the 3D flow graph until a feasible droplet routing is obtained. Then, by assigning edge weights that penalize the use of cells that are infrequently used in the obtained routing and solving a minimum-cost flow problem, it derives a droplet routing that uses fewer cells.
抄録全体を表示
-
Ji WU, Song WANG, Xiangqing WEI, Kazuteru NAMBA
原稿種別: PAPER
論文ID: 2026EDP7088
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
Deploying quantized neural networks on edge devices is challenged by transient voltage droops, which degrade standard 6-Transistor (6T) SRAM stability and cause catastrophic accuracy cliffs. While robust 8-Transistor (8T) SRAMs solve this, 28 nm silicon data reveals they incur a prohibitive +47.8% area penalty and +333.6% read power surge. To break this area-robustness trade-off, we propose an Algorithm-Hardware Co-Design featuring a heterogeneous 8T/6T SRAM architecture. Using a data-driven Protection Loss Score and 1D K-Means clustering, we mathematically partition weight bits into High-Protected Bits (HPB) and Low-Protected Bits (LPB) without human bias. This classification is seamlessly mapped onto a mixed memory array under a unified supply voltage, assigning 8T cells exclusively to critical HPBs and 6T cells to resilient LPBs. Evaluations show our Co-Design successfully flattens the accuracy cliff, extending the tolerable bit-error boundary by an order of magnitude. Ultimately, it achieves the system-level robustness of a pure 8T array with a negligible area overhead of only 6.0% to 11.9%.
抄録全体を表示
-
Shunta YAMAZAKI, Tomomi MATSUI
原稿種別: PAPER
論文ID: 2026FCP0001
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
In this research, we discuss a problem of calculating the Shapley value in bankruptcy games. We investigate the relationship between the Shapley value of bankruptcy games and the Shapley-Shubik index in weighted voting games. The relation naturally implies a dynamic programming technique for calculating the Shapley value. We further show that the parametric analysis of the Shapley value with respect to changes in the estate is computationally hard. We then present two recursive algorithms for computing the Shapley value: the first is the recursive completion method originally proposed by O'Neill, and the second is our novel contribution based on the dual game formulation. These recursive approaches offer conceptual clarity and computational efficiency, especially when combined with memoisation technique. Finally, we propose a Fully Polynomial-Time Randomized Approximation Scheme (FPRAS) based on Monte Carlo sampling, providing an efficient approximation method for large-scale instances.
抄録全体を表示
-
Keito MAENO, Mashiho MUKAIDA, Satoshi ONO
原稿種別: LETTER
論文ID: 2026MUL0001
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
Recent studies have shown that deep neural networks (DNNs) can misclassify inputs with adversarial perturbations. Most existing defenses rely on digital-domain processing and access to model internals, which restricts their applicability in practical systems. This study proposes a projection-based re-attacking defense method that physically mitigates the influence of adversarial attacks. The proposed method employs physics-in-the-loop (PITL) optimization, which evaluates candidate solutions in actual environments to account for device specifications and disturbances. Experimental results demonstrate successful defense against patch- and texture-based physical adversarial attacks when the ground-truth class is known.
抄録全体を表示
-
Keiichi MORI, Masaki KAWAMURA
原稿種別: PAPER
論文ID: 2026MUP0003
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
This paper proposes an enhanced feature extraction method that uses pre-masked weights to improve the robustness of associative watermarking for protecting convolutional neural networks (CNNs). Although the original associative watermarking framework offers a zerobit watermarking scheme, it remains vulnerable to structural modifications such as pruning. To address this issue, we introduce a masking process in which only a certain proportion of the weights are selected as features. Through computer simulations involving ten different CNNs, we demonstrate that the proposed method significantly improves robustness against weight-pruning attacks. Specifically, when the pruning rate of the attack is less than or equal to the masking rate R, the extracted features remain error-free. Furthermore, integrating a dual-structured associative memory consisting of hetero- and auto-associative models enables the system to reliably retrieve watermarks even under extreme pruning conditions. We also evaluated the applicability of our method in transfer learning scenarios involving fine-tuning attacks. The results show that, with an appropriate masking rate, the proposed method can effectively maintain a bit error rate of zero even when the network weights are significantly updated. Our findings confirm that combining masked weight features and associative memory provides a powerful, practical solution for securing deep learning models against unstructured and weight-based modifications.
抄録全体を表示
-
Ou EGAMI, Masaki KAWAMURA
原稿種別: PAPER
論文ID: 2026MUP0006
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
We propose FlexDT, a flexible and retraining-free encoder-decoder for latent diffusion model (LDM)-based image generation. FlexDT addresses the message length constraints of the conventional DiffuseTrace. Unlike DiffuseTrace, which requires reconstructing a variational autoencoder (VAE) for different message lengths, incurring significant computational costs, FlexDT uses a random network-based encoding and decoding mechanism that supports variable-length messages without VAE re-training. Specifically, we simplify the random network into a single-layer architecture, which reduces the matrix computation in encoding and decoding by at least 50% compared to the two-layer structure while maintaining high embedding flexibility. We evaluated the accuracy of extracting watermarks and the perceptual quality of generated images. The experimental results demonstrate that for message lengths up to 1024 bits, the single-layer FlexDT achieves superior extraction accuracy (Acc = 0.993) and outperforms both the conventional two-layer architecture and the DiffuseTrace baseline. This suggests that simplifying the network effectively suppresses computational error accumulation. Furthermore, evaluations using NIQE and CLIP scores confirm that FlexDT preserves high perceptual quality and semantic consistency with input prompts even after watermark embedding. FlexDT also exhibits robust resistance against various image processing attacks. These findings confirm that FlexDT is a highly accurate, flexible, and practical digital watermarking solution for generative AI applications.
抄録全体を表示
-
Michiharu NIIMI, Shoto YAMASHITA, Daishi TANAKA
原稿種別: PAPER
論文ID: 2026MUP0007
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
This paper describes methods for detecting images generated by diffusion models using sparse coding. In recent years, the rapid advancement of diffusion models has enabled the generation of highly realistic images, making it increasingly difficult to distinguish between real and generated images. Motivated by the differences in the image formation process between real and generated images, we exploit statistical properties of sparse coding coefficients as well as reconstruction-based features. Specifically, we extract image patches, represent them using sparse codes, and derive statistical features, while also utilizing reconstruction error as an additional cue for detection. We evaluate the proposed method on a dataset of 6,000 images consisting of real and diffusion-generated images under various conditions, including different dictionary construction settings and classification schemes. Experimental results demonstrate the effectiveness of the proposed approach and suggest the potential of sparse coding coefficients for robust diffusion-generated image detection.
抄録全体を表示
-
Haruka OKANO, Ryohei ORIHARA, Yasuyuki TAHARA, Akihiko OHSUGA, Yuichi ...
原稿種別: PAPER
論文ID: 2026MUP0009
発行日: 2026年
[早期公開] 公開日: 2026/10/16
ジャーナル
フリー
早期公開
Binaural audio generation that converts monaural recordings into spatial audio using visual cues can enable immersive video experiences without specialized equipment. However, most existing studies have been evaluated using pseudo-monaural audio derived from binaural recordings. Consequently, their applicability to real monaural recordings, which lack the interaural cues that these methods rely on, has remained unverified. To address this and enable verification in real-world scenarios, the RealBinaural dataset was constructed, containing real monaural and binaural recordings aligned with corresponding video clips. Experimental results with baseline methods revealed that while they function effectively with pseudo-monaural input, their spatial reproduction capability collapses with real monaural input, resulting in a sound source localization accuracy of 0%. To resolve the challenge of real-world application, we propose DiffBinaural, a conditional diffusion model for binaural audio generation. Crucially, this model incorporates the proposed Position-Aware Vision Features (PAVF) module, which integrates visual and positional information of sound sources. Experimental results demonstrate the clear superiority of the proposed method, achieving an approximate 40% reduction in mel-spectrogram reconstruction error and a nearly 30% increase in sound source localization accuracy compared to baseline methods. Further analysis on the RealBinaural dataset shows that, when switching from pseudo-monaural to real-monaural inputs, baseline methods degrade substantially, whereas our method is far more robust. This study demonstrates the limitations of applying conventional methods to real-world environments and presents a practical solution for enriching ordinary videos with immersive spatial audio experiences.
抄録全体を表示
-
Linh-Chi LE, Bima PRIHASTO, Wenny RAMADHA PUTRI, Manh QUAN BUI, Jia-Ch ...
原稿種別: LETTER
論文ID: 2026EDL8045
発行日: 2026年
[早期公開] 公開日: 2026/10/05
ジャーナル
フリー
早期公開
Voice conversion (VC) is the process of changing the speaker identity in a voice recording while keeping the spoken content unchanged. Traditional VC systems often require parallel speech data, which can be challenging and requires significant time to gather. To overcome this limitation, recent methods have focused on non-parallel data and the use of generative adversarial networks (GANs). In this work, we propose a new model called SICL-GVC, which combines GAN-based voice conversion with speaker-invariant contrastive learning. Our method is designed to better separate speaker identity from speech content, allowing the model to focus on what is being said rather than who is saying it. We evaluate our model on one-to-one and many-to-one voice conversion tasks using the VCTK dataset. Experimental results show that SICL-GVC achieves higher voice similarity compared to existing models, demonstrating better voice conversion quality and generalization across different speaker pairs.
抄録全体を表示
-
Jun ZHANG, Toru MIKAMI, Keiichi KANEKO
原稿種別: PAPER
論文ID: 2026EDP7015
発行日: 2026年
[早期公開] 公開日: 2026/10/05
ジャーナル
フリー
早期公開
In this study, we propose a metatorus, which is a topology obtained by reducing the number of links connected to each processing element to four, for the torus, a famous topology of interconnection networks of massively parallel computers. The diameter of the k-ary (n+1)-dimensional metatorus is n-2+⌊n/2⌋+n⌊k/2⌋, and the average distance between an arbitrary pair of nodes is n⌊k/2⌋⌈k/2⌉/k+⌊n/2⌋⌈n/2⌉/n+Θ(n). In addition, we propose a shortest-path routing algorithm. We measured the average distance by conducting a computer experiment, in which the algorithm is used. Furthermore, we propose stochastic link-fault-tolerant routing algorithms, and evaluated their performance by a computer experiment.
抄録全体を表示
-
Anam SHAHZADI, Kazuya KITANO, Yuki FUJIMURA, Takuya FUNATOMI, Yasuhiro ...
原稿種別: PAPER
論文ID: 2026EDP7100
発行日: 2026年
[早期公開] 公開日: 2026/10/05
ジャーナル
フリー
早期公開
Automatic shuttle hit detection is fundamental to badminton video analysis, yet existing methods produce coarse outputs that lack frame-level temporal precision. We propose ShuttleHitFormer, an end-to-end multimodal Transformer that integrates player poses, shuttle trajectory, and court-net geometry through hierarchical cross-modal interaction modeling with asymmetric gated player fusion and contrastive player-shuttle interaction learning. Sliding window inference with non-maximum suppression produces precise per-event detections directly comparable to ground-truth hit frame annotations. Evaluated on the ShuttleSet22 benchmark, ShuttleHitFormer achieves precision of 93.08%, recall of 97.24%, F1 of 95.11%, and mean absolute error of 0.44 frames, outperforming the best prior method by 5.1 F1 points. Ablation studies confirm the contribution of each input modality and architectural component. Shot-type analysis across ten stroke categories reveals clear performance stratification, with clear shots achieving 96.4% detection and short services presenting the primary challenge at 66.9%.
抄録全体を表示
-
Shogo EGUCHI, Akiyoshi SHIOURA
原稿種別: PAPER
論文ID: 2026FCP0004
発行日: 2026年
[早期公開] 公開日: 2026/10/05
ジャーナル
フリー
早期公開
We discuss a constrained version of the simple assembly line balancing problem of type 2 (SALBP-2), where tasks should be assigned to stations so that the cycle time is minimized, under the constraint that some tasks are already assigned to stations. We propose a polynomial-time approximation algorithm for this problem with a theoretical guarantee of additive error.
抄録全体を表示
-
Ryo SUZUKI, Yutaro YAMAGUCHI
原稿種別: PAPER
論文ID: 2026FCP0005
発行日: 2026年
[早期公開] 公開日: 2026/10/05
ジャーナル
フリー
早期公開
In the sliding window model, we are required to maintain the target statistics over the most recent n elements of a data stream, which is captured by a window of size n sliding over the data stream. Exact computation usually requires space linear in n, and the central goal is approximate maintenance using sublinear space. In this paper, we study the problem of maintaining the maximum subarray sum in the sliding window model. While the classical Kadane's algorithm computes the exact answer using constant space in the static setting, it does not extend directly, because a new element makes the oldest one expire, which may invalidate the optimal subarray so far. Our first observation is that the so-called Smooth Histogram framework can lead to a constant-factor approximation (in the sense of relative error) using O((log n)2) bits of space. We then refine this framework accordingly, which enables for any ϵ > 0 to maintain a (1±ϵ)-approximation using O(ϵ-1(log n)2) bits of space and O(ϵ-1 log n) operations per update. The space complexity is asymptotically optimal.
抄録全体を表示
-
Mingli WU, Bin LI, Yingyu ZHANG, Jianyong DUAN
原稿種別: PAPER
論文ID: 2025EDP7261
発行日: 2026年
[早期公開] 公開日: 2026/10/02
ジャーナル
フリー
早期公開
To overcome the limitations of existing datasets and large language models in reading comprehension tasks requiring logical reasoning, which often lack sufficient diversity and logical depth, we propose a method named ”Multi-dimensional Data Augmentation” (MDA) for training language models. First, we utilize open-source large language models to augment the dataset from various dimensions. This includes expanding the dataset by incorporating different question types across various domains, as well as performing implicit semantic expansion for each option in multiple-choice questions. We adopt a phased, differentiated fine-tuning strategy: initially fine-tuning a model to generate the implicit semantics of the options, followed by fine-tuning another model to perform unmasked decision-making based on these generated semantics. Experimental results show that our method achieves stable and enhanced performance compared to baseline models, achieving 85.50% accuracy in the ReClor open evaluation and outperforming other models with a comparable parameter count (8B scale). Our submission achieved sixth place on the ReClor leaderboard at the time of our evaluation submission. This demonstrates the effectiveness and potential of our method in enhancing the logical reasoning capabilities of large language models.
抄録全体を表示
-
Menglong WU, Shuyi REN, Ming HUANG
原稿種別: LETTER
論文ID: 2026EDL8040
発行日: 2026年
[早期公開] 公開日: 2026/10/02
ジャーナル
フリー
早期公開
Respiratory sound classification suffers from significant domain shift induced by heterogeneous recording equipment, substantially degrading model generalization. Existing methods either overlook this shift or apply fully fine-tuned domain adversarial neural networks (DANN), whose objectives remain difficult to optimize in high-dimensional spaces. We propose LoRA-DANN, a parameter-efficient domain adaptation framework combining low-rank adaptation (LoRA) with adversarial and contrastive objectives. LoRA limits trainable parameters to 0.3M and restricts optimization to a structured low-dimensional subspace, empirically improving adversarial training stability. Because DANN only aligns marginal distributions, risking class-conditional misalignment, we introduce cross-device contrastive learning (CDCL). By redefining positive pairs as same-class samples across devices, CDCL explicitly enforces class-conditional device invariance. Experiments on the ICBHI 2017 benchmark demonstrate our method achieves a 62.00% score, with ablation studies validating each component's independent contribution.
抄録全体を表示
-
Changhoon KIM
原稿種別: LETTER
論文ID: 2026EDL8030
発行日: 2026年
[早期公開] 公開日: 2026/09/24
ジャーナル
フリー
早期公開
Concept erasure methods suppress harmful content by removing entire semantic categories from vision-language embedding spaces, implicitly assuming that harmfulness is a property of objects. We argue it is not. A child holding a loaded firearm and a biathlete carrying a competition rifle share the same object-level representation yet differ categorically in danger; the harm resides in the situation, not the object. We introduce situational concept erasure, which redirects embeddings only when the full depicted situation is harmful while explicitly preserving the same object in legitimate contexts. Evaluated on the weapon domain in text-to-image generation downstream of the CLIP encoder, our method reduces harmful situation alignment without suppressing legitimate ones. We further identify cosine distance anchoring to frozen reference embeddings as the dominant structural factor: without it, naively fine-tuning with contrastive losses alone causes absolute embedding drift regardless of loss reweighting.
抄録全体を表示
-
Yaqian HUANG, Wenjun XU, Jiale WANG, Cheng ZENG, Weifeng PAN, Yun PAN, ...
原稿種別: PAPER
論文ID: 2026EDP7112
発行日: 2026年
[早期公開] 公開日: 2026/09/17
ジャーナル
フリー
早期公開
Key classes are considered as the optimal starting point for comprehending complex software. Although numerous key class prediction methods have been proposed, they still suffer from the following limitations: (1) Existing studies primarily rely on traditional machine learning algorithms, with relatively few approaches utilizing Graph Neural Networks (GNNs); (2) Existing studies depend solely on software networks composed of classes and their dependencies to extract structural features of classes, neglecting their semantic features; and (3) Existing GNN-based key class prediction methods rely on software networks that overlook many important inter-class dependencies, failing to accurately represent the true software structure. To address these limitations, we propose CAPE, a key class prediction method via fusing semantic and structural features. First, CAPE parses the source code of classes into abstract syntax trees, subsequently extracts token sequences, and utilizes convolutional neural networks to capture semantic features of the classes. Second, CAPE abstracts the class-level structure of software into a class dependency network and employs a graph convolutional network to extract structural features of the classes. Finally, CAPE constructs the final feature representation of classes by fusing the semantic and structural features. It then trains a random forest classifier for the key class prediction task. Experimental results on a benchmark dataset demonstrate that: (1) CAPE not only significantly outperforms baselines but also demonstrates moderate to substantial practical advantages, as evidenced by Cliff's delta values; and (2) The fusion of structural and semantic features yields the highest overall performance, with structural features serving as the primary determinant and semantic features offering marginal yet complementary gains.
抄録全体を表示
-
Yuki KOTAKA, Yu YOKOI
原稿種別: PAPER
論文ID: 2026FCP0007
発行日: 2026年
[早期公開] 公開日: 2026/09/17
ジャーナル
フリー
早期公開
We study congested assignment problems in which each agent ranks post-congestion pairs and is assumed to be congestion-averse; that is, for any fixed post, each agent strictly prefers lower congestion. Two central solution concepts are Nash stability and competitiveness. A Nash-stable outcome always exists, but it may be unfair; by contrast, a competitive outcome is envy-free and non-wasteful, but may fail to exist. We therefore study algorithmic questions that bridge the gap between these notions, including reducing envy within Nash-stable outcomes, requiring competitiveness for a designated subset of agents, and maximizing the number of competitive agents subject to overall Nash stability. For the hybrid existence problem, we prove polynomial-time solvability in two special cases: when the number of posts is constant and when the number of agents outside the competitive set is constant. For the maximization problem, we prove NP-hardness and W[1]-hardness when parameterized by the number of agents who are Nash-stable but not competitive.
抄録全体を表示
-
Makoto NAKAMURA, Yoshio OKAMOTO
原稿種別: PAPER
論文ID: 2026FCP0008
発行日: 2026年
[早期公開] 公開日: 2026/09/17
ジャーナル
フリー
早期公開
We study the number of vertices of an H-minor model in a graph G. We give two lower bounds on this quantity. Our bounds apply even when H is a connected graph that contains a cycle, but not only when H is a complete graph as usually studied in the literature. These lower bounds are tight; that is, we explicitly construct infinitely many pairs (G, H) of graphs such that the number of vertices of an H-minor model in G is within a constant factor of these lower bounds.
抄録全体を表示
-
Tsukasa SHIOTA, Tsukasa SAITO, Fanglu XIE, Motohiro TAKAGI
原稿種別: LETTER
論文ID: 2026EDL8016
発行日: 2026年
[早期公開] 公開日: 2026/09/11
ジャーナル
フリー
早期公開
We propose instruction-conditioned architectures that use LMs to detect and correct errors in procedural action recognition without manually implementing workflow-specific correction rules. On a first-person video dataset, our methods improved frame-wise accuracy over the baseline by more than 10 percentage points.
抄録全体を表示
-
Guidong ZHANG, Xin LIU, Donglan LIU, Honglei YAO, Yuhui JIN, Shanjie X ...
原稿種別: PAPER
論文ID: 2026EDP7102
発行日: 2026年
[早期公開] 公開日: 2026/09/11
ジャーナル
フリー
早期公開
Power load forecasting is essential for secure and economical power system operation, but the increasing penetration of renewable energy, changing user behavior, and extreme weather events make load patterns more volatile and condition-dependent. Existing spatio-temporal forecasting models often rely on fixed graph structures or fully data-driven adaptive graphs, which limits their ability to capture interpretable spatial dependency changes under different operating conditions. In this paper, we propose Topology and Condition Guided Dynamic Graph Network (TCDGNet) for multi-step power load forecasting. Its core component, the Topology and Condition Guided Dynamic Graph Generator, integrates physical or operational topology priors, the correlation graph constructed from the training set, node states, and condition variables to generate dynamic adjacency matrices. The generated graphs are further used by dynamic graph convolution, graph-constrained spatial attention, temporal attention, and a temporal decoder to model spatio-temporal dependencies. Experiments on a private power dataset and the EIA-930 public dataset show that TCDGNet consistently outperforms representative spatio-temporal graph, adaptive graph, graph Transformer, and power load forecasting baselines. Compared with the strongest external baseline, TCDGNet reduces mean absolute error by 4.8% on the private dataset and 6.0% on EIA-930, demonstrating its effectiveness and generalization ability.
抄録全体を表示
-
Yuma ITO, Yusuke SEKIKAWA, Satoshi IKEHATA, Ikuro SATO
原稿種別: PAPER
論文ID: 2026EDP7034
発行日: 2026年
[早期公開] 公開日: 2026/09/07
ジャーナル
フリー
早期公開
Large-scale deep learning models such as Vision Transformers (ViTs) have achieved strong performance across a wide range of vision tasks, making efficient inference increasingly important. Model pruning is a widely studied approach to reduce inference cost and can be broadly categorized into weight pruning and activation pruning, which has recently attracted growing attention. Recent studies have shown that intermediate activations in large-scale models exhibit substantial redundancy and sparsity, and that pruning the activations dynamically can lead to significantly small accuracy degradation. This work focuses on pruning activations in feed-forward network (FFN) modules. Among various pruning schemes, semi-structured N:M sparsity is especially appealing due to its hardware support on modern GPUs. However, existing post-training N:M pruning methods on activations implicitly assume a fixed ordering of feature dimensions. In practice, the alignment between feature dimensions and fixed N:M sparsity patterns strongly affects which activations are removed, indicating that this alignment warrants further investigation as an important design choice.
In this work, we propose Reconstruction-aware Feature Permutation Pruning (RFPP), a post-training activation pruning method that uses calibration data to determine the feature permutation so as to reduce reconstruction error under semi-structured N:M sparsity. Using calibration data, RFPP formulates the problem of finding a pruning-friendly feature dimension ordering as a minimum-cost perfect matching problem, enabling determination of feature permutations prior to inference. To ensure scalability to large models, we further introduce efficient heuristic solvers that approximate the optimal matching within practical time constraints. Extensive experiments on four image classification datasets, including ImageNet-1k, as well as the ADE20K semantic segmentation dataset demonstrate that RFPP consistently achieves smaller accuracy degradation than state-of-the-art post-training activation semi-structured pruning methods, while effectively accelerating inference on modern GPU hardware.
抄録全体を表示
-
Yuman HU, Zizheng KONG, Haozhe LV, Dongli HAN
原稿種別: PAPER
論文ID: 2025EDP7258
発行日: 2026年
[早期公開] 公開日: 2026/09/02
ジャーナル
フリー
早期公開
Stock movement prediction is a critical task in the financial sector, particularly in the highly volatile Chinese A-Share market. Traditional approaches often focus on isolated stock fundamentals and technical indicators while overlooking the dynamic interdependencies between stocks, such as industry linkages, supply chain relationships, and market-wide correlations. To address these issues, a large language model (LLM)-driven dynamic relationship network (DyLNet) is proposed in this paper. This novel framework integrates LLMs for dynamic relationship inference, temporal attention mechanisms, and time-series modeling. By dynamically constructing inter-stock relationship graphs from financial news and combining them with historical trading data, DyLNet captures temporal dependencies and evolving market structures. An empirical study conducted on the Chinese A-Share market demonstrates that DyLNet significantly outperforms traditional models in terms of prediction accuracy and investment returns, making it a robust methodology for informed stock investment decision-making.
抄録全体を表示
-
Yoshinao ISOBE
原稿種別: PAPER
論文ID: 2026EDP7012
発行日: 2026年
[早期公開] 公開日: 2026/09/02
ジャーナル
フリー
早期公開
In this paper, we aim to evaluate the stability of a classifier under weight perturbations that induce misclassification, and propose a method for estimating a probabilistically certified upper bound on the weight-perturbed generalization risk of the classifier. Our approach integrates empirical search for adversarial weight perturbations with probabilistic certification based on concentration inequalities. Here, stability is a type of generalization performance that describes a classifier's ability to produce reasonable outputs for inputs beyond the given data samples, and the term weight refers to the parameters of a classifier in a general sense. We provide theoretical guarantees for the proposed method and demonstrate through experiments that it is effective for evaluating the stability of practical deep neural classifiers.
抄録全体を表示
-
Han YU, Jianfeng LI, Shigang LI
原稿種別: PAPER
論文ID: 2026EDP7020
発行日: 2026年
[早期公開] 公開日: 2026/08/24
ジャーナル
フリー
早期公開
This paper presents a novel self-supervised framework for the online calibration of 360° spherical cameras using raw video streams. Our system comprises two back-to-back fisheye lenses, each boasting an ultra-wide 220° field of view (FoV) a hardware configuration consistent with professional-grade omnidirectional cameras such as the Ricoh THETA V. This specific configuration creates a significant 40° binocular overlap, providing redundant geometric constraints that are often underutilized in traditional pipelines. Unlike conventional methods that rely on static calibration patterns or independent lens optimization, our approach interprets calibration as a continuous spatio-temporal alignment task. By simultaneously exploiting photometric consistency (RGB) and geometric structural priors (Depth) within the expansive overlapping regions, our neural network effectively decouples intrinsic distortion from extrinsic pose errors. Experimental results demonstrate that our dual-camera joint optimization significantly outperforms independent calibration.
抄録全体を表示
-
Yongfei WU, Daisuke KATAYAMA, Tetsushi KOIDE, Toru TAMAKI, Shigeto YOS ...
原稿種別: PAPER
論文ID: 2026EDP7029
発行日: 2026年
[早期公開] 公開日: 2026/08/24
ジャーナル
フリー
早期公開
This study presents a lesion texture classification framework for Narrow Band Imaging (NBI) colorectal endoscopic images, aiming to enhance the reliability of local lesion recognition for computer-aided diagnosis. To address the challenges of small lesion regions and complex texture distributions, we propose an RoI-aware patch sampling strategy that leverages lesion mask annotations to focus data construction on diagnostically relevant regions, enabling effective local texture learning. To further improve data efficiency under limited annotated training samples, we introduce a mask-based CutMix augmentation method that preserves lesion semantics during region mixing, enhancing model robustness to heterogeneous and mixed lesion patterns commonly observed in clinical practice. In addition, we propose a progressive label representation aligned with the JNET classification criteria to explicitly model ordinal relationships among lesion categories. Combined with ordinal constraints imposed by the CORN framework, this design reduces severe misclassification errors while improving prediction consistency and interpretability. Experimental results demonstrate strong agreement between the proposed method and expert endoscopists, achieving a concordance rate of 87% and a correct diagnosis rate of 80%, approaching human-level performance (85%). These results indicate that the proposed progressive and ordinal-aware learning framework offers an effective and explainable solution for local lesion texture recognition in NBI images, with promising potential for clinical decision support.
抄録全体を表示
-
Takahiro NARUKO, Hiroaki AKUTSU
原稿種別: PAPER
論文ID: 2025EDP7207
発行日: 2026年
[早期公開] 公開日: 2026/08/19
ジャーナル
フリー
早期公開
Vision Transformer (ViT) models have made breakthroughs in image-embedding extraction and exhibit state-of-the-art performance in tasks such as zero-shot image classification. However, the models incur a high computational cost. We propose an acceleration method for ViT models called Attention-aware Token Filtering (ATF). ATF consists of two main ideas: a novel token-filtering module and token-filtering strategy. The token-filtering module is introduced between a tokenizer and transformer encoder of the ViT model without modifying or fine-tuning the transformer encoder. The module filters out tokens inputted to the encoder so that it keeps tokens in regions of specific object types dynamically and keeps tokens in regions that statically receive high attention in the transformer encoder. The filtering-strategy maintains task accuracy while filtering out tokens inputted to the transformer encoder. Evaluation results on retrieval tasks indicate that ATF provides up to a 2.8× speed-up for the ViT model SigLIP while maintaining the retrieval-recall rate.
抄録全体を表示
-
Jung Min Lim, Jong Wook Kwak
原稿種別: LETTER
論文ID: 2026EDL8026
発行日: 2026年
[早期公開] 公開日: 2026/08/19
ジャーナル
フリー
早期公開
Hybrid disk buffers combining dynamic random-access memory (DRAM) and non-volatile memory (NVM) are designed to exploit the high performance of DRAM and the high density of NVM in modern storage systems. However, existing hybrid approaches do not consider the efficiency of merge operations in the flash translation layer (FTL), leading to increased write amplification and reduced endurance. In this letter, we propose SeqSM, which stands for Sequentiality-Aware Switch Merge Acceleration, managing pages in logical block cluster units across DRAM and NVM to accelerate switch merge operations. SeqSM enables hybrid disk buffers to fully exploit their advantages in both performance and endurance, offering significant potential for both existing and emerging hybrid disk buffer architectures. By accelerating switch merge through unified cluster-level management and eviction, SeqSM exhibited a 58.0% higher hit ratio and reductions in write count and block erase count by 36.5% and 46.0%, respectively.
抄録全体を表示
-
Peiling LI, Fan YI, Wenjing XIONG, Huanchao GAO, Yue XIE
原稿種別: LETTER
論文ID: 2026EDL8039
発行日: 2026年
[早期公開] 公開日: 2026/08/19
ジャーナル
フリー
早期公開
Speech Emotion Recognition (SER) aims to automatically identify human emotional states from speech signals. The core challenge of SER lies in capturing subtle emotional details embedded in both local segments and global speech context. Existing methods often segment variable-length speech into fixed-length units with independent labels, thereby neglecting the interrelation of emotional information across segments, which limits recognition performance. To address these challenges, we propose a Multi-Scale DWT-Enhanced Transformer (MSDT) for speech emotion recognition. The model first partitions the raw speech waveform into multiple fixed-length segments. Each segment undergoes discrete wavelet transform (DWT) to extract frequency-specific components, which are subsequently processed by one-dimensional convolutional layers to reduce and unify feature dimensions. Segment-level DWT features are then aggregated into frame-level representations and fed into a Transformer to capture salient emotional details across frames. Finally, concatenated segment-level features are processed by an additional Transformer layer, followed by a softmax classifier for emotion prediction. Experimental evaluations on the EMO-DB and ABC datasets demonstrate that the proposed MSDT achieves recognition accuracies of 87.56% and 66.40%, respectively, validating the effectiveness of the proposed approach.
抄録全体を表示
-
Wan-Ting SUN, Zhe-Ming LU
原稿種別: LETTER
論文ID: 2026EDL8024
発行日: 2026年
[早期公開] 公開日: 2026/08/13
ジャーナル
フリー
早期公開
Existing image watermarking methods based on DNNs often rely on designs for fixed resolutions or pipelines with multiple stages, while recent methods based on INRs achieve flexibility across resolutions at the cost of separate optimization for each image. To address this limitation, we propose GInr-WM, a generalized implicit neural representation framework for image watermarking across resolutions. Instead of encoding each image into INR parameters that must be optimized separately, GInr-WM uses a fully convolutional encoder to extract a spatial latent map from the input image. A shared implicit predictor then predicts the RGB value of the watermarked image conditioned on the sampled latent feature, the watermark message, and the normalized pixel center coordinate. In this framework, image content is provided by latent features obtained through a forward pass, while the watermark embedding rule is learned by shared network parameters across the training set. Therefore, unseen images can be watermarked directly without gradient optimization for each image. Extensive experiments show that GInr-WM achieves strong visual quality and reliable watermark decoding over a wide range of resolutions while being substantially more efficient than prior methods based on INRs.
抄録全体を表示
-
Shinsei ENOKI, Takashi HARADA, Shogo TAKEUCHI
原稿種別: PAPER
論文ID: 2025EDP7225
発行日: 2026年
[早期公開] 公開日: 2026/08/12
ジャーナル
フリー
早期公開
The exact cover (XC) problem can be defined as follows: given a set S and family F of subsets of S, determine whether there is a subset of F that is a partition of S. XC can be used to solve assignment problems and puzzles, but practical instances often yield a huge F. The zero-suppressed binary decision diagram (ZDD) is a compressed data structure that can be used to represent F, but it cannot be used with preprocessing algorithms for XC problems without being decompressed. Thus, we propose preprocessing algorithms for finding forcing and blocking relations without decompressing ZDDs. Forcing and blocking relations are used to remove redundant elements from S and F so that an input ZDD z representing an instance of XC is converted into an output ZDD z′ containing no elements that cause forcing or blocking. Experimental results showed that our preprocessing algorithms reduced the amount of computational time required to solve XC problems in almost all instances.
抄録全体を表示
-
Yuki ARAI, Takamasa SHIMADA, Bunnoshin ISHIKAWA, Tadanori FUKAMI
原稿種別: PAPER
論文ID: 2026EDP7005
発行日: 2026年
[早期公開] 公開日: 2026/08/12
ジャーナル
フリー
早期公開
Event-related potentials (ERPs) are usually estimated by suppressing spontaneous electroencephalograms by averaging a large number of responses. Therefore, it is difficult to stably extract diagnostically useful features from the estimated waveform without collecting a large number of responses. In this study, we proposed a method for quantifying single responses and demonstrate its applicability in schizophrenia disease evaluation.
We used a deep neural network (EEGNet) to evaluate single responses to target and nontarget stimuli acquired in an auditory oddball task. Here we evaluate single responses, defining the data from young healthy subjects as the ideal response. The network was trained to output 1 and 0 for a single target and nontarget response that were correctly performed, respectively, in young healthy subjects only. We then used the output of this trained network as the evaluation value. Here, single responses from young healthy subjects, middle-aged healthy subjects, and middle-aged schizophrenic patients were input as test data, and the evaluation values among the three groups were compared.
A one-way analysis of variance with multiple comparisons revealed significant differences between the middle-aged schizophrenia patient group and the two healthy groups in the evaluation values of the target stimulus responses obtained by our method.
抄録全体を表示
-
Chihiro WATANABE, Jingyu SUN
原稿種別: PAPER
論文ID: 2026EDP7035
発行日: 2026年
[早期公開] 公開日: 2026/08/12
ジャーナル
フリー
早期公開
Large language models (LLMs) are increasingly used as agents to solve complex tasks such as question answering (QA), scientific debate, and software development. A standard evaluation procedure aggregates multiple responses from LLM agents into a single final answer, often via majority voting, and compares it against reference answers. However, this process can obscure the quality and distributional characteristics of the original responses. In this paper, we propose a novel evaluation framework based on the empirical cumulative distribution function (ECDF) of cosine similarities between generated responses and reference answers. This enables a more nuanced assessment of response quality beyond exact match metrics. To analyze the response distributions across different agent configurations, we further introduce a clustering method for ECDFs using their distances and the k-medoids algorithm. Our experiments on QA datasets of reading comprehension and mathematical problems demonstrate that ECDFs can distinguish between agent settings with similar final accuracies but different quality distributions. The clustering analysis also reveals interpretable group structures in the responses, offering insights into the impact of temperature, persona, bandwagon-effect involvement percentage, and question topics.
抄録全体を表示
-
Yuxuan WU, Hideki NAKAYAMA
原稿種別: PAPER
論文ID: 2025EDP7268
発行日: 2026年
[早期公開] 公開日: 2026/08/06
ジャーナル
フリー
早期公開
In recent years, neuro-symbolic methods have become a popular and powerful approach that augments artificial intelligence systems with the capability to perform abstract, logical, and quantitative deductions with enhanced precision and controllability. Recent studies successfully performed symbolic reasoning by leveraging various machine learning models to explicitly or implicitly predict intermediate labels that provide symbolic instructions. However, these intermediate labels are not always prepared for every task as a part of the training data, and pre-trained models, represented by Large Language Models (LLMs), also do not consistently generate valid symbolic instructions with their intrinsic knowledge. On the other hand, existing work developed alternative learning techniques that allow the learning system to autonomously uncover optimal symbolic instructions. Nevertheless, their performance also exhibits limitations when faced with relatively huge search spaces or more challenging reasoning problems. In view of this, in this work, we put forward an advanced practice for neuro-symbolic reasoning systems to explore the intermediate labels with weak supervision from problem inputs and final outputs. Our experiments on the Mathematics dataset illustrated the effectiveness of our proposals from multiple aspects.
抄録全体を表示
-
Ryoko ARITA, Joonyong PARK, Wataru NAKATA, Yuki SAITO, Hiroshi SARUWAT ...
原稿種別: LETTER
論文ID: 2026EDL8011
発行日: 2026年
[早期公開] 公開日: 2026/08/06
ジャーナル
フリー
早期公開
We propose SymphoMOS, a singing MOS prediction system that incorporates cross-domain self-supervised learning (SSL) fusion using speech and music SSL models for feature extraction. We instantiate our SymphoMOS using two representative models, wav2vec 2.0 for speech and MERT for music. The results demonstrate that SymphoMOS achieves the state-of-the-art system-level Spearman's rank correlation coefficient for singing MOS prediction on the SingMOS dataset.
抄録全体を表示
-
Ruiqi ZHANG, Naoko MISAWA, Chihiro MATSUI, Ken TAKEUCHI
原稿種別: PAPER
論文ID: 2026EDP7013
発行日: 2026年
[早期公開] 公開日: 2026/08/06
ジャーナル
フリー
早期公開
Deploying large Transformer models on resource-constrained Computation-in-Memory (CiM) edge devices is challenging due to limited memory and hardware non-idealities. This work proposes Quantization-Aware and Error-Tolerant Adapter Freezing (QEAF), which uses a dual-gradient criterion to freeze low-impact LoRA adapters and applies selective error exposure with scheduled noise injection to improve hardware robustness. Experiments on BERTBASE and Llama 3.2 show that QEAF achieves the best INT4 performance among LoRA-based baselines with substantially fewer trainable parameters. QEAF nearly matches full fine-tuning on BERTBASE with only 0.49% (0.53M) trainable parameters, and obtains the highest average benchmark scores on Llama 3.2 1B and 3B. It also reduces optimizer memory and robustness cost while maintaining stronger performance under severe device variation, demonstrating its effectiveness for reliable low-bit CiM deployment.
抄録全体を表示
-
Nina NAKAJIMA, Takatomi KUBO, Kazushi IKEDA, Kazunori OHNO
原稿種別: LETTER
論文ID: 2026DRL0003
発行日: 2026年
[早期公開] 公開日: 2026/08/05
ジャーナル
フリー
早期公開
Cats exhibit highly flexible body motions that challenge accurate pose tracking. We propose a refinement framework integrating DeepLabCut with TAPIR-based temporal tracking and SAM-based region filtering. Experiments on 14 videos show F1 score improvements of 0.08-0.13 for flexible motions compared to DLC, demonstrating improved robustness to deformation and occlusion.
抄録全体を表示