1

GateKD introduces a confidence-gated closed-loop reasoning distillation framework designed to transfer robust reasoning abilities from large language models into compact student models. Unlike conventional open-loop distillation approaches that assume uniformly reliable teacher supervision, GateKD dynamically modulates supervision quality through confidence-aware gating mechanisms. The framework integrates confidence-gated soft targets, gated hidden-state alignment, and reliability-filtered attention transfer to selectively preserve trustworthy reasoning trajectories while suppressing hallucinated or noisy intermediate representations. Extensive experiments on commonsense, logical, and symbolic reasoning benchmarks demonstrate that GateKD consistently improves reasoning fidelity, robustness, and low-resource generalization across T5 and Flan-T5 student models.

CAREF introduces a calibration-aware regularization framework designed to improve explanation faithfulness in large language models without requiring rationale supervision. The framework is built around the novel Sparsity-Calibrated Entropic Divergence (SCED) objective, which unifies entropy calibration and adaptive token-level sparsity within a single differentiable regularization term. By encouraging predictions to rely on compact, stable, and decision-relevant token subsets, CAREF strengthens the connection between model decisions and generated explanations while maintaining strong predictive performance. Extensive evaluations on four Natural Language Explanation benchmarks demonstrate consistent improvements in both accuracy and explanation alignment, with the CAREF-AQ variant achieving state-of-the-art results using only 6.43% of trainable parameters.

HERS presents a domain-adaptive diffusion framework for controllable, realistic, and trustworthy vehicle damage synthesis. The method decomposes complex damage generation into a set of risk-specific expert modules, each specializing in a particular damage type such as dents, scratches, broken lights, or cracked paint, and trained using self-supervised image–text pairs without manual annotation. These experts are later integrated into a unified diffusion model that balances specialization with generalization, enabling precise control over damage attributes while maintaining visual coherence. Extensive experiments across multiple diffusion backbones demonstrate consistent improvements in text–image alignment and human preference over standard fine-tuning baselines. Beyond visual fidelity, HERS highlights broader implications for auditability, fraud prevention, and the responsible deployment of generative models in high-stakes domains, underscoring the need for trustworthy and risk-aware diffusion systems in applications such as automated insurance assessment.

This study provides a rigorous and model-aware examination of the faithfulness and spatial reliability of Grad-CAM explanations for lung cancer CT image classification across both convolutional neural networks and Vision Transformer architectures. By systematically analyzing localization accuracy, perturbation-based faithfulness, and explanation consistency, the work reveals pronounced architecture-dependent disparities in how visual explanations align with true diagnostic evidence. While Grad-CAM often produces visually convincing heatmaps for convolutional models, these explanations can be spatially coarse or influenced by spurious correlations, raising concerns about shortcut learning and misleading interpretability. More critically, the analysis demonstrates that transformer-based models, despite strong predictive performance, exhibit a marked degradation in Grad-CAM reliability due to non-local attention mechanisms. Together, these findings underscore a central message, visually appealing explanations do not necessarily imply faithful model reasoning. The work highlights fundamental limitations of saliency-based XAI methods in high-stakes medical imaging and calls for more principled, model-aware interpretability approaches that can support genuinely trustworthy and clinically meaningful AI systems.

Political stance detection in low-resource and culturally complex settings poses a critical challenge for large language models (LLMs). In the Thai political landscape - marked by indirect language, polarized figures, and entangled sentiment and stance - LLMs often display systematic biases such as sentiment leakage and favoritism toward entities. These biases undermine fairness and reliability. We present ThaiFACTUAL, a lightweight, model-agnostic calibration framework that mitigates political bias without requiring fine-tuning. ThaiFACTUAL uses counterfactual data augmentation and rationale-based supervision to disentangle sentiment from stance and reduce bias. We also release the first high-quality Thai political stance dataset, annotated with stance, sentiment, rationales, and bias markers across diverse entities and events. Experimental results show that ThaiFACTUAL significantly reduces spurious correlations, enhances zero-shot generalization, and improves fairness across multiple LLMs. This work highlights the importance of culturally grounded debiasing techniques for underrepresented languages.

This paper dives into the cutting-edge world of road asset detection on Thai highways, showcasing a novel approach that combines an upgraded REG model with Generalized Focal Loss. Our focus is on identifying key road elements—like pavilions, pedestrian bridges, information and warning signs, and concrete guardrails—to boost road safety and infrastructure management. While deep learning methods have shown promise, traditional models often struggle with accuracy in tricky conditions, such as cluttered backgrounds and variable lighting. To tackle these issues, we’ve integrated REG with Generalized Focal Loss, enhancing its ability to detect road assets with greater precision. Our results are impressive, the REGx model led the way with a mAP50 of 80.340, mAP50-95 of 60.840, precision of 79.100, recall of 76.680, and an F1-score of 77.870. These findings highlight the REGx model’s superior performance, demonstrating the power of advanced deep learning techniques to improve highway safety and infrastructure maintenance, even in challenging conditions.

Colorectal cancer is one of the leading causes of cancer death worldwide. As of now, colonoscopy is the most effective screening tool for diagnosing colorectal cancer by searching for polyps which can develop into colon cancer. The drawback of manual colonoscopy process is its high polyp miss rate. Therefore, polyp detection is a crucial issue in the development of colonoscopy application. Despite having high evaluation scores, the recently published methods based on fully convolutional network (FCN) require a very long inferring (testing) time that cannot be applied in a real clinical process due to a large number of parameters in the network. In this paper, we proposed a compressed fully convolutional network by modifying the FCN-8s network, so our network is able to detect and segment polyp from video images within a real-time constraint in a practical screening routine. Furthermore, our customized loss function allows our network to be more robust when compared to the traditional cross-entropy loss function. The experiment was conducted on CVC-EndoSceneStill database which consists of 912 video frames from 36 patients. Our proposed framework has obtained state-of-the-art results while running more than 7 times faster and requiring fewer weight parameters by more than 9 times. The experimental results convey that our system has the potential to support clinicians during the analysis of colonoscopy video by automatically indicating the suspicious polyps locations.