Selected work

Representative Papers

HoloAegis: Frozen Representation, Topological Inference: Minimally Parametric Safety Manifolds for Zero-Shot LLM Guardrails

Aug 09, 2026

This work addresses the dilemma faced by current large language models, where safety alignment either distorts semantic representations through fine-tuning or incurs high inference costs. The authors propose a training-free geometric safety mechanism that freezes the pretrained encoder and maps text embeddings onto the unit hypersphere. Leveraging a precomputed library of topological anchor points, the method performs zero-shot safety judgments via Gibbs–Boltzmann free energy and a dual-timescale exponential moving average, effectively decoupling representation learning from inference. Requiring only a few fixed hyperparameters, the approach significantly enhances robustness against high-frequency perturbations and achieves state-of-the-art performance across eight benchmarks—e.g., AuthenHallu AUC = 1.0000 and HarmBench AUC = 0.9802—while offering sub-millisecond latency, zero cold-start overhead, and strong cross-lingual transferability, as demonstrated by CHIFRAUD AUC = 0.9758 on Chinese data.

0 citationsRead paper

Carrier Leakage Suppression and Power Difference Equalization for TFDMA Coherent PON

Aug 08, 2026

This work addresses the challenges of carrier leakage, inter-user power disparity, and semiconductor optical amplifier (SOA)-induced nonlinear distortions in TFDMA coherent passive optical networks. For the first time, the SOA is leveraged not only to suppress carrier leakage and equalize multi-user power levels but also in conjunction with a tailored digital signal processing algorithm designed specifically for nonlinear compensation. The proposed approach effectively mitigates carrier leakage, balances power variations among users, and significantly alleviates the performance degradation caused by SOA nonlinearities. By simultaneously tackling these critical impairments, the method offers substantial improvements in system stability and transmission quality, representing a notable advancement in the design and optimization of coherent passive optical access networks.

0 citationsRead paper

HFS-TriNet: A Three-Branch Collaborative Feature Learning Network for Prostate Cancer Classification from TRUS Videos

Apr 24, 2026

This study addresses the challenges in prostate cancer classification from transrectal ultrasound (TRUS) videos, including information redundancy, high intra- and inter-class similarity, and low signal-to-noise ratio, which hinder feature discriminability and diagnostic accuracy. To overcome these limitations, the authors propose HFS-TriNet, a novel architecture that integrates three parallel branches—leveraging the medical Segment Anything Model (SAM), wavelet-transform convolutional residual (WTCR) blocks, and ResNet50—augmented with a heuristic frame selection (HFS) mechanism and a normalized attention module. This design enables efficient extraction of multi-scale features that jointly capture edge details, semantic consistency, and spatiotemporal dynamics. The proposed method substantially mitigates redundancy and noise interference while maintaining computational efficiency, leading to significantly improved accuracy and robustness in prostate cancer classification.

0 citationsRead paper

KoopmanFlow: Spectrally Decoupled Generative Control Policy via Koopman Structural Bias

Mar 14, 2026

This work addresses the challenge that existing generative control policies struggle to simultaneously maintain stable global motion and perform high-frequency local corrections, as unified time integration often smooths out transient details. To overcome this, the authors propose a spectrally decoupled generative control architecture that incorporates a Koopman structural prior within a unified multimodal latent space. The macro branch models slowly varying trajectories via single-step consistency training, while the transient branch captures high-frequency residuals induced by visual discontinuities—such as contacts or occlusions—using flow matching. An asymmetric consistency objective enables joint modeling of low- and high-frequency dynamics. This approach avoids error accumulation across multiple stages and significantly outperforms current methods in contact-rich, disturbance-sensitive tasks, achieving both high control accuracy and parameter efficiency under real-time deployment constraints.

0 citationsRead paper

Learning Implicit Neural Degradation Representation for Unpaired Image Dehazing

Nov 17, 2025

Addressing the challenge of jointly modeling non-uniform haze distribution and preserving global consistency in complex scenes, this paper proposes an unsupervised image dehazing method. We introduce the Kolmogorov–Arnold representation theorem into visual degradation modeling for the first time, constructing a continuous haze concentration field via implicit neural representations (INRs), thereby eliminating reliance on explicit feature extraction or handcrafted physical priors. To enhance structural fidelity while suppressing redundancy, we further design a channel-decoupled learning mechanism and a densely connected residual enhancement module. The method operates without paired training data and achieves state-of-the-art performance across multiple public and real-world hazy datasets. Notably, it demonstrates superior robustness and reconstruction fidelity in challenging scenarios characterized by highly non-uniform haze distribution and complex illumination conditions.

0 citationsRead paper

Recent publications

Latest Papers

HoloAegis: Frozen Representation, Topological Inference: Minimally Parametric Safety Manifolds for Zero-Shot LLM Guardrails

Aug 09, 2026

This work addresses the dilemma faced by current large language models, where safety alignment either distorts semantic representations through fine-tuning or incurs high inference costs. The authors propose a training-free geometric safety mechanism that freezes the pretrained encoder and maps text embeddings onto the unit hypersphere. Leveraging a precomputed library of topological anchor points, the method performs zero-shot safety judgments via Gibbs–Boltzmann free energy and a dual-timescale exponential moving average, effectively decoupling representation learning from inference. Requiring only a few fixed hyperparameters, the approach significantly enhances robustness against high-frequency perturbations and achieves state-of-the-art performance across eight benchmarks—e.g., AuthenHallu AUC = 1.0000 and HarmBench AUC = 0.9802—while offering sub-millisecond latency, zero cold-start overhead, and strong cross-lingual transferability, as demonstrated by CHIFRAUD AUC = 0.9758 on Chinese data.

0 citationsRead paper

Carrier Leakage Suppression and Power Difference Equalization for TFDMA Coherent PON

Aug 08, 2026

This work addresses the challenges of carrier leakage, inter-user power disparity, and semiconductor optical amplifier (SOA)-induced nonlinear distortions in TFDMA coherent passive optical networks. For the first time, the SOA is leveraged not only to suppress carrier leakage and equalize multi-user power levels but also in conjunction with a tailored digital signal processing algorithm designed specifically for nonlinear compensation. The proposed approach effectively mitigates carrier leakage, balances power variations among users, and significantly alleviates the performance degradation caused by SOA nonlinearities. By simultaneously tackling these critical impairments, the method offers substantial improvements in system stability and transmission quality, representing a notable advancement in the design and optimization of coherent passive optical access networks.

0 citationsRead paper

HFS-TriNet: A Three-Branch Collaborative Feature Learning Network for Prostate Cancer Classification from TRUS Videos

Apr 24, 2026

This study addresses the challenges in prostate cancer classification from transrectal ultrasound (TRUS) videos, including information redundancy, high intra- and inter-class similarity, and low signal-to-noise ratio, which hinder feature discriminability and diagnostic accuracy. To overcome these limitations, the authors propose HFS-TriNet, a novel architecture that integrates three parallel branches—leveraging the medical Segment Anything Model (SAM), wavelet-transform convolutional residual (WTCR) blocks, and ResNet50—augmented with a heuristic frame selection (HFS) mechanism and a normalized attention module. This design enables efficient extraction of multi-scale features that jointly capture edge details, semantic consistency, and spatiotemporal dynamics. The proposed method substantially mitigates redundancy and noise interference while maintaining computational efficiency, leading to significantly improved accuracy and robustness in prostate cancer classification.

0 citationsRead paper

KoopmanFlow: Spectrally Decoupled Generative Control Policy via Koopman Structural Bias

Mar 14, 2026

This work addresses the challenge that existing generative control policies struggle to simultaneously maintain stable global motion and perform high-frequency local corrections, as unified time integration often smooths out transient details. To overcome this, the authors propose a spectrally decoupled generative control architecture that incorporates a Koopman structural prior within a unified multimodal latent space. The macro branch models slowly varying trajectories via single-step consistency training, while the transient branch captures high-frequency residuals induced by visual discontinuities—such as contacts or occlusions—using flow matching. An asymmetric consistency objective enables joint modeling of low- and high-frequency dynamics. This approach avoids error accumulation across multiple stages and significantly outperforms current methods in contact-rich, disturbance-sensitive tasks, achieving both high control accuracy and parameter efficiency under real-time deployment constraints.

0 citationsRead paper

Learning Implicit Neural Degradation Representation for Unpaired Image Dehazing

Nov 17, 2025

Addressing the challenge of jointly modeling non-uniform haze distribution and preserving global consistency in complex scenes, this paper proposes an unsupervised image dehazing method. We introduce the Kolmogorov–Arnold representation theorem into visual degradation modeling for the first time, constructing a continuous haze concentration field via implicit neural representations (INRs), thereby eliminating reliance on explicit feature extraction or handcrafted physical priors. To enhance structural fidelity while suppressing redundancy, we further design a channel-decoupled learning mechanism and a densely connected residual enhancement module. The method operates without paired training data and achieves state-of-the-art performance across multiple public and real-world hazy datasets. Notably, it demonstrates superior robustness and reconstruction fidelity in challenging scenarios characterized by highly non-uniform haze distribution and complex illumination conditions.

0 citationsRead paper