Institution profile

Heritage Institute of Technology

Academic institutionasia · in
Official website
Research library8linked papers
Opportunities0open roles
Selected work

Representative Papers

Position, Not Provenance: Separating Reasoning Mediation from Sycophancy in Medical Vision-Language Models

Jul 29, 2026

It remains unclear whether the reasoning generated by current medical vision-language models (VLMs) genuinely influences their predictions or merely reflects deference to authoritative sources. To address this, this work proposes the CoT-Mediate framework, which disentangles the location and source of reasoning by perturbing clinical attributes within chain-of-thought (CoT) sequences and integrating a two-arm protocol with controlled source interventions. Experiments on LLaVA-Med and MedGemma reveal that prefix-forced continuation more faithfully captures the causal effect of reasoning than prompt-based weighting; removing visual evidence increases reliance on injected reasoning; and crucially, the model’s use of reasoning is primarily governed by its position in the context rather than its declared source—challenging conventional assumptions about reasoning faithfulness in VLMs.

0 citationsRead paper

Ellipse Meets Bit-Planes: A Novel Approach to RNFL based Glaucoma Detection Using Advanced Image Processing and Deep Learning

Jun 14, 2026

This study addresses the challenge of analyzing retinal nerve fiber layer (RNFL) due to positional variations of the optic disc and macula by proposing a dual-framework automated glaucoma detection method based on elliptical polar coordinate transformation. The approach integrates adaptive elliptical polar coordinate transformation with deep learning-based feature fusion to construct a high-accuracy model, while also incorporating bit-plane slicing image processing to develop a lightweight alternative. Experimental results demonstrate that the high-accuracy framework achieves a detection rate of 99.3%, and the lightweight framework attains an accuracy of 92.31%. By balancing performance with computational efficiency, the proposed method offers a scalable and cost-effective solution for early glaucoma screening.

0 citationsRead paper

SolarTformer: A Transformer Based Deep Learning Approach for Short Term Solar Power Forecasting

Apr 27, 2026

This study addresses the challenge of insufficient accuracy in short-term solar power forecasting by introducing, for the first time, the Transformer architecture to this task. Leveraging its self-attention mechanism, the proposed model effectively captures both temporal dependencies and spatial variations in solar irradiance, while incorporating power plant metadata to enhance generalization. Within a unified framework, the method achieves high-precision predictions across diverse sites and seasons, demonstrating consistently superior robustness and generalization performance compared to existing models under varied weather conditions—including clear skies and overcast days.

0 citationsRead paper
Recent publications

Latest Papers

Position, Not Provenance: Separating Reasoning Mediation from Sycophancy in Medical Vision-Language Models

Jul 29, 2026

It remains unclear whether the reasoning generated by current medical vision-language models (VLMs) genuinely influences their predictions or merely reflects deference to authoritative sources. To address this, this work proposes the CoT-Mediate framework, which disentangles the location and source of reasoning by perturbing clinical attributes within chain-of-thought (CoT) sequences and integrating a two-arm protocol with controlled source interventions. Experiments on LLaVA-Med and MedGemma reveal that prefix-forced continuation more faithfully captures the causal effect of reasoning than prompt-based weighting; removing visual evidence increases reliance on injected reasoning; and crucially, the model’s use of reasoning is primarily governed by its position in the context rather than its declared source—challenging conventional assumptions about reasoning faithfulness in VLMs.

0 citationsRead paper

Ellipse Meets Bit-Planes: A Novel Approach to RNFL based Glaucoma Detection Using Advanced Image Processing and Deep Learning

Jun 14, 2026

This study addresses the challenge of analyzing retinal nerve fiber layer (RNFL) due to positional variations of the optic disc and macula by proposing a dual-framework automated glaucoma detection method based on elliptical polar coordinate transformation. The approach integrates adaptive elliptical polar coordinate transformation with deep learning-based feature fusion to construct a high-accuracy model, while also incorporating bit-plane slicing image processing to develop a lightweight alternative. Experimental results demonstrate that the high-accuracy framework achieves a detection rate of 99.3%, and the lightweight framework attains an accuracy of 92.31%. By balancing performance with computational efficiency, the proposed method offers a scalable and cost-effective solution for early glaucoma screening.

0 citationsRead paper

SolarTformer: A Transformer Based Deep Learning Approach for Short Term Solar Power Forecasting

Apr 27, 2026

This study addresses the challenge of insufficient accuracy in short-term solar power forecasting by introducing, for the first time, the Transformer architecture to this task. Leveraging its self-attention mechanism, the proposed model effectively captures both temporal dependencies and spatial variations in solar irradiance, while incorporating power plant metadata to enhance generalization. Within a unified framework, the method achieves high-precision predictions across diverse sites and seasons, demonstrating consistently superior robustness and generalization performance compared to existing models under varied weather conditions—including clear skies and overcast days.

0 citationsRead paper