Selected work

Representative Papers

Survey on Publicly Available Sinhala Natural Language Processing Tools and Research

Jun 05, 2019arXiv.org

The Sinhala NLP community suffers from fragmented resources, absence of systematic surveys, and lack of collaborative benchmarks. Method: This paper introduces the first dynamically updated, open-source panoramic survey of Sinhala NLP. Leveraging bibliometric analysis, automated crawling and classification of open-source tools, multidimensional metadata annotation, and continuous arXiv tracking, it systematically catalogs dozens of global Sinhala NLP projects and tools. Contribution/Results: It identifies critical technical gaps and reuse pathways, and innovatively establishes a sustainably maintained knowledge graph and collaborative benchmark suite—filling a key void in unified surveys for low-resource language NLP. The survey significantly enhances community visibility, reproducibility, and interoperability, and has become the central reference and coordination hub for Sinhala NLP researchers in Sri Lanka and worldwide.

39 citations3 influentialRead paper

Comparing the Performance of LLMs in RAG-based Question-Answering: A Case Study in Computer Science Literature

Nov 05, 2025

This study systematically evaluates performance disparities among diverse large language models (LLMs) in retrieval-augmented generation (RAG) question answering over computer science literature. We benchmark both open-source (Mistral-7B-Instruct, Orca-mini-v3-7B, LLaMA2-7B-Chat) and closed-source (GPT-3.5, Gemini) LLMs within a unified RAG framework, measuring accuracy, human evaluation scores, cosine similarity to reference answers, and response latency. Results show that Mistral-7B-Instruct and GPT-3.5 achieve top-tier answer accuracy post-RAG enhancement; Orca-mini-v3-7B exhibits the lowest latency, whereas LLaMA2-7B-Chat incurs the highest; and open-source models collectively attain performance comparable to commercial closed-source counterparts. To our knowledge, this is the first fine-grained empirical analysis of RAG–LLM co-performance specifically on CS literature, revealing fundamental accuracy–latency trade-offs. The study establishes a reproducible, domain-specific benchmark for informed model selection in CS-oriented RAG system design.

1 citationsRead paper

Enhancing Small Dataset Classification Using Projected Quantum Kernels with Convolutional Neural Networks

Dec 18, 20242024 8th SLAAI International Conference on Artificial Intelligence (SLAAI-ICAI)

This work addresses the significant performance limitations of conventional convolutional neural networks (CNNs) in few-shot image classification due to extreme data scarcity. To overcome this challenge, the authors propose a novel architecture that integrates a projected quantum kernel (PQK) into the CNN’s feature extraction module for the first time, leveraging the quantum kernel’s expressive power in capturing high-dimensional and complex data structures to enhance representational capacity under minimal training data. Experimental results demonstrate that with only 1,000 training samples, the proposed method achieves 95% accuracy on MNIST and 90% on CIFAR-10, substantially outperforming classical CNNs, which attain merely 60% and 12% accuracy, respectively. This approach effectively breaks through the performance bottleneck typically encountered in data-scarce scenarios.

1 citationsRead paper

Deep Learning Based Relative Transfer Matrix Estimation for Multiple Sources and Multiple Microphones

Aug 12, 2026

This study addresses the challenge of accurately estimating the relative transfer matrix (ReTM) in multi-source, multi-microphone scenarios, where conventional covariance-based methods suffer from performance limitations that hinder speech enhancement in noisy environments. To overcome this, the work introduces, for the first time, a deep learning approach to ReTM estimation, proposing an end-to-end trainable supervised learning framework. The model employs time-domain convolution, short-time Fourier transform (STFT)-domain convolution, and LSTM networks to directly learn the ReTM from multi-channel recordings. By circumventing the reliance on covariance matrices inherent in traditional methods, the proposed approach achieves significantly better performance across five objective metrics and demonstrates speech enhancement results comparable to state-of-the-art baselines.

0 citationsRead paper

Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages

Aug 04, 2026

This study investigates how language models implement conditional rules of the form “if P(x) then A else B,” specifically examining whether such behavior relies on dissociable “test” and “routing” modules. Through activation patching, residual stream analysis, and cross-lingual and cross-rule transfer experiments across multiple models and languages, the authors systematically probe the underlying mechanisms. They find that predicate truth values are locally encoded in intermediate residual layers, with test signals exhibiting high stability and strong transferability across predicate families. In contrast, routing directions are highly token-dependent and generally non-transferable across models, except in Gemma-3-4B, which demonstrates remarkable cross-lingual consistency (≈0.98). This work provides the first empirical evidence for the separability of the “test” component in conditional reasoning and reveals that “routing” does not operate as a universal abstract mechanism.

0 citationsRead paper

Recent publications

Latest Papers

Deep Learning Based Relative Transfer Matrix Estimation for Multiple Sources and Multiple Microphones

Aug 12, 2026

This study addresses the challenge of accurately estimating the relative transfer matrix (ReTM) in multi-source, multi-microphone scenarios, where conventional covariance-based methods suffer from performance limitations that hinder speech enhancement in noisy environments. To overcome this, the work introduces, for the first time, a deep learning approach to ReTM estimation, proposing an end-to-end trainable supervised learning framework. The model employs time-domain convolution, short-time Fourier transform (STFT)-domain convolution, and LSTM networks to directly learn the ReTM from multi-channel recordings. By circumventing the reliance on covariance matrices inherent in traditional methods, the proposed approach achieves significantly better performance across five objective metrics and demonstrates speech enhancement results comparable to state-of-the-art baselines.

0 citationsRead paper

Test, then Route: How Language Models Execute In-Context Conditional Rules Across Models and Languages

Aug 04, 2026

This study investigates how language models implement conditional rules of the form “if P(x) then A else B,” specifically examining whether such behavior relies on dissociable “test” and “routing” modules. Through activation patching, residual stream analysis, and cross-lingual and cross-rule transfer experiments across multiple models and languages, the authors systematically probe the underlying mechanisms. They find that predicate truth values are locally encoded in intermediate residual layers, with test signals exhibiting high stability and strong transferability across predicate families. In contrast, routing directions are highly token-dependent and generally non-transferable across models, except in Gemma-3-4B, which demonstrates remarkable cross-lingual consistency (≈0.98). This work provides the first empirical evidence for the separability of the “test” component in conditional reasoning and reveals that “routing” does not operate as a universal abstract mechanism.

0 citationsRead paper

Multimodal Skin Lesion Classification with Swin Transformer and Clinical Metadata Fusion

Aug 04, 2026

This work addresses the diagnostic challenges in skin lesion classification arising from class imbalance, inter-class similarity, and intra-class variability by proposing a multimodal approach that fuses dermoscopic images with structured clinical metadata. Visual features are extracted using a Swin Transformer and jointly learned with contextual clinical information. To enhance model reliability and interpretability, the framework incorporates temperature scaling calibration and uncertainty estimation. Evaluated on public datasets, the method achieves an accuracy of 92.55% and a macro F1-score of 91.33%, demonstrating significantly improved recognition of minority classes and reduced calibration error. These advances collectively strengthen the trustworthiness and practical utility of automated dermatological diagnosis systems.

0 citationsRead paper

Does Explainability Transfer? A Controlled Benchmark of Attribution Methods on Vision Transformers and CNNs

Aug 03, 2026

This work addresses the limited generalizability of existing explainable AI attribution methods, which are predominantly validated on CNNs and lack systematic evaluation on modern architectures like Vision Transformers. The authors establish a standardized benchmark to comprehensively assess 13 attribution methods across five dimensions—faithfulness, localization, robustness, complexity, and computational cost—on eight representative visual backbones, including CNNs, isotropic ViTs, hierarchical Transformers, hybrid models, and linear attention architectures. Their analysis reveals that attribution performance is highly architecture-dependent, challenging the transferability of CNN-based conclusions to Transformer-based models. They propose a multidimensional, architecture-aware evaluation paradigm that questions conventional single-metric assessments. Notably, CAM-based methods excel on traditional metrics yet suffer from saturation and fail on linear attention models; attention rollout demonstrates robustness but poor localization; and faithfulness alone proves insufficient for discriminating method quality.

0 citationsRead paper

VespaSeg: A Resource-Aware Ground-then-Segment Pipeline for Referring Expression Segmentation

Aug 02, 2026

This work addresses the high deployment cost of monolithic models in referring expression segmentation by proposing a resource-aware, two-stage framework. It first employs a lightweight vision-language model (Florence-2 or Moondream2) for text-guided object localization and then leverages MobileSAM to convert the predicted bounding box into a pixel-level mask, with both stages fine-tuned independently. Evaluated on the RefCOCO validation set, the method achieves 73.64 mIoU and 84.60 P@0.5, requiring only 2.20 GB of GPU memory per query and attaining a throughput of 22.8 FPS on an RTX 6000 Ada GPU. These results demonstrate that the proposed output token compression and ground-truth box adaptation strategies effectively maintain high accuracy while substantially reducing computational overhead.

0 citationsRead paper