Selected work

Representative Papers

The Illusion of Cross-Lingual Safety in Low-Resource Languages

Aug 11, 2026

Current safety alignment of large language models is predominantly based on English, and their cross-lingual generalization to low-resource languages remains poorly understood, posing potential risks. This work introduces the LoDNA dataset, comprising both literal translations and culturally localized prompts, to systematically evaluate the transferability of safety mechanisms across four African languages. We propose a probing method grounded in the geometric structure of the model’s latent space to analyze internal representations underlying refusal behaviors. Our study reveals, for the first time, significant limitations in cross-lingual safety alignment: in most language–model combinations, harmful prompts retain less than 10% of the refusal signal observed in English, indicating that semantic alignment does not ensure consistent safety routing. These findings challenge the assumption of a language-invariant harm manifold.

0 citationsRead paper

Majority Vote Silences Minority Values: Annotator Disagreement at the Hate/Offensive Boundary in HateXplain

Jun 27, 2026

This study addresses the limitations of majority voting in annotating hate and offensive speech, which obscures substantial annotator disagreements on subjective boundaries and leads models to treat contested judgments as objective truths. Focusing on the HateXplain dataset, the authors systematically analyze disagreement patterns and evaluate three modeling approaches: hard-label BERT, soft-label models, and per-annotator multi-head architectures. Through chi-square tests and confidence analysis, they demonstrate that all models suffer a 22–28 percentage point accuracy drop on contentious samples, with the multi-head model achieving only 0.245 accuracy on offense-related disagreements. Critically, standard evaluation metrics fail to capture this degradation, as models often exhibit spuriously high confidence in incorrect predictions. The work exposes the structural bias inherent in majority voting for sensitive content annotation and its detrimental impact on model evaluation reliability.

0 citationsRead paper

Physics-Informed Residuals for Adaptive Mesh Refinement in Finite-Difference PDE Solvers

Jun 01, 2026

This work addresses the inefficiency of traditional finite difference methods on uniform grids in resolving localized complex features—such as shocks or steep gradients—without excessive degrees of freedom. The authors propose a hybrid adaptive strategy that leverages the residual of a physics-informed neural network (PINN) as an off-grid indicator to generate cell-wise refinement metrics. These metrics drive non-uniform mesh refinement through a combination of Dörfler and threshold-based marking criteria, while high-fidelity solutions are still computed using finite differences. Innovatively, the PINN serves not as a solver but as an adaptive guidance tool, effectively integrating physical priors with classical numerical methods. In one-dimensional viscous Burgers’ equation, the approach achieves superior accuracy with only 60 degrees of freedom—outperforming a uniform 192-point grid by reducing error by 67.5%—and demonstrates marked advantages over random refinement strategies in two- and three-dimensional proxy problems.

0 citationsRead paper

Machine Learning for Detection and Severity Estimation of Sweetpotato Weevil Damage in Field and Lab Conditions

Feb 06, 2026

This study addresses the limitations of traditional manual assessment of sweetpotato weevil damage, which relies on subjective judgment and suffers from low efficiency and poor consistency, thereby hindering progress in insect-resistant breeding. To overcome these challenges, the authors propose a computer vision approach integrating field and laboratory settings, leveraging YOLOv12 for object detection, image segmentation, and a tiling strategy to establish a two-stage laboratory detection pipeline that significantly enhances the identification of minute feeding holes. The method achieves 71.43% accuracy in classifying field-based damage severity levels and attains a mean average precision of 77.7% for detecting microscopic holes in controlled laboratory conditions. Notably, this work presents the first application of YOLOv12 to sweetpotato weevil phenotyping, enabling efficient, objective, and automated damage evaluation.

0 citationsRead paper

Intelligent Traffic Surveillance for Real-Time Vehicle Detection, License Plate Recognition, and Speed Estimation

Jan 01, 2026arXiv.org

This study addresses the high incidence of traffic accidents caused by speeding in resource-constrained regions such as Uganda by proposing a lightweight, real-time intelligent traffic monitoring system. The system integrates YOLOv8 for license plate detection and a hybrid CNN-Transformer model for character recognition, achieving a remarkably low error rate of 1.79%. It further introduces an innovative region-of-interest-based method for vehicle speed estimation with an accuracy within ±10 km/h. By leveraging the Africa’s Talking communication API, the system enables automatic linkage to vehicle owner databases and sends real-time violation alerts via SMS, establishing an end-to-end automated enforcement pipeline. Experimental results demonstrate a license plate detection mAP of 97.9%, significantly outperforming existing solutions designed for low-resource settings.

0 citationsRead paper

Recent publications

Latest Papers

The Illusion of Cross-Lingual Safety in Low-Resource Languages

Aug 11, 2026

Current safety alignment of large language models is predominantly based on English, and their cross-lingual generalization to low-resource languages remains poorly understood, posing potential risks. This work introduces the LoDNA dataset, comprising both literal translations and culturally localized prompts, to systematically evaluate the transferability of safety mechanisms across four African languages. We propose a probing method grounded in the geometric structure of the model’s latent space to analyze internal representations underlying refusal behaviors. Our study reveals, for the first time, significant limitations in cross-lingual safety alignment: in most language–model combinations, harmful prompts retain less than 10% of the refusal signal observed in English, indicating that semantic alignment does not ensure consistent safety routing. These findings challenge the assumption of a language-invariant harm manifold.

0 citationsRead paper

Majority Vote Silences Minority Values: Annotator Disagreement at the Hate/Offensive Boundary in HateXplain

Jun 27, 2026

This study addresses the limitations of majority voting in annotating hate and offensive speech, which obscures substantial annotator disagreements on subjective boundaries and leads models to treat contested judgments as objective truths. Focusing on the HateXplain dataset, the authors systematically analyze disagreement patterns and evaluate three modeling approaches: hard-label BERT, soft-label models, and per-annotator multi-head architectures. Through chi-square tests and confidence analysis, they demonstrate that all models suffer a 22–28 percentage point accuracy drop on contentious samples, with the multi-head model achieving only 0.245 accuracy on offense-related disagreements. Critically, standard evaluation metrics fail to capture this degradation, as models often exhibit spuriously high confidence in incorrect predictions. The work exposes the structural bias inherent in majority voting for sensitive content annotation and its detrimental impact on model evaluation reliability.

0 citationsRead paper

Physics-Informed Residuals for Adaptive Mesh Refinement in Finite-Difference PDE Solvers

Jun 01, 2026

This work addresses the inefficiency of traditional finite difference methods on uniform grids in resolving localized complex features—such as shocks or steep gradients—without excessive degrees of freedom. The authors propose a hybrid adaptive strategy that leverages the residual of a physics-informed neural network (PINN) as an off-grid indicator to generate cell-wise refinement metrics. These metrics drive non-uniform mesh refinement through a combination of Dörfler and threshold-based marking criteria, while high-fidelity solutions are still computed using finite differences. Innovatively, the PINN serves not as a solver but as an adaptive guidance tool, effectively integrating physical priors with classical numerical methods. In one-dimensional viscous Burgers’ equation, the approach achieves superior accuracy with only 60 degrees of freedom—outperforming a uniform 192-point grid by reducing error by 67.5%—and demonstrates marked advantages over random refinement strategies in two- and three-dimensional proxy problems.

0 citationsRead paper

Machine Learning for Detection and Severity Estimation of Sweetpotato Weevil Damage in Field and Lab Conditions

Feb 06, 2026

This study addresses the limitations of traditional manual assessment of sweetpotato weevil damage, which relies on subjective judgment and suffers from low efficiency and poor consistency, thereby hindering progress in insect-resistant breeding. To overcome these challenges, the authors propose a computer vision approach integrating field and laboratory settings, leveraging YOLOv12 for object detection, image segmentation, and a tiling strategy to establish a two-stage laboratory detection pipeline that significantly enhances the identification of minute feeding holes. The method achieves 71.43% accuracy in classifying field-based damage severity levels and attains a mean average precision of 77.7% for detecting microscopic holes in controlled laboratory conditions. Notably, this work presents the first application of YOLOv12 to sweetpotato weevil phenotyping, enabling efficient, objective, and automated damage evaluation.

0 citationsRead paper

Intelligent Traffic Surveillance for Real-Time Vehicle Detection, License Plate Recognition, and Speed Estimation

Jan 01, 2026arXiv.org

This study addresses the high incidence of traffic accidents caused by speeding in resource-constrained regions such as Uganda by proposing a lightweight, real-time intelligent traffic monitoring system. The system integrates YOLOv8 for license plate detection and a hybrid CNN-Transformer model for character recognition, achieving a remarkably low error rate of 1.79%. It further introduces an innovative region-of-interest-based method for vehicle speed estimation with an accuracy within ±10 km/h. By leveraging the Africa’s Talking communication API, the system enables automatic linkage to vehicle owner databases and sends real-time violation alerts via SMS, establishing an end-to-end automated enforcement pipeline. Experimental results demonstrate a license plate detection mAP of 97.9%, significantly outperforming existing solutions designed for low-resource settings.

0 citationsRead paper