Selected work

Representative Papers

Diagnosing Generalization Failures in Fine-Tuned LLMs: A Cross-Architectural Study on Phishing Detection

Jan 15, 2026

This work addresses the challenge of diagnosing generalization failures in fine-tuned large language models (LLMs) for high-stakes tasks such as phishing detection. We propose a multi-layer diagnostic framework that integrates SHAP value analysis with mechanistic interpretability techniques to conduct a cross-architecture investigation of generalization behavior across diverse datasets, focusing on prominent models including Llama 3.1, Gemma 2, and Mistral. Our study uncovers a synergistic interaction between model architecture and data diversity, identifies architecture-dependent failure modes, and establishes a reproducible diagnostic paradigm. Experimental results demonstrate that Gemma 2 9B achieves an F1 score above 91% under data diversity, while Llama 3.1 8B exhibits significant performance degradation due to insufficient fusion capabilities; in contrast, Mistral displays robust generalization across varying training paradigms.

0 citationsRead paper

Data-Driven Extreme Response Estimation

Mar 27, 2025

Predicting extreme surge responses of ships under Sea State 5 (significant wave height (H_s = 4.0) m, mean period (T_m = 15.0) s) remains computationally expensive and inaccurate with conventional methods. To address this, we propose an extreme-event-oriented LSTM correction framework: it takes low-fidelity hydrodynamic predictions from SimpleCode as input and is supervised by high-fidelity nonlinear time-domain simulations from LAMP. Crucially, we introduce a novel temporal local weighting strategy that focuses training exclusively on short-duration segments around response peaks, enabling targeted optimization for extreme values. Compared to standard LSTM training, our method significantly improves prediction accuracy for extreme surge amplitudes—reducing peak error by approximately 42% under Sea State 5—while retaining the computational efficiency of the low-fidelity solver. This approach establishes a new paradigm for rapid, yet accurate, assessment of full-scale ship extreme responses in operational sea states.

0 citationsRead paper

Model-free system identification of surface ships in waves via Hankel dynamic mode decomposition with control

Feb 17, 2025

Modeling the motion of freely navigating vessels in irregular waves using physics-based equations remains challenging, particularly under data-limited conditions. To address this, this paper proposes a model-free, data-driven system identification and prediction framework. Leveraging hull states, historical wave height sequences, and rudder angle as inputs, it constructs a low-order dynamical model. A novel Bayesian Hankel-Dynamic Mode Decomposition with control (Hankel-DMDc) method is introduced, integrating delay embedding to enhance nonlinear representation and inherently quantifying parametric and predictive uncertainties. Validated under SS7-level beam–oblique irregular wave conditions, the method achieves high-accuracy motion prediction over 15 wave periods, with stable, non-decaying errors and low computational cost. This approach breaks away from traditional equation-dependent paradigms, offering a robust, data-driven foundation for ship design optimization and real-time navigation decision support.

0 citationsRead paper

Issues with Neural Tangent Kernel Approach to Neural Networks

Jan 19, 2025

This paper investigates whether the Neural Tangent Kernel (NTK) accurately characterizes the actual training dynamics of deep neural networks, particularly how its predictive error scales with network depth. Method: We conduct rigorous theoretical re-derivation, full-batch gradient descent simulations, trajectory tracking of generalization error across multilayer networks, and systematic comparisons against Gaussian process kernels. Contribution/Results: We find that NTK kernel regression exhibits significant deviation from the true training trajectories in both optimization and generalization behavior; remarkably, a simple Gaussian kernel achieves comparable performance, indicating that the NTK fails to deliver its theoretical advantages in practice. This work provides the first empirical evidence that the NTK’s theoretical equivalence to infinite-width networks breaks down substantially under standard training settings, challenging its ability to model path-dependent optimization effects. Our findings establish critical empirical bounds on the practical applicability of the NTK framework.

0 citationsRead paper

Recent publications

Latest Papers

Diagnosing Generalization Failures in Fine-Tuned LLMs: A Cross-Architectural Study on Phishing Detection

Jan 15, 2026

This work addresses the challenge of diagnosing generalization failures in fine-tuned large language models (LLMs) for high-stakes tasks such as phishing detection. We propose a multi-layer diagnostic framework that integrates SHAP value analysis with mechanistic interpretability techniques to conduct a cross-architecture investigation of generalization behavior across diverse datasets, focusing on prominent models including Llama 3.1, Gemma 2, and Mistral. Our study uncovers a synergistic interaction between model architecture and data diversity, identifies architecture-dependent failure modes, and establishes a reproducible diagnostic paradigm. Experimental results demonstrate that Gemma 2 9B achieves an F1 score above 91% under data diversity, while Llama 3.1 8B exhibits significant performance degradation due to insufficient fusion capabilities; in contrast, Mistral displays robust generalization across varying training paradigms.

0 citationsRead paper

Data-Driven Extreme Response Estimation

Mar 27, 2025

Predicting extreme surge responses of ships under Sea State 5 (significant wave height (H_s = 4.0) m, mean period (T_m = 15.0) s) remains computationally expensive and inaccurate with conventional methods. To address this, we propose an extreme-event-oriented LSTM correction framework: it takes low-fidelity hydrodynamic predictions from SimpleCode as input and is supervised by high-fidelity nonlinear time-domain simulations from LAMP. Crucially, we introduce a novel temporal local weighting strategy that focuses training exclusively on short-duration segments around response peaks, enabling targeted optimization for extreme values. Compared to standard LSTM training, our method significantly improves prediction accuracy for extreme surge amplitudes—reducing peak error by approximately 42% under Sea State 5—while retaining the computational efficiency of the low-fidelity solver. This approach establishes a new paradigm for rapid, yet accurate, assessment of full-scale ship extreme responses in operational sea states.

0 citationsRead paper

Model-free system identification of surface ships in waves via Hankel dynamic mode decomposition with control

Feb 17, 2025

Modeling the motion of freely navigating vessels in irregular waves using physics-based equations remains challenging, particularly under data-limited conditions. To address this, this paper proposes a model-free, data-driven system identification and prediction framework. Leveraging hull states, historical wave height sequences, and rudder angle as inputs, it constructs a low-order dynamical model. A novel Bayesian Hankel-Dynamic Mode Decomposition with control (Hankel-DMDc) method is introduced, integrating delay embedding to enhance nonlinear representation and inherently quantifying parametric and predictive uncertainties. Validated under SS7-level beam–oblique irregular wave conditions, the method achieves high-accuracy motion prediction over 15 wave periods, with stable, non-decaying errors and low computational cost. This approach breaks away from traditional equation-dependent paradigms, offering a robust, data-driven foundation for ship design optimization and real-time navigation decision support.

0 citationsRead paper

Issues with Neural Tangent Kernel Approach to Neural Networks

Jan 19, 2025

This paper investigates whether the Neural Tangent Kernel (NTK) accurately characterizes the actual training dynamics of deep neural networks, particularly how its predictive error scales with network depth. Method: We conduct rigorous theoretical re-derivation, full-batch gradient descent simulations, trajectory tracking of generalization error across multilayer networks, and systematic comparisons against Gaussian process kernels. Contribution/Results: We find that NTK kernel regression exhibits significant deviation from the true training trajectories in both optimization and generalization behavior; remarkably, a simple Gaussian kernel achieves comparable performance, indicating that the NTK fails to deliver its theoretical advantages in practice. This work provides the first empirical evidence that the NTK’s theoretical equivalence to infinite-width networks breaks down substantially under standard training settings, challenging its ability to model path-dependent optimization effects. Our findings establish critical empirical bounds on the practical applicability of the NTK framework.

0 citationsRead paper