Selected work

Representative Papers

End-to-End Autonomous Recursive Arborescence Deformable Flow and Non-Linear Hemodynamics for Patient-Specific Coronary Centerline Extraction

Oct 05, 2026

This study addresses the limitations of coronary centerline extraction, including severed bifurcations, non-anatomical shortcuts, and ischemia underestimation by linear flow models. We propose an end-to-end framework that decouples geometric tree generation from nonlinear hemodynamics. Methodologically, we introduce autonomous ostium localization to eliminate seed dependence, alongside a recursive tree state machine and Tree-NMS to ensure topological connectivity. Furthermore, a 3D feature pyramid deformable-flow architecture is integrated with a Picard-iterative nonlinear Kirchhoff solver to accurately capture functional ischemia. Experimental results demonstrate a mean localization error of 7.63 mm and mass conservation residuals reaching machine precision. Crucially, the proposed method successfully identifies ischemic lesions missed by linear models, significantly enhancing diagnostic sensitivity.

0 citationsRead paper

Beyond Diagonal State Space Models: Exact Non-Abelian Group Tracking, Solvability Barriers, and Geometric Physical Manifolds

Sep 29, 2026

This study addresses the limitations of diagonal state space models (SSMs), which cannot track non-abelian groups due to commutativity constraints and suffer from optimization degeneracy. To overcome these issues, this work proposes NC-SSM, elevating state transitions to compact Lie groups. Methodologically, it introduces Hopf fibration readouts, quaternion parallel scans, and identity gating mechanisms, combined with Euler–Rodrigues mapping and rational Cayley transforms, thereby transcending solvable group restrictions while eliminating sign ambiguities and phase drift. Experimental results demonstrate that the proposed model achieves 100% tracking accuracy on simple groups such as A5, surpassing the theoretical upper bound of S5. Furthermore, it attains Riemannian manifold errors below 3e-6, significantly outperforming existing baselines.

0 citationsRead paper

Port-Hamiltonian Latent Deliberation: Mitigating the Deliberation Drift Cliff in Test-Time Compute Scaling

Sep 29, 2026

This study addresses the "deliberation drift cliff" and deep-thinking collapse arising from iterative reasoning in continuous latent spaces during test-time compute scaling. To this end, it introduces DG-HHD, a novel port-Hamiltonian latent deliberation architecture. The proposed method leverages Helmholtz–Hodge decomposition to orthogonally decouple rotational and gradient flows, integrating tangential projection tensor networks with an RK45 integrator to eliminate automatic differentiation dependencies and overcome accuracy bottlenecks inherent in conservative flows. Experimental results demonstrate that DG-HHD achieves monotonic compute scaling for multi-hop reasoning in small language models, improving peak accuracy by 25.94% and accelerating inference by over 1.8× while effectively suppressing out-of-distribution drift.

0 citationsRead paper

An Adaptive Heterogeneous Architecture for High-Ratio, High-Throughput Lossless Compression

Sep 28, 2026

This study addresses the inherent conflict in lossless compression between industrial-grade high throughput and the superior compression ratios achieved by context-mixing algorithms. To reconcile this trade-off, we propose GPX, an adaptive heterogeneous architecture that integrates GPU-based invertible domain transforms, AVX2-accelerated multi-hypothesis sequence optimization, and a sub-15-microsecond structural decision probe to enable efficient compression of structured data streams. This approach establishes a Pareto frontier advantage under RFC 8878-compatible configurations. Evaluated on the Silesia benchmark, GPX achieves a compressed size of 54.46 MiB and a throughput of 191.90 MiB/s, strictly dominating the official Zstandard levels 10–14 while demonstrating zero-tuning generalization capability.

0 citationsRead paper

Beyond Embedding Transfer: Component Roles in Grokking Transfer and Stability

Sep 15, 2026

研究通过转移内部注意力/MLP权重及读出组件来加速算法任务的泛化，并探讨了继续优化时的稳定性问题。

0 citationsRead paper

Recent publications

Latest Papers

End-to-End Autonomous Recursive Arborescence Deformable Flow and Non-Linear Hemodynamics for Patient-Specific Coronary Centerline Extraction

Oct 05, 2026

This study addresses the limitations of coronary centerline extraction, including severed bifurcations, non-anatomical shortcuts, and ischemia underestimation by linear flow models. We propose an end-to-end framework that decouples geometric tree generation from nonlinear hemodynamics. Methodologically, we introduce autonomous ostium localization to eliminate seed dependence, alongside a recursive tree state machine and Tree-NMS to ensure topological connectivity. Furthermore, a 3D feature pyramid deformable-flow architecture is integrated with a Picard-iterative nonlinear Kirchhoff solver to accurately capture functional ischemia. Experimental results demonstrate a mean localization error of 7.63 mm and mass conservation residuals reaching machine precision. Crucially, the proposed method successfully identifies ischemic lesions missed by linear models, significantly enhancing diagnostic sensitivity.

0 citationsRead paper

Beyond Diagonal State Space Models: Exact Non-Abelian Group Tracking, Solvability Barriers, and Geometric Physical Manifolds

Sep 29, 2026

This study addresses the limitations of diagonal state space models (SSMs), which cannot track non-abelian groups due to commutativity constraints and suffer from optimization degeneracy. To overcome these issues, this work proposes NC-SSM, elevating state transitions to compact Lie groups. Methodologically, it introduces Hopf fibration readouts, quaternion parallel scans, and identity gating mechanisms, combined with Euler–Rodrigues mapping and rational Cayley transforms, thereby transcending solvable group restrictions while eliminating sign ambiguities and phase drift. Experimental results demonstrate that the proposed model achieves 100% tracking accuracy on simple groups such as A5, surpassing the theoretical upper bound of S5. Furthermore, it attains Riemannian manifold errors below 3e-6, significantly outperforming existing baselines.

0 citationsRead paper

Port-Hamiltonian Latent Deliberation: Mitigating the Deliberation Drift Cliff in Test-Time Compute Scaling

Sep 29, 2026

This study addresses the "deliberation drift cliff" and deep-thinking collapse arising from iterative reasoning in continuous latent spaces during test-time compute scaling. To this end, it introduces DG-HHD, a novel port-Hamiltonian latent deliberation architecture. The proposed method leverages Helmholtz–Hodge decomposition to orthogonally decouple rotational and gradient flows, integrating tangential projection tensor networks with an RK45 integrator to eliminate automatic differentiation dependencies and overcome accuracy bottlenecks inherent in conservative flows. Experimental results demonstrate that DG-HHD achieves monotonic compute scaling for multi-hop reasoning in small language models, improving peak accuracy by 25.94% and accelerating inference by over 1.8× while effectively suppressing out-of-distribution drift.

0 citationsRead paper

An Adaptive Heterogeneous Architecture for High-Ratio, High-Throughput Lossless Compression

Sep 28, 2026

This study addresses the inherent conflict in lossless compression between industrial-grade high throughput and the superior compression ratios achieved by context-mixing algorithms. To reconcile this trade-off, we propose GPX, an adaptive heterogeneous architecture that integrates GPU-based invertible domain transforms, AVX2-accelerated multi-hypothesis sequence optimization, and a sub-15-microsecond structural decision probe to enable efficient compression of structured data streams. This approach establishes a Pareto frontier advantage under RFC 8878-compatible configurations. Evaluated on the Silesia benchmark, GPX achieves a compressed size of 54.46 MiB and a throughput of 191.90 MiB/s, strictly dominating the official Zstandard levels 10–14 while demonstrating zero-tuning generalization capability.

0 citationsRead paper

Beyond Embedding Transfer: Component Roles in Grokking Transfer and Stability

Sep 15, 2026

研究通过转移内部注意力/MLP权重及读出组件来加速算法任务的泛化，并探讨了继续优化时的稳定性问题。

0 citationsRead paper