Selected work

Representative Papers

Delta-Position Estimation-Based IMU Odometry: A Comparison of MLP and Kolmogorov-Arnold Networks

Jun 24, 2026

This work addresses the issue of cumulative error in learning-based inertial odometry caused by direct regression of absolute position. To mitigate this, the authors propose estimating incremental displacement (Δp) over a 50 ms sliding window and reconstructing the trajectory via numerical integration. The study introduces Kolmogorov–Arnold Networks (KANs) into IMU odometry for the first time, leveraging their learnable B-spline activation functions to effectively suppress long-term error accumulation. Experiments on the EuRoC MAV dataset demonstrate that, compared to conventional multilayer perceptrons (MLPs), KANs reduce cumulative drift error by 44% while using only approximately one-sixth the number of parameters, and exhibit superior stability under both P₅₀ and P₉₀ metrics.

0 citationsRead paper

SPD-RAG: Sub-Agent Per Document Retrieval-Augmented Generation

Mar 09, 2026

This work addresses the challenge of answering complex real-world queries that require synthesizing information across numerous heterogeneous documents—a task where traditional retrieval-augmented generation (RAG) systems suffer from insufficient evidence coverage and long-context large language models exhibit unreliable reasoning. To overcome these limitations, the authors propose a hierarchical multi-agent framework in which each document is assigned a dedicated sub-agent for focused retrieval, while a coordinator distributes tasks and aggregates local answers. A token-bounded recursive synthesis layer then integrates these partial results. This architecture introduces, for the first time, a document-wise sub-agent design that enables modular and scalable multi-document question answering, achieving both broad coverage and enhanced reasoning accuracy. Evaluated on the LOONG benchmark, the method attains an average score of 58.1, substantially outperforming standard RAG (33.0) and Agentic RAG (32.8), with only 38% of the API cost of full-context baselines.

0 citationsRead paper

RAGSmith: A Framework for Finding the Optimal Composition of Retrieval-Augmented Generation Methods Across Datasets

Nov 03, 2025

Traditional RAG systems suffer from suboptimal performance due to tight coupling among retrieval, reranking, prompt rewriting, and generation modules, hindering holistic optimization. Method: This paper proposes the first end-to-end RAG architecture search framework, modeling the RAG configuration space as an evolvable search problem and employing genetic algorithms to jointly optimize multi-objective metrics—including recall@k/nDCG for retrieval and LLM-Judge/semantic similarity for generation. The framework encompasses nine component types across vector retrieval, reranking, and prompt rewriting. Contribution/Results: Evaluated across six domains, the framework achieves an average 3.8% performance gain (up to +12.5% in retrieval, +7.5% in generation) while converging after exploring only 0.2% of the configuration space. It identifies robust architectural patterns transferable across datasets and quantifies how domain characteristics and question types systematically influence optimal configurations.

0 citationsRead paper

Artificial Intelligence Based Predictive Maintenance for Electric Buses

Oct 27, 2025

Electric bus powertrain and battery systems exhibit high complexity, rendering conventional time-based maintenance inadequate for real-time detection of multidimensional CAN bus anomalies. Method: This paper proposes a hybrid graph-structured feature selection method integrating statistical filtering and optimized community detection to enhance identification of critical fault parameters and model interpretability. It first screens features using Pearson correlation, Cramér’s V, and ANOVA; constructs a parameter association graph via InfoMap; and employs an ensemble of SVM, Random Forest, and XGBoost. Data imbalance is addressed via SMOTEEN and binary undersampling; LIME augments model explainability. Contribution/Results: Experiments demonstrate significant improvements in alarm prediction accuracy and fault feature localization precision, alongside reduced response latency. The approach validates the feasibility and industrial applicability of AI-driven predictive maintenance in real-world public transit operations.

0 citationsRead paper

A Large-Scale Dataset and Citation Intent Classification in Turkish with LLMs

Sep 26, 2025

Addressing the challenge of citation intent classification in agglutinative languages such as Turkish, this paper introduces the first large-scale, manually annotated, publicly available Turkish citation intent dataset. We propose a programmable classification framework built on DSPy, which jointly leverages automatic prompt optimization and in-context learning to enhance generalization, while integrating multi-model feature extraction with an XGBoost meta-learner for stacked ensemble classification. Our approach achieves 91.3% accuracy on the held-out test set—significantly outperforming handcrafted prompt baselines. This work constitutes the first high-accuracy, reproducible solution for automated citation intent recognition in Turkish, establishing a new benchmark and methodological paradigm for scholarly text analysis in low-resource languages.

0 citationsRead paper

Recent publications

Latest Papers

Delta-Position Estimation-Based IMU Odometry: A Comparison of MLP and Kolmogorov-Arnold Networks

Jun 24, 2026

This work addresses the issue of cumulative error in learning-based inertial odometry caused by direct regression of absolute position. To mitigate this, the authors propose estimating incremental displacement (Δp) over a 50 ms sliding window and reconstructing the trajectory via numerical integration. The study introduces Kolmogorov–Arnold Networks (KANs) into IMU odometry for the first time, leveraging their learnable B-spline activation functions to effectively suppress long-term error accumulation. Experiments on the EuRoC MAV dataset demonstrate that, compared to conventional multilayer perceptrons (MLPs), KANs reduce cumulative drift error by 44% while using only approximately one-sixth the number of parameters, and exhibit superior stability under both P₅₀ and P₉₀ metrics.

0 citationsRead paper

SPD-RAG: Sub-Agent Per Document Retrieval-Augmented Generation

Mar 09, 2026

This work addresses the challenge of answering complex real-world queries that require synthesizing information across numerous heterogeneous documents—a task where traditional retrieval-augmented generation (RAG) systems suffer from insufficient evidence coverage and long-context large language models exhibit unreliable reasoning. To overcome these limitations, the authors propose a hierarchical multi-agent framework in which each document is assigned a dedicated sub-agent for focused retrieval, while a coordinator distributes tasks and aggregates local answers. A token-bounded recursive synthesis layer then integrates these partial results. This architecture introduces, for the first time, a document-wise sub-agent design that enables modular and scalable multi-document question answering, achieving both broad coverage and enhanced reasoning accuracy. Evaluated on the LOONG benchmark, the method attains an average score of 58.1, substantially outperforming standard RAG (33.0) and Agentic RAG (32.8), with only 38% of the API cost of full-context baselines.

0 citationsRead paper

RAGSmith: A Framework for Finding the Optimal Composition of Retrieval-Augmented Generation Methods Across Datasets

Nov 03, 2025

Traditional RAG systems suffer from suboptimal performance due to tight coupling among retrieval, reranking, prompt rewriting, and generation modules, hindering holistic optimization. Method: This paper proposes the first end-to-end RAG architecture search framework, modeling the RAG configuration space as an evolvable search problem and employing genetic algorithms to jointly optimize multi-objective metrics—including recall@k/nDCG for retrieval and LLM-Judge/semantic similarity for generation. The framework encompasses nine component types across vector retrieval, reranking, and prompt rewriting. Contribution/Results: Evaluated across six domains, the framework achieves an average 3.8% performance gain (up to +12.5% in retrieval, +7.5% in generation) while converging after exploring only 0.2% of the configuration space. It identifies robust architectural patterns transferable across datasets and quantifies how domain characteristics and question types systematically influence optimal configurations.

0 citationsRead paper

Artificial Intelligence Based Predictive Maintenance for Electric Buses

Oct 27, 2025

Electric bus powertrain and battery systems exhibit high complexity, rendering conventional time-based maintenance inadequate for real-time detection of multidimensional CAN bus anomalies. Method: This paper proposes a hybrid graph-structured feature selection method integrating statistical filtering and optimized community detection to enhance identification of critical fault parameters and model interpretability. It first screens features using Pearson correlation, Cramér’s V, and ANOVA; constructs a parameter association graph via InfoMap; and employs an ensemble of SVM, Random Forest, and XGBoost. Data imbalance is addressed via SMOTEEN and binary undersampling; LIME augments model explainability. Contribution/Results: Experiments demonstrate significant improvements in alarm prediction accuracy and fault feature localization precision, alongside reduced response latency. The approach validates the feasibility and industrial applicability of AI-driven predictive maintenance in real-world public transit operations.

0 citationsRead paper

A Large-Scale Dataset and Citation Intent Classification in Turkish with LLMs

Sep 26, 2025

Addressing the challenge of citation intent classification in agglutinative languages such as Turkish, this paper introduces the first large-scale, manually annotated, publicly available Turkish citation intent dataset. We propose a programmable classification framework built on DSPy, which jointly leverages automatic prompt optimization and in-context learning to enhance generalization, while integrating multi-model feature extraction with an XGBoost meta-learner for stacked ensemble classification. Our approach achieves 91.3% accuracy on the held-out test set—significantly outperforming handcrafted prompt baselines. This work constitutes the first high-accuracy, reproducible solution for automated citation intent recognition in Turkish, establishing a new benchmark and methodological paradigm for scholarly text analysis in low-resource languages.

0 citationsRead paper