Selected work

Representative Papers

When Less Is Enough: Context Selection and Prompting Strategies for Bengali News Headline Generation

Aug 16, 2026

This study addresses the sensitivity of large language models to context and prompting in Bengali news headline generation by systematically evaluating mainstream models and few-shot strategies. Results demonstrate that curated lead paragraphs outperform full-text inputs, confirming that context quality supersedes length. Furthermore, cross-lingual prompting combined with context augmentation yields significant improvements over native prompts, albeit with model-dependent efficacy. While one-shot learning substantially enhances Gemini’s performance, it offers limited benefits for Llama. These findings underscore the critical roles of context selection and prompt engineering in low-resource language scenarios, providing effective optimization pathways for deploying large language models in such settings.

0 citationsRead paper

HybridRAG-BN: A Retrieval-Augmented Framework with Fine-Tuned Verification for Bangla KBQA

Aug 13, 2026

This work addresses the challenges of weak retrieval capability, scarce linguistic resources, and difficulty in accurately grounding answers to external knowledge in Bengali knowledge base question answering (KBQA). To tackle these issues, we propose the first end-to-end framework integrating hybrid retrieval, large language model generation, and lightweight fine-tuning for verification. Our approach combines BM25 with BGE-M3 for hybrid retrieval, leverages Gemma-4-31B-Instruct (in GGUF format) for answer generation, and employs LoRA fine-tuning to enhance generation accuracy, further augmented by DuckDuckGo fallback search for robustness. The system achieves token-level F1 scores of 0.71654 and 0.72912 on public and private test sets, respectively—both ranking first in the competition—and substantially advances the performance frontier for KBQA in low-resource languages.

0 citationsRead paper

Hybrid Semantic and Spectral Ensemble for Robust Synthetic Image Source Attribution

Jul 24, 2026

This work addresses the challenge of distribution shift and degraded source-tracing accuracy in synthetic images caused by unknown post-processing operations such as JPEG compression or blurring. To this end, the authors propose a dual-branch ensemble framework that efficiently integrates semantic deep learning—using EfficientNet-B0 enhanced with exponential moving average (EMA) and label smoothing—with lightweight mathematical forensic features derived from high-pass noise residuals. These handcrafted features comprise a 126-dimensional vector incorporating singular value decomposition (SVD) spectra, local binary patterns, and dimensionality reduction via truncated SVD. The combined representations are fused through an XGBoost classifier. Requiring only CPU computation, the method achieves 95.60% accuracy on a challenging benchmark where 55% of test samples undergo degradation, completing the entire pipeline within 6.5 hours. This approach demonstrates remarkable robustness to post-processing while offering high accuracy, computational efficiency, and practical deployability.

0 citationsRead paper

RMS@CC-MMD 2026: Multimodal Misogyny Detection via Geometric Interaction and Multi-View Consensus

Jul 20, 2026

This work addresses the challenge of detecting implicit misogynistic content in internet memes, which often relies on semantic incongruity between images and text as well as culturally specific contexts. To this end, the authors propose GeoMVC, a novel approach featuring a geometric interaction layer that dynamically models alignment between image and text embeddings using Hadamard product and cosine similarity, replacing conventional static concatenation. Additionally, a multi-view consensus mechanism is introduced to aggregate predictions from original, length-filtered, and English-translated text views, thereby mitigating distributional shifts caused by OCR noise and transcription errors. Evaluated on Task A of the ICMI 2026 CC-MMD Challenge, GeoMVC achieved second place for Malayalam (Macro F1: 0.892), third for Chinese (0.895), and fifth for Tamil (0.521), demonstrating its effectiveness and robustness in cross-lingual multimodal misogyny detection.

0 citationsRead paper

Transferable Low-Rank Convolutional Bases for Onboarding Unseen Medical Imaging Modalities

Jul 18, 2026

This work addresses the challenge of efficiently adapting medical imaging models to unseen modalities post-deployment while avoiding catastrophic forgetting. The authors propose a parameter-efficient adaptation method that, under a strict leave-one-domain-out setting, freezes the pre-trained convolutional backbone and leverages transferable low-rank convolutional bases learned from source modalities. Adaptation is achieved solely through projection parameters atop these bases, constituting only 0.78% of the total model parameters. By integrating convolutional LoRA, low-rank decomposition, and Mahalanobis anomaly detection, the approach improves adaptation accuracy on new modalities by 6.11 percentage points over random bases, while incurring zero performance loss on source modalities (Δ = 0.00 pp), substantially outperforming full fine-tuning and decision-level adaptation strategies.

0 citationsRead paper

Recent publications

Latest Papers

When Less Is Enough: Context Selection and Prompting Strategies for Bengali News Headline Generation

Aug 16, 2026

This study addresses the sensitivity of large language models to context and prompting in Bengali news headline generation by systematically evaluating mainstream models and few-shot strategies. Results demonstrate that curated lead paragraphs outperform full-text inputs, confirming that context quality supersedes length. Furthermore, cross-lingual prompting combined with context augmentation yields significant improvements over native prompts, albeit with model-dependent efficacy. While one-shot learning substantially enhances Gemini’s performance, it offers limited benefits for Llama. These findings underscore the critical roles of context selection and prompt engineering in low-resource language scenarios, providing effective optimization pathways for deploying large language models in such settings.

0 citationsRead paper

HybridRAG-BN: A Retrieval-Augmented Framework with Fine-Tuned Verification for Bangla KBQA

Aug 13, 2026

This work addresses the challenges of weak retrieval capability, scarce linguistic resources, and difficulty in accurately grounding answers to external knowledge in Bengali knowledge base question answering (KBQA). To tackle these issues, we propose the first end-to-end framework integrating hybrid retrieval, large language model generation, and lightweight fine-tuning for verification. Our approach combines BM25 with BGE-M3 for hybrid retrieval, leverages Gemma-4-31B-Instruct (in GGUF format) for answer generation, and employs LoRA fine-tuning to enhance generation accuracy, further augmented by DuckDuckGo fallback search for robustness. The system achieves token-level F1 scores of 0.71654 and 0.72912 on public and private test sets, respectively—both ranking first in the competition—and substantially advances the performance frontier for KBQA in low-resource languages.

0 citationsRead paper

Hybrid Semantic and Spectral Ensemble for Robust Synthetic Image Source Attribution

Jul 24, 2026

This work addresses the challenge of distribution shift and degraded source-tracing accuracy in synthetic images caused by unknown post-processing operations such as JPEG compression or blurring. To this end, the authors propose a dual-branch ensemble framework that efficiently integrates semantic deep learning—using EfficientNet-B0 enhanced with exponential moving average (EMA) and label smoothing—with lightweight mathematical forensic features derived from high-pass noise residuals. These handcrafted features comprise a 126-dimensional vector incorporating singular value decomposition (SVD) spectra, local binary patterns, and dimensionality reduction via truncated SVD. The combined representations are fused through an XGBoost classifier. Requiring only CPU computation, the method achieves 95.60% accuracy on a challenging benchmark where 55% of test samples undergo degradation, completing the entire pipeline within 6.5 hours. This approach demonstrates remarkable robustness to post-processing while offering high accuracy, computational efficiency, and practical deployability.

0 citationsRead paper

RMS@CC-MMD 2026: Multimodal Misogyny Detection via Geometric Interaction and Multi-View Consensus

Jul 20, 2026

This work addresses the challenge of detecting implicit misogynistic content in internet memes, which often relies on semantic incongruity between images and text as well as culturally specific contexts. To this end, the authors propose GeoMVC, a novel approach featuring a geometric interaction layer that dynamically models alignment between image and text embeddings using Hadamard product and cosine similarity, replacing conventional static concatenation. Additionally, a multi-view consensus mechanism is introduced to aggregate predictions from original, length-filtered, and English-translated text views, thereby mitigating distributional shifts caused by OCR noise and transcription errors. Evaluated on Task A of the ICMI 2026 CC-MMD Challenge, GeoMVC achieved second place for Malayalam (Macro F1: 0.892), third for Chinese (0.895), and fifth for Tamil (0.521), demonstrating its effectiveness and robustness in cross-lingual multimodal misogyny detection.

0 citationsRead paper

Transferable Low-Rank Convolutional Bases for Onboarding Unseen Medical Imaging Modalities

Jul 18, 2026

This work addresses the challenge of efficiently adapting medical imaging models to unseen modalities post-deployment while avoiding catastrophic forgetting. The authors propose a parameter-efficient adaptation method that, under a strict leave-one-domain-out setting, freezes the pre-trained convolutional backbone and leverages transferable low-rank convolutional bases learned from source modalities. Adaptation is achieved solely through projection parameters atop these bases, constituting only 0.78% of the total model parameters. By integrating convolutional LoRA, low-rank decomposition, and Mahalanobis anomaly detection, the approach improves adaptation accuracy on new modalities by 6.11 percentage points over random bases, while incurring zero performance loss on source modalities (Δ = 0.00 pp), substantially outperforming full fine-tuning and decision-level adaptation strategies.

0 citationsRead paper