TAAI 2026 議程總覽

Dynamic Economic Emission Dispatch (DEED) is a challenging constrained multiobjective optimization problem that seeks to simultaneously minimize power generation cost and pollutant emissions. In this paper, we propose EAGAMES-MOEA/D, an enhanced decomposition-based multiobjective evolutionary algorithm featuring an adaptive search framework with two complementary selection mechanisms. The proposed adaptive mating selection mechanism allocates computational resources by jointly considering convergence and diversity, while the adaptive environmental selection mechanism dynamically switches between weighted sum and Tchebycheff scalarizing functions according to the distribution of the current population. The proposed algorithm is evaluated on six representative DEED benchmark problems ranging from 5 to 40 generating units and compared with 26 existing algorithms. Experimental results demonstrate that EAGAMES-MOEA/D consistently produces well-converged and well-distributed approximation fronts under competitive computational budgets.

This paper investigates graph neural network-based reinforcement learning for deterministic job-shop scheduling (JSSP), aiming to minimize makespan. Each instance is represented as a disjunctive graph using a five-feature operation representation combining scheduling status and workload information. A two-layer graph convolutional network (GCN) with proximal policy optimization (PPO) is used to compare two policy architectures under this common representation. GCN-RL1 employs a mean-pooled graph embedding with a fixed-dimensional policy head, whereas GCN-RL2 uses candidate-wise scoring to support variable-size action sets. On 100 unseen random instances per problem size, GCN-RL1 achieves average makespans of 558.88 and 982.77 on $6\times6$ and $10\times10$ JSSPs, respectively. It also obtains the lowest aggregate relative scheduling error (RSE) of 17.5\% among the evaluated methods on selected Ft, Orb, and La benchmarks. GCN-RL2 trained on $15\times15$ instances transfers directly to Taillard instances up to $100\times20$, achieving an RSE of 13.3\% on the $100\times20$ group. These results illustrate differences in same-size scheduling performance and cross-size applicability. The implications of non-delay scheduling are also discussed.

Standard reinforcement-learning formulations generally assume that a selected action affects the environment immediately. In practical control systems, however, actions may be executed only after a fixed delay, making it difficult to associate an action with its subsequent effects. State augmentation addresses this problem by appending the pending action sequence to the physical state, while $n$-step returns incorporate multiple sampled rewards before bootstrapping from an action-value estimate. This study empirically compares backup lengths in state-augmented Double Deep Q-Network (DDQN) under fixed action delays using CartPole-v1 and independent training seeds. The results show that backup length affects evaluation performance, learning speed, and consistency across training seeds. Larger tested backup lengths often achieve higher mean evaluation returns over a wider range of delays, although performance does not improve monotonically with $n$. All tested settings yield low returns under very long delays. Matching the backup length to the earliest reward that can reflect the selected action does not consistently yield the highest mean evaluation return among the tested settings. These findings suggest that backup length should be considered alongside action delay.

Our research group has been developing a Mental-Physical States Estimation System that estimates the degradation of the respective users' mental-physical states and then supports the system users' healthy daily lives continuously by promoting awareness and behavioral improvements. To achieve the above goal, this paper aims to investigate the relationships among three kinds of data regarding participants' biometric, weather, and activity log(life-log) information collected through experiments, and to clarify valuable features available for the prediction of their mental-physical states.

Experiments are conducted to evaluate relationships among participants' mental-physical states, biometric information, activity logs, and environmental information. In the experiments, participants were instructed to wear a wearable device (watch-shaped activity tracker) to record their biometric data and input their activity logs into the activity-log application we developed in this research. Experimental results showed relationships between sleep patterns and mental-physical load, between weather information and the load, and so on. However, further investigations are needed to achieve mental-physical state estimation.

In the era of information overload, recommendation systems play a critical role by efficiently guiding users toward desired information or products. Although personalized recommendations often enhance user experiences, excessive personalization can result in user discomfort and negative reactions. To address these concerns, this paper proposes a Graphically Illustrated Recommendation System that integrates AI-generated recommendations, image generation technologies, and psychological and behavioral theories. In addition, the system also employs culturally intuitive interactive elements—including selecting traditional Japanese charms (Omamori) and comic-style character illustrations—to implicitly detect user preferences and individual status in a comfortable and engaging manner. Finally, the recommendation content, which includes strategically embedded recommended items, is seamlessly integrated into visually appealing tarot-style images and positively oriented tarot-themed textual interpretations, both generated through generative AI tools including ChatGPT and Stable Diffusion. Experimental results using a virtual E-Mall with 12 participants showed that Hit Rate increased from 33.3% to 75.0% by experiencing interaction with the proposed system. In addition, their emotional responses after interacting with the system indicated that participants accepted the recommended items while showing generally positive emotional responses toward the recommendation process. By aligning with the FAD principles of Shikakeology—ensuring fairness through user-controlled interactions, enhancing attractiveness via culturally familiar and gamified elements, and embedding dual purposes of entertainment and recommendation—the proposed system offers an user-centered design that balances personalization with psychological comfort and cultural resonance.

This paper proposes a method for generating synthetic playlist dataset for context-aware session-based music recommendation. In music recommendation, users' preferences may vary depending on contexts such as activity, mood, and season, but constructing large-scale sequential dataset with contextual information is difficult. To address this problem, the proposed method extracts context-related playlists from the Spotify Playlists Dataset based on playlist names. Using the collected data as a seed, synthetic playlists dataset are generated considering context-specific track distributions. The quality of the synthetic dataset generated from Spotify playlist dataset is evaluated by training recommendation models with three conditions: training with real playlist dataset associated with target contexts (Real matched), training with context-based synthetic dataset generated by the proposed method (ContextSynth), and randomly generated dataset (Random). GRU4Rec and SASRec are trained on each condition. Recommendation performance is evaluated on test data from Spotify. Experimental results show that ContextSynth consistently outperforms Random, although it does not reach the performance of Real matched. These results suggest that incorporating contextual information into synthetic playlist dataset is effective for constructing training data for session-based music recommendation.

Explainable artificial intelligence (XAI) methods are increasingly used to derive feature-importance rankings, yet a feature that explains a trained model well is not necessarily the most useful one once the model is retrained on a reduced input set. This study evaluates XAI rankings from a performance-based feature-selection perspective. Six methods---SHAP, LIME, PDP, ALE, permutation importance (PI), and tree-based feature importance (FI)---are compared on WDBC, WPBC, and Spambase with XGBoost and LightGBM: within each cross-validation fold, rankings are generated from the training partition only, top-$k$ subsets are retrained with five seeds, and AUC is measured on the held-out fold. MIP reranking, rank aggregation, and the ANOVA and mutual information (MI) filters serve as comparisons. Across the 24 settings (three datasets, two models, and four top-$k$ values for each dataset), no XAI method is universally best: FI and PI are most often the strongest ranking, but each leads in fewer than a third of the settings. The choice of method matters mainly when very few features are kept, except on the small WPBC dataset. At larger top-$k$ the six rankings yield similar AUC and, with one exception, exceed the filter baselines; at the smallest top-$k$, ANOVA and MI are comparable to at least half of them. MIP yields no consistent gain, and simple rank aggregation is comparable to the best single ranking on two of the three datasets. These results support evaluating explanation-derived rankings through downstream retraining rather than treating explanation importance as a proxy for feature-selection quality.

Large language models (LLMs) can support spoken language practice, but it is unclear how corrections should be displayed during a conversation. We built a Japanese speaking-practice system that recognizes learner speech, checks grammar, vocabulary, and politeness, generates a corrected sentence, and replies as a conversation partner. Twelve learners used three display conditions: no text (C1), subtitles (C2), and subtitles plus panel (C3). The partner's voice is played in every condition. C2 adds text to C1: the partner's replies, and each learner turn as two lines, the kana the system heard and the corrected sentence, without marking what was changed. C3 adds to C2 a feedback panel that shows error details and repeats the corrected sentence. Each learner completed the same scenario twice under each condition. We checked whether errors flagged in the first conversation were repaired in the second. Repair rates were 74% under C1, 45% under C2, and 67% under C3. Words from the corrected sentences reappeared most often under C3, and 10 of the 12 learners ranked C3 first. Learners rated C2 positively despite its lowest repair rate. With 12 learners, most differences between conditions were not clear; the clearest was the lower repair rate under C2 than under C1. These exploratory results suggest that showing the corrected sentence without marking what was changed may not help learners use the correction.

We study rank-conditioned next-move prediction on the ten skill levels of the AI CUP 2026 Go training set (A–D = 1–3, 4–6, 7–9 and 10–12 kyu; 1d–6d). The nested-residual network of an open-source AICUP 2023 entry is extended from 17 to 30 input planes with ladder, 10-way rank and move-number features, initialized from its pretrained dan model in a function-preserving way, and fine-tuned on 10.6M positions. On 2.21M test positions from players unseen in training, the rank-aware model reaches 55.60% top-1 and 86.33% top-5 accuracy, 0.22 points above an otherwise identical control without rank planes (95% CI [0.19, 0.24]). No level gets worse, and the gain is concentrated at both ends of the rank axis. A counterfactual rank-swap test shows that, although one-hot labels carry no order, the model recovers the true order A > B > C > D.

本研究探討閘控路徑規劃網路（GPPN）應用於無人機局部路徑規劃的可行性。研究使用八方向 Dijkstra 演算法產生教師價值圖，以監督式學習訓練 GPPN，並於 10 張未參與訓練的 BEV 地圖上測試 3,000 組地圖與目標組合。實驗結果顯示，預測價值圖與教師價值圖的平均皮爾森相關係數為0.9147，任務成功率與路徑合法率皆達 100%，平均路徑成本比為 1.0001。GPPN 整體規劃時間為 213.99 毫秒，Dijkstra 則為 133.08 毫秒。結果顯示，GPPN 雖未具速度優勢，仍能在未見地圖中產生接近最佳解的有效路徑。

In this paper, we have used a modular driving system in CARLA for lane selection and braking. Dueling Double DQN (D3QN) selects lanes using 13 traffic features, a deterministic executor carries out the decision, and a separate controller handles braking. Recovery can reconsider an unfinished lane change after braking. We have evaluated navigation, braking, and recovery separately. In 200 Town05 navigation cases, our system achieved 200 contact-free arrivals, compared with 13 for a Safe Navigation-based implementation. In 200 Town06 static braking cases at 20–90 km/h, the extended-speed model achieved 200 safe stops, compared with 99 for the original model. In 200 paired repetitions of one fixed blockage configuration, recovery enabled resulted in 200 arrivals, while recovery disabled resulted in none. The framework completed the tested tasks and handled the stopped-maneuver condition, although the comparisons did not isolate the route learner.

Keywords: autonomous driving, Dueling Double DQN, CARLA, lane selection, braking arbitration, recovery

重大建設與住宅價格之關聯可能隨空間距離及發展階段而變化。本研究以新北市住宅交易為基礎，依行政區、建物型態及坪數建立歷史價格基準，以相對溢價率為預測目標，結合捷運距離衰減、商場 Push-Pull 及建設生命週期特徵建立 CatBoost 模型。2024-2025 年共 30,007 筆時間外測試結果顯示，房價 RMSE 為 6.8989 萬元/坪、R²為 0.7857，較僅使用住宅屬性之模型降低 13.87% RMSE。SHAP 分析顯示，捷運距離貢獻呈遞減趨勢，商場則呈現非單調駝峰型態；生命週期特徵僅提供小幅預測增益。研究結果反映模型預測關聯，不代表直接因果效果。

Buddhist stories contain rich wisdom and educational value, but their abstract language can make them difficult to understand. This study uses generative AI to transform Buddhist stories into animations. ChatGPT handles text segmentation, character features, and prompt generation, while Qwen maintains character consistency and Wan 2.2 converts images into animations. The results show that generative AI can support the animation of Buddhist stories.

Enterprise AI agents require evidence that distinguishes approval from execution and observed effects. We propose a process-to-execution governance continuity model with five non-equivalent stages: Decision, Authorization, Dispatch, Runtime, and Outcome. Governance Event Core (GEC) v0.1 supplies an implementation-neutral event representation and semantic qualification contract. An OpenCode vocabulary case classifies seven claims as three PROVEN, three CONTRADICTED, and one NOT_OBSERVED. Paired isolated executions distinguish matching from mismatched request authorization. Post-submission fixtures address context drift, observation completeness, and five-stage control flow while preserving rejected mappings and post-effect failures. Seven workflow classifications replay consistently, but only five meet all original expectations. The contribution is bounded claim qualification over governance evidence, not a production authorization or enforcement system. The study does not establish cross-platform interoperability, scalability, or general agent safety.

This paper presents a fully on-device method for citrus pest and disease recognition with grounded diagnosis, combining lightweight object detection, retrieval-augmented generation (RAG), a quantized small language model, and a safety guardrail. On a nine-class Ministry of Agriculture dataset (7,264/401/401 training/validation/test images), the detector achieves a test mAP50 of 0.861 and an mAP50–95 of 0.641; the eight classes unaffected by an annotation-rule change reach a pooled mean AP50 of 0.8508. With 33 atomic knowledge chunks and adaptive N-gram retrieval, a 0.5B-parameter model reaches 85.05% Faithfulness, 100.0% Context Precision, and 93.06% Context Recall on a 40-question benchmark and defends against all four boundary questions. A five-stage ablation shows that atomic chunking and the guardrail contribute most. The mean latency per question on a Samsung Galaxy A33 is 29.98 s. Reliable on-device agricultural AI depends on annotation consistency, generalization, retrieval evidence, and safety boundaries rather than a single accuracy score.

失眠已成為現代社會普遍的健康問題，長期藥物治療伴隨依賴性與副作用等疑慮。雙耳節拍（Binaural Beats, BB）為一種非侵入性介入手段，可改善睡眠品質，然而其神經夾帶效應強弱因人而異，為既有研究尚未解決之缺口。本研究提出個人化雙耳節拍音樂誘導睡眠系統，透過分析腦電圖自動篩選對使用者有效之BB刺激組合，並結合自動睡眠分期即時判定誘導效果是否持續，若失效則自動切換下一組合，系統包含BB自動篩選、睡眠分期、入睡品質評估三部分。實驗招募3位受試者，結果顯示有效BB組合具明顯個體差異，其中兩組合對所有受試者皆有效，另五組合僅對個別受試者有效；睡眠分期模型測試準確率達0.89以上。受試者聆聽個人化BB音樂後入睡時間較雙和醫院資料集為短。本研究初步驗證個人化策略可提升睡眠誘導效果，未來將擴大樣本進一步驗證。 Insomnia has become a prevalent health issue in modern society, and long-term pharmacological treatment raises concerns such as dependency and side effects. Binaural beats (BB) represent a non-invasive intervention that can improve sleep quality; however, the strength of their neural entrainment effect varies across individuals, a gap that existing research has yet to address. This study proposes a personalized binaural-beat-induced sleep system that automatically screens BB stimulus combinations effective for each user by analyzing electroencephalography (EEG) signals, and combines this with real-time automatic sleep staging to determine whether the induction effect persists—automatically switching to the next combination if it fails. The system consists of three components: automatic BB screening, sleep staging, and sleep onset time (SOT) assessment. Three subjects were recruited for the experiment, and the results showed marked individual differences in effective BB combinations: two combinations were effective for all subjects, while five combinations were effective for only specific individuals. The sleep staging model achieved a test accuracy above 0.89. Subjects exhibited a shorter sleep onset time (SOT) after listening to personalized BB music compared to sleep onset times recorded in a dataset from Shuang Ho Hospital. This study provides preliminary validation that a personalized strategy can enhance sleep induction effectiveness; future work will expand the sample size for further validation.

Convolutional neural networks commonly reduce spatial resolution at the front end, but the effect of aggressive early downsampling on radiographic grading of mandibular third molar surgical difficulty remains unclear. We investigated this effect using 682 panoramic radiograph samples. Under the same patient-level split and training protocol, two opposite-direction front-end interventions were evaluated: the aggressive ResNet-style stem was replaced by a milder staged design in ResNet50-GN, while the same aggressive design was introduced into VGG16. Prediction concentration was quantified using the majority prediction rate (MPR), where MPR = 1.0 denotes complete single-class collapse. In the ResNet pair, milder downsampling reduced Angle MPR from 1.0000 to 0.8472 and Ramus MPR from 0.8472 to 0.7917. In the reverse VGG intervention, aggressive downsampling increased Angle MPR from 0.7222 to 1.0000 and Ramus MPR from 0.7778 to 1.0000, resulting in complete single-class predictions, whereas Depth showed a different response. These paired results indicate that aggressive early downsampling can substantially alter prediction concentration in mandibular third molar radiographic grading and that its effect is task dependent, highlighting front-end spatial compression as an important architectural consideration.

健檢資料屬高度敏感之個人隱私，於雲端進行心臟病風險預測時，明文外洩風險極高。本系統提出一套以全同態加密（FHE）為核心之隱私保護心臟病風險預測系統，採用 TenSEAL 函式庫實作CKKS 加密方案，使伺服器於不接觸明文之情形下完成羅吉斯迴歸心臟病風險推論。本 研 究以Framingham Heart Study 資料集訓練，選用八項臨床特徵，測試集 AUC 達 0.7371，準確率 82.79%，惟因類別不平衡不宜作為主要目標。鑑於標準 FHE僅保證機密性，本研究進一步整合 HMAC-SHA256完整性驗證與密文域可逆資訊隱藏技術，使系統同時具備機密性、完整性與可追溯性，實驗證實三面向皆具可行性與有效性。

本研究探討Wasserstein生成對抗網路（WGAN）資料增強應用於腦電圖（EEG）神經退化性疾病診斷之效果，並系統性比較常見的機器學習與深度學習模型在資料增強前後的分類效能。實驗採用公開EEG資料集（Dataset B），涵蓋阿茲海默症（AD）、輕度認知障礙（MCI）及正常認知狀態（CN）三類，原始資料存在嚴重類別不平衡問題（AD:MCI:CN≈33:4:2）。本研究以Progressive WGAN with Gradient Penalty生成合成EEG訊號，將各類別訓練資料補足至4,000筆，並評估RF、SVM、1D CNN、2D CNN及ViT五種分類模型在資料增強前後的表現。實驗結果顯示，WGAN生成資料在有效頻段內與真實EEG訊號具有高度吻合性；SVM與1D CNN改善最顯著，2D CNN與ViT則效能略有下滑或持平。整體而言，WGAN對因類別不平衡受限的模型具有顯著修復效果，但對效能已飽和的模型效益有限，MCI始終是最難辨識的類別。

企業資源規劃系統(Enterprise Resource Planning, ERP)中之 Agent 任務多以正常流程為評估情境，對於交易遭既有後端拒絕後之恢復過程仍缺乏觀察證據證明。為探討 LLM Agent 在 ERP 既有交易約束下之失敗恢復能力，本研究將單一 Agent 接入商用 ERP 系統，並以請購單、採購單與收貨單之採購流程作為研究情境；並進一步設計輸入與參數、商業規則、狀態依賴、跨文件一致性，以及工具與系統五種受控異常情境，每種異常情境進行三次驗證，共有 15 筆觀察。當異常發生後，Agent 僅依實際系統回饋自主查詢、修正與重新提交交易，且系統同步保存公開回應、工具操作、參數與後端回傳結果，並於實驗結束後回讀交易文件及其關聯，以驗證恢復結果。本研究結果發現，於15 筆觀察中，其中有 14 筆解除原始錯誤並完成採購流程，唯一未完成者，雖辨識出過帳日期問題並提出修正方向，卻未實際執行收貨。本研究結果顯示，在保留 ERP 商業規則與交易控制之條件下，Agent 可於多類異常後完成恢復，但問題辨識、處理決策與實際交易完成並不相同。因此，於評估企業流程中之 Agent 恢復能力時，除最終結果外，亦需整合可追溯之執行軌跡與後端狀態驗證。

In this work, we designed a learning framework using liquid time constant neural networks (LTC-RNNs), a brain-inspired class of continuous-time neural models to perform vision-based obstacle avoidance. The learning framework is based on training data generated by the decision tree as control policy. The vision-based obstacle avoidance is achieved with monocular depth estimation using optical flow with the image streaming from the nano quadcopter, Bitcraze Crazyflie 2.1 (with a weight of 34 grams). The usage of genetic algorithm successfully identifies optimized LTC-RNN with interpretable decision making network motif. The LTC-RNN generalized well to previously unseen environments with drastic scenery changes.

As privacy regulations increasingly require the removal of specific training data from deployed models, machine unlearning has become a crucial model maintenance mechanism for privacy-aware intelligent systems. This challenge is particularly significant for Vision Transformers (ViTs), whose non-local self-attention and distributed representations make selective forgetting difficult without full retraining. This work studies machine unlearning in ViTs through a weight-resetting unlearning framework. Specifically, we compare three resetting strategies: magnitude-based reset (as an unstructured pruning baseline), gradient-based reset, and attention head reset. We formally decouple the conceptual multi-objective unlearning goal from the practical two-stage resetting-and-recovery pipeline, and formulate parameter importance at an element-wise scalar resolution. Experiments are conducted on CIFAR-100 under the forget-10 setting alongside benchmark baselines and efficiency profiling. The results demonstrate that while all three resetting methods achieve near-complete forgetting after post-fine-tuning, gradient-based reset achieves the most favorable balance between forgetting effectiveness, retain accuracy recovery, and reference model alignment, yielding an ≈ 8.3× speedup over full retraining. Furthermore, intervention scope analysis indicates that targeted parameter resetting in late-stage transformer blocks is sufficient to induce zero-retrain forgetting, offering practical insights for sustainable and privacy-compliant vision deployments.

This paper proposes a learning model for classifying the presence or absence of fouls during basketball layup shots using multi-view still images. In the proposed method, still images obtained at the same time from six cameras positioned at different viewpoints are treated as a single input set, and the presence or absence of fouls is classified by fusing the features extracted from each image. To verify the effectiveness of the proposed method, its performance was compared with that of a conventional method using single-view images. The results showed that the proposed method outperformed the conventional method on all evaluation metrics: Accuracy, Recall, Precision, Specificity, and F1 Score. These results demonstrate that fusing and utilizing multi-view information is effective for improving foul classification performance during layup shots.

While most video-camera-based gaze estimation methods target a single user, few multi-user studies have investigated per-user per-sonalized facial-landmark models or quantified how anatomical feature partitioning and screen-region granularity affect estimation performance. This study proposes a video-camera-based system that simultaneously estimates the gaze points of two users sharing a screen. The proposed system applies principal component analysis (PCA) to 3D facial landmarks extracted from video images to construct compact facial features. A non-person-specific model (base model) is built before applying each user’s 9-point personal calibration, and the gaze estimation results of multiple users are output simultaneously. We evaluated the proposed system with 18 participants across evaluation screens divided into 9 regions (evaluating accuracy at calibrated positions) and 16 and 25 regions (evaluating spatial generalization to unseen positions). The method separating the eye region yielded the highest sample mean gaze-region accuracies of 95.06%, 88.12%, and 78.44% under the single-user condition and 91.98%, 77.71%, and 64.00% under the two-user condition, respectively.

Persistent homology is effective for analyzing the structure of handwritten characters, which often exhibit diverse forms. In our previous research, we proposed elliptical filtration and CHD filtration as methods for calculating persistent homology to accurately extract the invariant structure of handwritten characters. However, the persistent intervals extracted using these methods represent the lifespan of each hole or cluster of black pixels. Consequently, if components differ in size or position but have the same number, these features can potentially lead to misrecognition. In this study, we define the persistent property by incorporating information on the position and size of holes or clusters of black pixels into the conventional persistent interval. We prove that this information is contained within specific basis vectors of the representation matrix of the first boundary homomorphism in persistent homology. We also propose an algorithm to extract this information. The proposed persistent property better represents invariant features of characters, and we found that these features improve recognition accuracy.

Anime production archives group intermediate assets by production cut; individual files may provide incomplete visual evidence. This study examines how to represent cut-aligned scripts on the index side of a natural-language retrieval system. With a fixed multilingual encoder and ranking procedure, we compare nine conditions based on raw script text, neighboring context, structured metadata, retrieval-oriented rewriting, text concatenation, and score fusion. The evaluation covers Little Witch Academia (LWA1) and Little Witch Academia: The Enchanted Parade (LWA2), with 373 and 269 indexed production-cut units and 53,620 asset files. We report mean pooled nDCG@10 over 23, 12, and 8 queries for the two work-specific settings and the combined corpus. Several representations yielded numerical gains over raw script text, but none of the nDCG@10 comparisons with the baseline reached statistical significance after Holm correction.

Instruction-tuned large language models have shown strong instruction-following and reasoning capabilities, but their deployment remains costly due to high computational and memory requirements. To improve the practicality of smaller language models, this work investigates parameter-efficient knowledge distillation for transferring instruction-following capability from an 8B teacher model to a 3B student model. Specifically, we investigate a training framework that combines supervised fine-tuning, logits-based knowledge distillation, and DoRA-based parameter-efficient adaptation. To support training, we construct a 20k-sample instruction-style corpus, named Curated Instruction Mixture 20K (CIM-20K), by combining and filtering samples from WizardLM, Magpie-Pro, and ShareGPT. We refer to the resulting distilled 3B student as DoRA-KD-3B and evaluate it against Llama-3.2-3B, DoRA-3B, and LoRA-KD-3B, together with Llama-3.2-3B-Instruct and the Llama-3.1-8B-Instruct teacher as instruction-tuned reference models. Across three random seeds, DoRA-KD-3B achieves mean scores of 3.787 on MT-Bench, 0.474 on ARC-C, 0.293 on GSM8K, 0.552 on MMLU, and 0.456 on TruthfulQA, exceeding the base Llama-3.2-3B model on all five aggregate benchmarks. Among the trained non-instruction-tuned 3B models, DoRA-KD-3B achieves the highest mean scores on MT-Bench, ARC-C, MMLU, and TruthfulQA, while DoRA-3B achieves the highest mean score on GSM8K. Overall, the results suggest that logits-based distillation can complement DoRA-based adaptation, although its benefit depends on the evaluation task.

Software-Defined Networking (SDN) centralizes control-plane intelligence, exposing it to both volumetric DDoS floods and stealthy, low-rate poisoning attacks. Existing defenses address these threats in isolation---volume-based classifiers miss covert low-rate patterns, while purely entropy-based methods suffer elevated false positives---and both lack the semantic contextual awareness needed to reliably distinguish benign fluctuations from adversarial behavior. To address these limitations, we propose GuardianNet, an LLM-assisted multi-agent framework deployed within a Mininet and Ryu SDN testbed. Our system comprises a hierarchical three-agent architecture: (1) a Statistical Agent employing a K-Nearest Neighbor (KNN) classifier for instantaneous volumetric screening; (2) an Entropy Agent that computes Shannon Entropy over the packet-delta sequence of each flow to detect the mechanical regularity characteristic of low-rate injection; and (3) a Reasoning Supervisor powered by a locally deployed open-source Large Language Model (Llama-3.1). By converting raw telemetry and entropy signals into structured natural-language contexts, the LLM supervisor enables reasoning over attack semantics that numerical classifiers cannot capture. Compared against four established baselines---KNN, Random Forest, LSTM, and an Entropy Only detector---GuardianNet attains 90.0\% accuracy with a 0.0\% FPR. The ablation study across six agent configurations shows that rule-based fusion of the two statistical agents reaches 68.1\% accuracy with a 68.2\% false positive rate (FPR), whereas adding the LLM supervisor raises accuracy to 83.2\% and eliminates false positives entirely.

Learned atmospheric compressors are usually assessed by aggregate distortion, which does not reveal how much residual error remains predictable from decoded fields. We apply a fixed 1.65\,M-parameter U-Net to CRA5 reconstructions over CONUS without adding per-frame bits. On the full 2018 test year, mean normalised RMSE decreases by 25.8\% over 19 directly represented FourCastNet variables (95\% temporal block-bootstrap interval [25.32, 26.18]\%). On the complete 2022 year outside CRA5's published training window, the same frozen checkpoint and normalisation give a 26.4\% reduction [25.90, 26.93]\%, without fine-tuning. Near-surface versus upper-air gains are 32.2\% versus 3.5\% in 2018 and 32.8\% versus 3.6\% in 2022. A mean-rate-matched MPCA-only control has a smaller near/upper gain ratio, while additional MPCA+PCA controls are explicitly non-matched. A parameter-matched dilation ablation retains a roughly eightfold split and concentrated gradient sensitivity. Using ERA5 orography and land--sea masks, western-land correction gains increase from 18.1\% to 36.9\% between the lowest and highest terrain-roughness deciles, while corrected errors remain larger on rough ground. Elevation-stratified comparisons support an association but do not isolate a terrain mechanism. These results measure model-relative predictability; terrain causality, corrector architecture effects, and broader distribution-shift robustness remain open. Shared weights and inference add deployment costs.

This work presents a preliminary weed stem candidate localization method for RealSense-assisted laser weeding. Instead of using full polygon segmentation masks, the proposed approach uses point-level annotations placed near visible weed stem centers. The annotated points are converted into Gaussian heatmaps and used to fine-tune a U-Net-based model for stem candidate prediction. A manually prepared dataset containing 114 cropped field images and 974 weed-stem point labels was used. The data were split into 79 training images, 17 validation images, and 18 test images. On the 18 test images, an exploratory threshold of 0.85 produced 137 matched points from 155 ground-truth stem points, with 25 false positives and 18 missed points. The precision, recall, and F1-score were 0.8457, 0.8839, and 0.8644, respectively, with a mean localization error of 3.7629 pixels on a 512-pixel scale. Preliminary RealSense RGB camera experiments using printed weed images demonstrated live stem-candidate localization. However, false detections on crop or non-target plants remain a limitation. Future work will focus on crop-weed filtering, height-aware RealSense calibration, and laser shooting efficiency tests on weeds with different heights.

FreshKG is evaluated as a mission-level supervisor for UAV routes whose corridor permissions can age, arrive late, or conflict across sources. Experiments in gym-pybullet-drones and safe-control-gym comprise 30,000 mission episodes and 2,000 fixed-reference tests, with 100 paired seeds per condition. With the safe-control-gym PID controller, FreshKG safely completes 464/600 missions, compared with 107/600 for LatestArrival; the paired improvement is 59.5 percentage points (95% cluster-bootstrap CI, 57.7-61.2). An equal-information table implementation and the graph implementation make identical supervisory decisions, indicating that the gain arises from the information rules rather than graph representation alone. We repeat the mission study with PID, LQR, MPC, and six fixed PPO/SAC residual checkpoints, and use separate fixed-reference trials to characterize tracking and attitude behavior. Ablations isolate observation order, expiry, and source filtering, while recorded failures show where replanning is insufficient. The evidence is limited to simulation and supports a mission-supervision claim, not a new flight controller or formal safety guarantee.

This paper presents a skeleton-based Taiwanese Sign Language (TSL) recognition and sentence reconstruction system. MediaPipe Holistic landmarks are normalized, cropped by wrist motion, and resampled to 30 frames. A task-specific hierarchy routes inputs to five motion families and activates a specialist whose features match the observed motion structure. On a video-level 80/20 split, the 124-dimensional Uni-LSTM with temporal attention achieves 99.42% gatekeeper accuracy, while specialist accuracies are 81.40%, 100.00%, 92.40%, 91.60%, and 95.80% on 79 labels and 398 videos. These are component results; the approximately 91.7% cascade value is an independence-based estimate rather than a measured end-to-end accuracy. A local language reconstruction module uses seven TSL-oriented few-shot patterns and a three-turn LangChain memory to reconstruct fragmented lexical predictions into Chinese sentences. The language results are prototype-scale qualitative evidence, and signer-independent testing, controlled baselines, sentencelevel metrics, human evaluation, and full-pipeline latency remain future work. All data processing and document preparation for this study were performed locally.

Real-time visual surveillance on unmanned aerial vehicles (UAVs) is severely constrained by onboard battery storage and thermal dissipation limits, rendering continuous 30 FPS deep neural network (DNN) forward passes unsustainable. Conventional background subtraction fails catastrophically in aerial video due to 6-DoF camera ego-motion. In this paper, we propose an Ego-Motion Compensated Adaptive Spatio-Temporal Difference (ASTD) gating architecture for energy-efficient edge detection. ASTD dynamically neutralizes drone flight motion via real-time RANSAC homography alignment and scale-invariant spatial filtering, serving as an intelligent compute gate that awakens downstream DNN detectors only upon genuine dynamic events. Evaluated on a continuous 1,821-frame 1080P real-world construction flight sequence, ASTD successfully skips 97.91% of redundant background frames at an ultra-low edge latency of 22.04 ms (45.37 FPS). ASTD slashes weighted per-frame inference latency by 91.0% (serving as an effective proxy for edge computational energy reduction) while achieving 100% event recall across all active machinery, demonstrating exceptional power efficiency for persistent autonomous UAV surveillance.

現有將 VLM 應用於工業異常檢測的研究多偏重邏輯性異常，較少顧及結構性異常。本研究透過資料庫式異常檢測與 VLM 的結合來完善兩類異常的偵測平衡：PatchCore 將區域異常分數以文字提供給 VLM，VLM 回答邏輯性與結構性問題集後評分，最後整合兩者的分數判斷是否異常。我們發現僅用邏輯性問題集時，VLM 在結構異常上與 PatchCore 的貼合度不足，因此加入了結構性問題集讓 VLM 回答更多結構性異常原因；而實驗中也發現 VLM 擁有最終判定權時會否決 PatchCore 的資訊，因此改採等權重分數融合。本架構不需訓練模型參數，在 MVTec LOCO-AD 上達到 logical F1-max 0.870、logical AUROC 0.916、structural AUROC 0.917 與 overall AUROC 0.915，並能輸出自然語言的異常原因說明。

Existing VLM-based work on industrial anomaly detection focuses mainly on logical anomalies and pays less attention to structural ones. We balance the two by combining database-based anomaly detection with a VLM: PatchCore supplies regional anomaly scores to the VLM as text, the VLM answers a logical and a structural question set and returns scores, and the two scores are fused to decide whether an image is anomalous. We found that with logical questions alone the VLM aligned poorly with PatchCore on structural anomalies, so we added a structural question set that lets the VLM report more structural causes; we also found that the VLM overrides PatchCore when given final decision authority, so we fuse the scores with equal weights. The framework trains no model parameters. On MVTec LOCO-AD it achieves a logical F1-max of 0.870, logical AUROC of 0.916, structural AUROC of 0.917, and overall AUROC of 0.915, with natural-language explanations of detected anomalies.

語言模型可透過模型上下文協定（Model Context Protocol, MCP）調用外部工具，但通過身分與角色授權，並不代表每次操作都符合使用者原有的行為脈絡。研究原型在 MCP 工具呼叫包裝層建立事件：呼叫前取得使用者、工作階段、工具、時間與工作流程等脈絡，呼叫後補入目的地類別及執行狀態，再選取分析所需欄位形成研究資料；完整提示詞、憑證、敏感工具參數及完整輸出不納入本文分析資料。固定合成測試集共 290 筆事件，工作流程感知規則模型的 Recall、Precision、F1 與 FPR 分別為 0.724、0.974、0.831 與 0.050，ROC-AUC 及 PR-AUC 分別為 0.986 與 0.994。在同一測試集的回溯消融中，加入工作流程脈絡欄位後，F1 由 0.794 增至 0.831。結果顯示，依 MCP 工具呼叫結構形成的操作脈絡可轉換為可追溯的使用者行為資料，並用來辨識時間、工具、目的地與工作流程等偏離。所有結果均來自合成資料，尚不能視為真實企業環境的異常偵測成效；行為分數也不直接改變既有授權決策或工具執行結果。

黑色素瘤資料高度不平衡，且既有多模態研究常採不同模型與實驗設定，難以釐清各架構元件的實際影響。本研究以 ISIC 2020 資料集進行四階段比較，依序評估影像特徵提取器、臨床資料模型、融合策略與損失函數，並以惡性類別 Recall 與 F1-score 為重點。最終架構（EfficientNet + MLP + Ori-concat + Focal Loss）於驗證集達 ROC-AUC 0.9108、Accuracy 0.9470、惡性 Recall 0.5043 與 F1-score 0.2516，其惡性 Recall 高於影像單模態 (0.0342)，Precision (0.1676) 則在 Recall 相近下高於僅用臨床資料之模型 (0.0498)。結果顯示，在本實驗設定下，較複雜的融合機制未提升惡性檢出，而多模態資訊與 Focal Loss 有助於改善極度不平衡情境下的少數類別辨識。

本研究針對海鱺種魚外觀性徵不易由岸上觀察之問題，自行蒐集養殖場水下影像並進行人工標註，建立海鱺種魚性徵成熟度資料集，提出以腹部骨架與條紋幾何特徵為基礎的影像分析方法。系統以 YOLOR 定位魚體並選取側面影像，再以重新訓練之 MobileNet v1 作為特徵擷取網路，利用自建海鱺關鍵點資料微調 OpenPose，預測依養殖專家經驗定義之性徵位置與腹部骨架，透過主成分分析量化條紋特徵長度，將條紋與腹部長度比例 R 及腹部角度 D 輸入 FHRCNN，完成 0 至 3 期外觀性徵分類。研究貢獻包括自建性徵成熟度資料集、海鱺專用關鍵點與骨架設計、OpenPose 任務微調、腹部與條紋幾何表示，以及水下魚體偵測至性徵分期之流程整合。成熟度分類資料包含訓練 1,200 筆、驗證 150 筆及測試 150 筆，依十次實驗結果計算，FHRCNN 平均測試準確率為 74.30%，線性 SVM 平均測試準確率為 74.10%。本研究將養殖人員依經驗觀察的腹部與條紋特徵轉換為量化資料，建立海鱺種魚外觀性徵分期之影像分析基礎。

組合料號圖，又稱為料號說明列表，常存在於電子零件規格書中表示一組相似零件料號的基本特性，而各廠商組合料號圖也常具有不同的樣式，甚至同一家廠商就可能有多種組合料號表示。通常一份組合料號圖上至少會含有下列三部分的資料：樣板、特性、及屬性值(含說明)，而這三個部分的資料內容常會是電子設計自動化公司之工程師所需要參考的資料。而組合料號的形式眾多，而且都是以圖形形式儲存，而不是簡單的以表格儲存，因此更增添了處理時的困難度。由於一份PDF電子規格書可能包含很多頁，因此我們先結合關鍵字判斷及物件偵測的技術找出組合料號圖所在的頁數及該圖在所找到頁數的位置，並把它切割出來。接著我們利用深度學習的方法去擷取一組合料號圖的樣板、特性、及屬性值(含說明)區塊，並使用傳統圖形識別的方法找出組合料號圖的特徵如有無方框、線條方向等精確地框定並配對子名稱與其對應的屬性值以獲得更準確的結果。接下來，我們再運用光學文字辨識技術對所割出來的樣板、特性、及屬性值(含說明)區塊作內容辨識，並利用正規表達式以及黑白名單策略將其分類為子名稱、屬性、值和值集合。實驗顯示不受OCR影響時準確率可達到87%。

Melody–accompaniment separation in solo piano audio requires distinguishing musical roles that share timbre and overlapping spectra. We evaluated two independent spectral two-dimensional (2D) U-Nets with complementary-source masking and mixture-phase reconstruction, trained on piano audio rendered from Classical MIDI files. A composite objective combined spectral L1, multi-scale pooling, waveform scale-invariant signal-to-distortion ratio (SI-SDR), and similarity losses. On 80 held-out test pieces, mean SI-SDR was 8.4 dB for melody and 10.7 dB for accompaniment. In a supplementary re-rendering test with two alternative SoundFonts, mean SI-SDR improvement over the input mixture remained positive, although scores decreased. An exploratory listening study with 126 respondents detected differences among accompaniment excerpts but no significant differences among melody outputs. The findings are limited to synthesized audio; controlled model comparisons, loss ablations, and evaluation on real recordings remain necessary.

Patient-level weak supervision reduces frame-level annotation in gastroscopy but propagates each patient's diagnosis to many nondiagnostic images. We compared eight Vision Transformer backbones for Helicobacter pylori prediction under fine-grained image-level and patient-level weak supervision. Models were developed with five-fold patient-level cross-validation in 327 patients and evaluated in an independent temporal cohort of 97 patients. Under fine-grained supervision, all backbones achieved AUROCs of 92.71%-94.11%. Under weak supervision, seven retained 91.92%-93.77%, whereas native DINOv2 ViT-B/14 declined to 78.29% ± 17.92%. Gastrointestinal continuation self-supervision on 480,000 to 4.8 million GastroNet-5M images restored AUROC to 91.64%-93.25%. The 480,000-image, 50,000-step configuration reached 92.92% ± 0.50%, only 0.33 percentage points below the highest mean. Intermediate supervised fine-tuning on 10,662 HyperKvasir images raised native DINOv2 ViT-B/14 AUROC to 91.76% ± 0.97% but changed already robust backbones little. Patient-level weak supervision therefore exposes representation-specific transfer instability. Effective domain adaptation was possible with sub-million-scale unlabeled data or a smaller labeled gastrointestinal task, but its benefit depended on the starting representation.

心臟電腦斷層可提供高解析度之三維心臟結構資訊，但人工逐層標註心肌、瓣膜與鈣化區域耗時且高度仰賴專業經驗。本研究目的為初探以 nnU-Net v2 建立心臟 CT 多類別自動分割流程，採 3D Full Resolution 組態與五折交叉驗證進行 1000 epochs 訓練。訓練資料共 50 例，每折使用 40 例訓練與 10 例驗證；五折 Mean Validation Dice 分別為 0.5615、0.5562、0.5640、0.5619 與 0.5627，平均為 0.5613。另以五折集成模型對 50 例具參考標準標註之測試影像進行評估，心肌、瓣膜 與 鈣化之 Dice 分別為 0.9226±0.0144、0.7388±0.0424 與 0.0000±0.0000，三類別平均 Dice 為 0.5538。進一步分析發現 calcium 僅出現在 10 例測試資料中，且前景體素數遠少於 muscle 與 valve，模型未預測出任何 calcium voxel，顯示極端類別不平衡仍是本資料集之主要挑戰。研究結果說明 nnU-Net 對心肌與瓣膜具有良好分割能力，但鈣化區域仍需透過資料增補、損失函數調整或取樣策略改善。

Small edits, such as replacing the word dog with or ap- pending the informal phrase no cap to a hypothesis, should not change the label of a natural language inference (NLI) pair, yet they can change the prediction of a fine-tuned classifier. We study two kinds of edit on SNLI and MultiNLI with six fine-tuned encoders: the substitution of 16 nouns by emoji, and the appending of nine informal markers. On average, emoji substitution lowers accuracy by 2.54 to 4.94 points per encoder and appended markers lower it by 4.91 to 12.83 points, but these averages conceal where the accuracy is lost. For emoji, the loss depends on how the replaced nouns are distributed over the two sentences and much less on which noun is replaced. On SNLI it is below one point when the same noun is replaced in both sentences. It is 18 to 24 points for five of the encoders when the premise and the hypothesis contain different replaced nouns. For markers, the loss comes from a few phrases and not from informality as such. Five of nine informal markers cost no more than control phrases, whereas by heaven, which is not informal, changes 39.2% of predictions, more than any marker. A measure that needs no labels ranks the phrases as accuracy does. The two losses also need different fixes: mapping emoji back to nouns at inference removes the emoji loss, and training on other markers recovers 72.1% of the loss on held-out markers. Robustness evaluations should therefore report effects within an edit type and not only its average.

This pilot study examined how information availability relates to role-inference performance by novice humans and large language models (LLMs) in the Werewolf game. Using six logs from five-player games, we compared Target-Role Accuracy at two game stages and classified four logs into Types I–IV accord-ing to their information structures. We evaluated three OpenAI models: gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna, with 10 trials per model. In Quiz 4, where ex-plicit evidence was more abundant, the LLM group showed descriptively higher accuracy than the human group in both temporal settings, although neither differ-ence was statistically significant after Holm adjustment. In Quiz 2 for the specific Type IV log, where objective cues were scarce, the human-group mean was 48.3%, compared with 16.7% for the three-model mean. This result does not establish a general human advantage under low-information conditions. Among the seven par-ticipants analyzed for Type IV, mean accuracy increased from 43% on Day 1 to 86% on Day 3 after game experience and repeated presentation. However, the change was not statistically significant (p = 0.25), and only three participants im-proved. Exploratory analysis suggested that LLM errors often reflected interpreta-tions of active behavior as Villager-like, whereas some humans came to consider the strategic intentions behind utterances. Because each type was represented by one log and the human sample was small, these findings are preliminary and lim-ited to the conditions studied. Further research using more participants and inde-pendent game logs is needed to determine whether these patterns generalize.

This paper studies how large language models (LLMs) can enable natural-language workflow construction and editing on the n8n automation platform. We compare six prompt-engineering strategies on GPT-4.1 against three domain fine-tuning input representations across GPT-4.1, Qwen2.5-Coder 32B, and Gemma 2 27B on a shared 300-template benchmark. Structured, plan-based fine-tuning substantially improves schema-grounded generation over the best prompting approach, raising Connection F1 from 0.254 to 0.543 and Parameter Accuracy from 0.273 to 0.450. Building on this, we develop an agentic system that routes natural-language instructions to creation, insertion, or deletion pipelines, pairing the fine-tuned generator with deterministic programmatic control, and compare it with n8n's own Cloud AI Workflow Builder. The proposed system performs competitively on creation and substantially outperforms the official tool on node deletion (100\% vs. 46.7\%, $N=30$); insertion success is low for all three (5--15\%) from omitted parameter values, though ours best preserves untouched nodes (100\% vs. 66.7--86.7\%).

With the rapid advancement of deep learning, artificial intelligence (AI) applications in predicting financial asset price trends have diversified significantly. Accurate forecasts reduce risks and boost returns, particularly in volatile markets like cryptocurrencies, which offer abundant profit opportunities. However, high-frequency trading data, with its second-level fluctuations, poses challenges for models in handling fine-grained details and adaptability. To tackle these issues, this study introduces a novel hybrid architecture integrating a fine-tuned large language model (LLM) with a daily-retrained deep learning model. The LLM excels at extracting short-term market patterns from vast second-level historical data, addressing traditional models' limitations in retaining past patterns amid massive datasets. Conversely, the deep learning component adapts to long-term trends. To enhance prediction accuracy, we incorporate first-, third-, and fourth-order statistical moments to capture distributional characteristics of price fluctuations. The workflow involves feeding historical data into the LLM for signal generation, then combining these with statistical moments and cross-asset data to support the deep learning model's price movement predictions. Experiments demonstrate that our approach outperforms standard deep learning models, improving accuracy by 5\% to 25\%, confirming the architecture's efficacy.

Assessing argument quality in earnings call statements is essential for interpreting the specificity, persuasiveness, strength, and objectivity of managerial communication. Yet strong performance on individual dimensions does not guarantee that an automated judge preserves the relationships among them. To address this gap, this study investigates the structural fidelity of multidimensional LLM judges and examines whether higher task performance corresponds to better preservation of human-referenced score relationships. We compare three judging systems, namely few-shot prompting, fine-tuning, and a post-hoc correction pipeline, using overall and OBJECTIVE-specific correlation distances, a conditional independence null model, and robustness analyses. The primary comparison is an exploratory analysis of these preexisting systems on a shared 359-statement development split, on which the pipeline's post-hoc rules were tuned. Human annotations exhibit a two-block pattern in which OBJECTIVE is comparatively weakly associated with SPECIFIC, PERSUASIVE, and STRONG. Relative to the observed human labels, most OBJECTIVE-related correlations are higher in the automated outputs, although this direction is sensitive to uncertainty and attenuation in the human reference, and on this split the fine-tuned system shows the largest OBJECTIVE-specific deviation, an ordering not reproduced on a 131-statement held-out subset. The observed comparisons suggest that higher Macro-F1 need not coincide with smaller reference-relative structural discrepancies. Controlled manipulations further show that joint elicitation raises overall but not detectably OBJECTIVE-specific coupling, while false-score disclosure produces only modest, dimension-dependent shifts. These findings extend multidimensional evaluation beyond criterion-level accuracy and provide a reference-based diagnostic for detecting cross-dimensional distortion. In practice, automated judges should be assessed using both predictive performance and structural behavior before they are used in evaluation or decision-support settings.

While digital natives frequently read massive amounts of text on social media without fatigue, they often experience severe psychological aversion and cognitive load when confronted with dense, full-length book pages. Existing e-book designs primarily tweak low-level typography (e.g., font size, line spacing) but fail to address the initial visual dread of ``walls of text.'' In this study, we propose a novel reading support approach that transforms full-text literary novels into familiar social media interfaces---specifically modeling LINE (messaging app), X (microblog feed), and 5ch (bulletin board system, BBS)---alongside an e-book baseline. We modernized ten public-domain Japanese novels and evaluated reading behavior across 20 participants in an experiment with balanced novel--interface assignments, measuring page dwell time, reading speed, text comprehension, and subjective cognitive load (NASA-TLX-inspired 14-item questionnaire). The results reveal that while mean comprehension scores were descriptively similar across interfaces, the LINE-style messaging app interface received favorable subjective evaluations among reluctant readers. These findings suggest that familiar chat and feed layouts may make literary reading more approachable for digital-generation readers.

Persuasion is becoming a machine capability. Randomized studies now find that language-model-written messages change political and factual attitudes at least as much as human-written ones, and this capability couples naturally to the behavioral and affective forecasting that a decade of engagement optimization already produces. We call this coupling AI neuro-persuasion and argue that it is best understood not as a new kind of manipulation but as a multiplier on manipulation that already exists. The contribution is a way to reason about where that multiplier does the most damage. We state a general principle, the steepest derivative: the marginal harm from each increment of persuasion capability is largest where three conditions coincide, a high baseline of organized manipulation, high population vulnerability, and the absence of \emph{model-level} auditing, the ability to test the systems themselves rather than only the content they emit. We treat the ``derivative, not level'' framing as an organizing device rather than a fitted model, and we show the three conditions are general by instantiating them across contexts, synthetic-media financial fraud, electoral information operations, personalized political targeting, and a focused multi-region case study where all three conditions are acute at once. We then add a measurable, cross-cutting vector: models aligned to a principal carry directional loyalties, and as users come to treat a chatbot as a neutral expert those loyalties become standing influence infrastructure. A secondary analysis of published refusal-rate data quantifies one instance, a state-aligned model refuses origin-sensitive prompts at a risk difference of $26$ to $31$ percentage points over comparable models (Cohen's $h = 0.74$ to $1.06$), with no matching elevation on the other side. We engage the counter-capacity that already exists, fact-checking, platform moderation, and media literacy, and argue why it does not reach the model level. As an existence proof that a defense is buildable and deployable by any body that needs one, we describe a small-model instrument coupling a behavioral forecaster with an abstaining cognitive-security check, validated on synthetic controls only and released as open, inspectable code. We make no claim to have measured persuasion in any named campaign.

Blockchain-based Federated Learning (BCFL) remains vulnerable to gradient mimicry attacks, in which malicious clients imitate the behavior of honest ones. We study how such attackers degrade a BCFL system protected by FoolsGold and Multi-Krum, and propose ShardedGold, a defense that combines gradient sharding with an enhanced FoolsGold algorithm adding an Order-based penalty score and a Consistency score, and that stores shards in the InterPlanetary File System (IPFS) with on-chain Content Identifiers (CIDs) to overcome ledger storage limits. We evaluate the full attacker range from 0 to 50% of clients on MNIST and CIFAR-10 and compare against FoolsGold+Multi-Krum run through an identical pipeline. Under the evaluated protocol, ShardedGold keeps malicious clients' aggregation weight below that of honest clients and preserves accuracy up to moderate attacker fractions, degrading gradually beyond them, with accuracy on par with FoolsGold+Multi-Krum. Repeated runs further reveal a bimodal convergence instability that concentrates at intermediate attacker fractions rather than at the maximum, a pattern visible only across multiple runs.

Large language models are increasingly deployed as smart-contract vulnerability auditors. We ask whether such auditors can be evaded by rewrites that provably do not change what a contract does. Using 130 labelled contracts from SmartBugs-Curated and SolidiFI across five vulnerability classes, seven LLM auditors spanning three model families, and six semantic-preserving source transformations (identifier renaming, dead-code insertion, comment stripping, control-flow rewriting, inlining, and reformatting) verified behaviour-preserving at a 99.7% rate by bytecode identity or differential execution, we report two findings. First, obfuscation does not degrade auditors on average - mean evasion is within +/-3 percentage points - but it opens specific, reproducible blind spots: on arithmetic bugs, stripping comments alone collapses the recall of a well-calibrated cheap auditor (Qwen2.5-7B) from 0.88 to 0.48 while its false-positive rate stays at 0.00, with non-overlapping bootstrap confidence intervals at n=33. Comments, not code, were carrying the detection. Second, and more troubling for the field, most auditors we test are degenerate: the largest model (Llama-3.3-70B) attains perfect reentrancy recall while flagging 93% of non-vulnerable contracts, i.e. precision 0.33. Because recall rewards indiscriminate flagging, a recall-only evaluation both overstates auditor quality and makes robustness-to-obfuscation unmeasurable - an apparent "capability confers robustness" correlation (rho = -0.82) is confounded by exactly this effect and is not statistically significant (p = 0.13, five models). Finally, the obvious mitigation fails: escalating to the largest model restores recall (0.46 to 0.96) but inflates false positives (0.00 to 0.26, and to 0.94 on reentrancy). We argue that LLM smart-contract auditors must be evaluated with precision alongside recall, and that cosmetic-robustness testing belongs in that evaluation.

Can observed signal quality and sensor-removal sensitivity reveal activity-classification errors under a sensor-fault family excluded from training and calibration? We test this hypothesis on UCI HAR while preserving its nine official test participants. For each of five simulated fault families, a compact CNN, three logistic correctness predictors, temperature scaling, entropy, and MC dropout exclude that family from every fitting stage. Three classifier seeds and paired corruptions cover 30 held-out family severity sensor-group conditions. Fault augmentation improves mean held-out accuracy from 67.7% to 71.0%, versus 89.9% on clean test windows. However, adding quality and removal features to the confidence-only correctness predictor worsens the prespecified area under the risk-coverage curve (AURC, lower is better) from 0.1473±0.0084 to 0.1681 ± 0.0110. The paired participant-bootstrap difference is +0.0208 (95% interval [+0.0116, +0.0297]). The expanded score helps for held out missing segments and noise, but fails for bias, gain, and freeze; it also raises error AUPRC from 0.560 to 0.574 and slightly lowers risk at 80% coverage, showing that ranking summaries need not agree. A 14k-parameter GRU and a reduced reconstruction-pretraining control, each run on a single seed, are consistent with the same negative AURC pattern. These results reject the broad hypothesis that inexpensive quality and removal signals reliably improve selective HAR under an unseen fault family, and identify family-dependent score transfer as the central obstacle.

A weak audio classifier can correct some errors of a stronger text classifier, but selecting audio indiscriminately also causes damage. We study this problem on the Multimodal EmotionLines Dataset (MELD). We compare trained fusion models with a large language model (LLM) judge that reads dialogue text and the two classifiers’ candidate labels. The judge’s relative support selects between those labels, increasing support-weighted F1 (WF1) from 63.76 to 66.46. Combining its full class scores with text instead reaches 68.41 WF1 and increases macro-averaged F1 (MF1) from 46.49 to 55.33. Output composition improves four backbone–text references, whereas shuffled scores and the tested training-time injections do not reproduce those gains. Prediction-change counts distinguish useful corrections from damage. The results support testing decision-level integration in this setting; they remain exploratory because earlier model development consulted the test set.

This study integrates a Large Language Model (LLM) with a Retrieval-Augmented Generation (RAG) architecture to build a cloud-based dashboard that combines Internet of Things (IoT) data—including underwater cameras, temperature probes, and driftwood-tracking records—with AI prediction models, providing real-time queries and intelligent question-answering services. This strengthens science communication and cross-domain decision support, complementing ongoing efforts in ecological restoration, material innovation, and education outreach to jointly advance data-driven governance. The study site is the Penghu area, where coral in nearby waters has suffered massive mortality in recent years due to climate change. Our team's coral restoration work in the waters of Penghu can help the region develop tourism, offering new approaches for promoting climate resilience and sustainable ocean governance.

Getting humanoid robots to turn natural-language emotional expressions into physical motion is still hard, especially on low-cost hardware that cannot run large end-to-end models. We present HMPA (Hybrid Motor-Proprioception Architecture), a training-free neuro-symbolic pipeline with four layers. A locally hosted large language model (LLM) maps an utterance to a small set of posture parameters (Semantic Motor Imagery); these are interpolated into a trajectory and solved into servo commands by an analytic inverse-kinematics layer (Morphological Negotiation); every frame then passes a rule-based Geometric Safety Valve that enforces joint limits, prevents arm-body self-collision, and checks static stability before the execution driver sends it to the robot. A "Physics of Emotion" prompt tells the LLM how emotions map onto body geometry (height, lean, arm aperture, and timing) instead of asking it for raw joint values. We deployed HMPA on a 16-servo PlayRobot EDE0256 humanoid and generated motions for 40 emotional prompts. A human evaluator rated each recorded video on a three-point scale. The mean semantic-match score was 1.32 and the mean emotional-intensity score was 1.32; only 4 motions clearly matched the intended emotion and 22.5% matched it at least partially. The pipeline produced an executable motion for every prompt on inexpensive hardware, and prototypical emotions such as happiness, sadness, and calm were conveyed clearly, but most emotions were not recognizable from posture alone. These results locate the main bottleneck in expressiveness rather than in feasibility.

Large language models (LLMs) make companion robots conversational, but an LLM that can command motion turns a wrong answer into a physical risk. Our quadruped system, built on ROS 2 and a Unitree Go2 Pro with an 8 GB Jetson Orin Nano Super, separates the two: the LLM may propose at most one skill from a fixed allowlist, and a deterministic Interaction Executive alone decides whether any motion is dispatched. Each gate decision is recorded with its reason, and an operator Studio makes these records inspectable. A 26-person exploratory survey, mostly of demonstration viewers, found perceived intelligence higher than perceived safety (3.76 vs 3.40, paired Wilcoxon p = .015), which motivates, but does not test, safety control that is visible as well as enforced.

Academic publication registration is critical to university administration, yet manual entry is time-consuming and error-prone across 30+ structured fields. This paper presents a non-intrusive automated registration system, implemented as a Chrome Extension, that integrates parallel queries to CrossRef (with DataCite as fallback), OpenAlex, Semantic Scholar, and Unpaywall, together with DBLP-based indexing checks, and applies LLM-based field mapping, using National Chengchi University's publication catalog system as a case study. An ORCID two-stage matching mechanism periodically detects unregistered works and labels them for user confirmation rather than automatic filtering. A three-layer PDF extraction strategy (binary DOI scan, metadata parsing, PDF.js text extraction up to two pages with garbled-text detection) further handles diverse PDF types. Model selection together with two prompt optimization strategies reduced AI field-mapping response time from approximately 26 to 8 seconds, yielding a total end-to-end wait time of about 12 to 16 seconds.

General-purpose LLM sentence embeddings lack domain-specific semantic sensitivity. We propose a hybrid strategy that freezes the pretrained encoder and trains a lightweight Dense adaptation network on Ward-clustering pseudo-labels, mapping embeddings into a 64-dimensional domain-specific space. On 400 UK visa regulations, stratified 5-fold cross-validation shows that the test-fold Silhouette rises from 0.15 to 0.51 (MPNet) and from 0.12 to 0.48 (OpenAI), outperforming PCA, UMAP and LDA. On 20 Newsgroups, evaluated against original categories never used in training, the Silhouette of true categories rises from 0.07 to 0.27, while an oracle trained on true categories reaches 0.41, showing that the adaptation network is effective and performance is bounded by pseudo-label quality. Keywords: Domain-specific Embedding, Supervised Embedding Adaptation, Knowledge Graph, Semantic Clustering, Pseudo-labels.

本研究以台灣加權指數（TAIEX）已實現波動度為預測標的，結合技術面與中文新聞情緒特徵，以技術面（含當日 RV）為基準，並以 walk-forward 逐折配對檢定取代單次切分評估。實證發現新聞情緒對短窗口（次日、次週）波動度具顯著貢獻，長窗口（h=10、22）則各檢定結論不一致；此效應已由Audrino et al. [1]於美股證實，本研究貢獻在於：於散戶成交比重遠高於美股之台灣市場重新驗證，並發現單一衰減設定皆無法穩定重現此效應，以多維半衰期特徵輸入表現較佳。

本研究探討翻譯語言應用於垃圾簡訊辨識之可行性。以 UCI SMS Spam Collection經前處理後之 5,572 筆英文簡訊為資料來源，透過 Qwen3-8B翻譯為繁體中文，並保留正常訊息（ham）與垃圾訊息（spam）標籤。研究採用 BERT-tiny、Chinese RoBERTa與 XLM-RoBERTa，分別以中英文語料進行微調（fine-tuning），並以四種條件比較跨語言的辨識表現，以Accuracy、Precision、Recall 與 F1-Score 進行評估。實驗結果顯示，各模型於原驗證集皆具有良好表現，其中 XLM-R的垃圾訊息 F1-Score 達 0.95 以上；然而，在不同訊息情境的測試集中，模型效能明顯下降。結果顯示，翻譯既有英文標註語料可作為建立繁體中文訓練資料的可行方法，但辨識效能仍受到模型預訓練背景、翻譯品質及資料分布等因素影響。

在法律文本分類領域，部分分類模型存在黑箱問題，使用者難以理解其分類依據。過去多以單字或單句解釋分類模型，雖能標註關鍵內容，但解釋單位較為零碎，可能使法律論述脈絡難以理解。本文提出基於修辭角色與語意一致性片段擾動之法律文本分類解釋，將相同語意詞彙的句子合併，提升解釋效果與可讀性。為提升分類結果的可解釋性，本文提出 SPS-LIME，以修辭角色將相鄰同角色句子合併為語意片段，透過片段擾動與局部代理模型估計各片段對預測的貢獻，再定位片段內的關鍵句子，呈現由論證區塊至句子的雙層解釋，協助使用者回查影響預測的原文內容。SCOTUS 原設定下，片段擾動的 Consistency 為 0.7591、COMP 為 0.0859，均高於詞級與句級擾動；ECtHR 輔助分析亦呈相近趨勢。然而，擾動粒度與文字覆蓋率尚未受控，跨法域角色標註及不同分類器上的適用性仍待驗證。本文提供結構化的局部解釋流程；其歸因反映模型行為，不等同法律推理正確性，亦未經使用者研究驗證理解效果。

臺灣地熱探勘的早期潛能區篩選仍高度依賴地質師人工判釋。本文整理同一地熱資料整合計畫下的三項 GeoAI 應用：以 U-Net 從數值地形模型反演的 12 項地形因子判釋地表線型（滲透性通道）；以 LightGBM 整合 Landsat-8 波段、地形因子與大地電磁法電阻率判釋地表溫度異常區（熱源顯示）；以地端部署之大型語言模型結合地質詞彙庫與檢索增強生成建置地質智慧問答模組。前兩項為獨立的概念驗證研究，第三項已納入上線運作之空間資訊整合系統。以大屯火山群標記訓練的線型模型套用至臺東紅葉與花蓮瑞穗兩處變質岩案場，初步產生與地形紋理相符的線型分布，可歸因於模型僅以地形因子為輸入，惟其構造意義仍待地質驗證；線型模型訓練區測試集 F1 為 0.57；溫度異常模型驗證集 AUC 為 0.91、F1 為 0.86，改以 6 公尺光達地形重算地形因子後召回率提高至 0.92，特徵重要度顯示近紅外與紅光波段及淺層電阻率主導判斷。本文並說明三項應用的分工與定位、現階段的驗證缺口與後續發展。

Decentralized multi-agent systems increasingly rely on shared information, yet repeated interactions allow locally plausible actions to accumulate into distributed harm. We propose an automated trust-management framework that converts noisy monitor scores into directed local trust, aggregates them into a global reputation via an EigenTrust-style process, and removes persistently low-reputation agents. Building on this framework, we study an adaptive Red syndicate using shared Q-learning to farm trust, attack benign agents, smear reputations, or boost malicious teammates. We evaluate pretrust, removal threshold population, defense calibration, and Red observation regimes on attacks, Red survival and eradication, and Blue false isolation. In evaluated regimes, including pretrust anchors in threshold computation can prevent removal; excluding them restores operation but reveals an attack-false isolation trade-off. Full manipulation generally increases attacks and Red survival, while post-attack credence has limited aggregate impact despite affecting greedy action preferences. Results characterize adversarial trust dynamics under uncertain monitoring.

本研究從 NLP representation learning 與 AI emotion modeling 的角度，探討不同認知與語意抽象層次的情緒在 sentence embedding space 中的幾何表徵特性。研究建立包含 12 類情緒、共 1,200 筆英文語句之語料集，並提出 Basic、Reflective 與 Inner 三層操作性情緒架構。實驗採用 all-mpnet-base-v2 與 e5-base-v2 產生句向量，結合 UMAP、silhouette score、centroid cosine distance 與 one-sided permutation test，分析不同情緒層次的群聚、重疊與分離情形。結果顯示，MPNet 與 E5 均呈現一致的 Basic > Reflective > Inner 語意分離度趨勢；其中 Basic 具有較明確的語意邊界，而 Reflective 與 Inner 則呈現較高的重疊與連續性。整體而言，本研究顯示 embedding-space geometry 可提供傳統 emotion classification 指標以外的 representation-level 分析觀點，並可作為後續 hierarchy-aware 或 continuous AI emotion modeling 之研究基礎。

本研究結合可信任人工智慧對話引擎（Trustworthy AI Dialogue Engine, TAIDE）大型語言模型（Large Language Model, LLM）的檢索增強生成（Retrieval-Augmented Generation, RAG）回答、人工評分與Quantum Computational Intelligence and Artificial Intelligence（QCI&AI）模糊推論，評估醫療衛教問答。兩位護理學系四年級學生（R1、R2）對741題文字回答的正確性、匹配度、完整性及語言與表達各評分1至5分，共5,928筆評分。四項平均依序為3.89、3.82、3.84、3.56分，語言與表達較低，可優先改善。為評估兩位評分者對相同回答的給分一致程度，本研究採用線性加權Cohen’s κ進行分析。四項評分項目的κ值介於0.218至0.264，顯示評分一致程度仍有提升空間。QCI&AI結合答案語意相似度與人工評分，並計算系統判定的等級與部分輸入評分的平均值所換算的等級相同的比例，R1、R2原為85.8%、89.7%。以基因演算法（Genetic Algorithm, GA）及粒子群最佳化演算法（Particle Swarm Optimization, PSO）調整參數後，R1採GA達94.7%，R2採PSO達97.6%。人工評分結果可供後續重新檢查回答內容；模糊推論結果則反映此批資料的等級匹配情形，尚不能證明模型能獨立判斷回答品質，或在其他題目上有相同表現。

隨著大型語言模型的發展，自然語言處理的效能已顯著提升。然而，其仍面臨高計算成本與高硬體需求等挑戰。混合專家模型（Mixture of Experts, MoE）能在固定計算成本下擴展模型容量，因而有效降低這些需求。MoE 模型相較於具相同計算複雜度的密集模型，具有更高的記憶體需求，進而導致專家冗餘與有效容量低等問題。本研究旨在透過分析訓練期間各專家的激活頻率，判定專家對模型的重要性，並據此調整其參數。激活頻率作為縮放因子，可擴大高頻率專家並縮小低頻率專家，以最佳化模型的資源配置。於 SQuAD資料集上的測試結果顯示，三種BERT模型經雙向縮放後的準確率均略微提升，但部分F1分數有所下降。僅縮減參數的策略則可減少約17%至42%的參數量，同時維持或略微提升BERT模型效能。然而，GPT-2模型在兩種縮放策略下的準確率均略微下降。實驗結果顯示，本方法具有減少模型參數冗餘的潛力，但其效能改善仍因模型架構而異，未來需進一步驗證其資源效率與泛化能力。

本研究提出一套以結構化 Prompt 為核心的多輪適性評量題目生成流程。系統依教材建立核心概念與知識點，生成含診斷式誘答的起始題目（BQ），並將各誘答對應至可能錯誤理解與追問策略，再據此產生後續追問題（FQ）。題目於施測前經生成、審查與重新生成流程後納入題庫。期末施測共 124 位學生、985 筆有效 BQ 題次，初答正確率為 61.52%，追問與重答後最終正確率為 73.50%；初答錯誤題次中，31.13% 於追問後改答正確。結果顯示，本流程可依學生實際錯誤選項提供差異化追問；由於本次未設無 FQ 控制組，相關結果僅作描述性解讀。

Study-abroad advising draws on official requirements, community-reported experiences, and faculty research information, which differ in authority and structure. Retrieval-augmented generation ranks evidence by relevance, yet a relevant community report about a GPA threshold is not an official requirement. Our insight is that source authority should shape retrieval routing, not be considered only after retrieval. We present a source-aware agentic RAG system built with LangGraph that routes official requirements to text-to-SQL, experience questions to retrieval over community reports, and faculty questions to query-time search; a bounded evidence check allows one retry and otherwise abstains. On 34 official-data queries at five universities, the system answered all correctly; on 60 threshold questions for 20 universities without indexed official data, it abstained on all and gave no unsupported numerical threshold, whereas a ReAct baseline with different prompting and answer model gave one for 55%. An ablation attributes this mainly to intent-based routing rather than the evidence check. The adversarial set does not test over-refusal or source conflicts.

刑事案件的對話紀錄數量龐大且脈絡分散，又常夾雜代稱與錯字。本文提出以檢索增強生成（RAG）整理線索的流程。流程以單則訊息作為引用與稽核單位，以同一聊天室的連續訊息作為檢索脈絡，再透過主張與證據對照及人工檢核限制模型推論。原型已在一件案件的數位證物工作複本上完成功能展示，並以去識別化的真實對話進行前導測試。量化評估使用220則合成對話與32道標註題目，比較BM25、BGE-M3向量與兩者融合的檢索，並以qwen3:8b評估受限生成。三種檢索方法的Top-5關鍵訊息找回率為0.833、0.667與0.792，代稱題為0.250至0.500。生成結果沒有超出取回範圍的引用，應拒答題的拒答率為0.875至1.000，三次重複輸出完全相同。代稱與暗語查詢仍需搭配偵查人員的經驗判讀。

預測學生學業表現是教育資料探勘的重要目標之一，可協助及早辨識可能需要額外支援的學生，並提供資料驅動的教育決策資訊。本研究應用資料探勘技術預測學生學業表現，並辨識與分類結果具有較高關聯性的特徵。研究以學生行為與學業資料為基礎，在 RapidMiner 中建立決策樹模型，並使用 Optuna 進行超參數優化。資料集被劃分為相互獨立的訓練集與測試集，超參數優化僅於訓練集上透過 10 折交叉驗證進行。優化後模型在獨立測試集上的準確率為 68.42%，相較於 RapidMiner 預設模型的 65.26%，提升 3.16 個百分點。特徵重要性分析顯示，Visited Resources 與 Discussion 為最具影響力的預測特徵，其次為 Announcements View 與 Raised Hands。結果顯示，超參數優化可使決策樹分類效能獲得小幅改善，並顯示學生使用學習資源與課堂互動等行為特徵對學業表現預測具有重要性。

本研究提出一套以強弱人工智慧決策差異驅動的適應式五子棋教學系統。強AI採用Policy-Value Network搭配Monte Carlo Tree Search（MCTS），弱AI採用深度3之Alpha-Beta Minimax與棋型評估。兩者的走法會以400次搜尋預算的MCTS統一重新評估並計算分數差距；若分數損失超過門檻且玩家在該局面下的走法不屬於可接受集合時會觸發提示；此時系統會將程式辨識的棋型和強AI的推薦座標等盤面資訊交由大型語言模型產生說明。實驗驗證則由20位參與者完成3個主題共10題的問卷，主題分別為提示時機、教學內容與整體知覺效用，平均分數分別為4.25、3.90與4.07，十題總平均為4.06。結果顯示參與者肯定系統的提示時機、教學內容與整體教學價值。

隨著生成式人工智慧與大型語言模型快速發展，如何確認學生是否理解生成之程式碼，成為程式設計教學的重要議題。為此，本研究開發整合於 Visual Studio Code（VS Code）的 IntriCode，將既有程式評測與程式理解測驗整合於單一開發環境。系統包含程式碼評測與智慧出題兩大模組，前者提供程式提交、執行與結果回傳，後者利用 Tree-sitter 建立抽象語法樹（AST），進行函式特徵分析與篩選，並透過機器學習預測三種題型的適合程度，再結合大型語言模型生成程式理解題目。此外，本研究透過實際課堂問卷評量，蒐集學生對系統操作、程式檢測及智慧出題功能的使用回饋，作為系統評估與後續改進之依據。

隨著商用無人機應用逐步普及，帶動無人機操作員人力市場的發展，然而當前市場極度缺乏可信賴的能力指標：各國發照制度著重於法規知識與基本操作能力，職業所需的實務飛行操作水準往往未能完整呈現（美國 Part 107 航空人員檢定標準之技能與風險管理欄位全部標示為 [Reserved] 與 [Not Applicable]），以致市場僅能以總飛行時數作為判別指標。本研究採設計科學研究（DSR）典範，建構並評估一套六構面、總分 1,000 分之無人機操作員操作水平評分成品，其核心設計主張為：複合分數之權重應可追溯至既有的效標關聯效度證據，而非由設計者直覺設定。權重依 Sackett 等（2022）之修正效度估計推導，並疊加時間衰減、部分信度與非補償性閘門三項機制。四項模擬顯示：臨界交互相關 r* = 0.548 而模擬所得實際平均相關 r = 0.398，故六構面複合架構成立；權重 ±20% 擾動下等級改變率 P95 僅 0.50%；非補償性閘門亦阻止 9.1% 高風險操作員取得 B 級以上評等

親密暴力問卷的無監督分析面臨題項依存關係、缺乏真實群體標籤與小樣本等問題。本文提出結合題項關聯表徵、深度分群與注意力導向資料擴增的架構：先將 30題、0–4分的問卷轉換為三種30×30題項關聯矩陣，再以Encoder–Decoder建立低維表徵，並結合KL散度、群內一致性與群間差異性損失進行分群；最後以分群結果訓練分類器，利用 Grad-CAM 找出重要的題項關聯區域並產生擴增資料。實驗使用 180 份訓練資料、2,831 份線上測試資料與 331份由專業輔導員協助填寫的測試資料。結果顯示，注意力導向資料擴增在多數設定下提升平均 Cronbach’s alpha，但效果仍受題項關聯表示與資料分布影響。