Selected work

Representative Papers

Benchmarking the CoW with the TopCoW Challenge: Topology-Aware Anatomical Segmentation of the Circle of Willis for CTA and MRA

Dec 29, 2023arXiv.org

The Circle of Willis (CoW) suffers from a scarcity of high-quality voxel-level annotations in CTA/MRA imaging, reliance on labor-intensive expert manual segmentation, and poor guarantee of topological consistency. Method: We introduce the first publicly available voxel-level multi-class CoW dataset—comprising 13 vascular structures with paired MRA/CTA volumes—and propose a topology-aware segmentation framework: (i) a novel VR-assisted annotation paradigm ensuring anatomical plausibility; (ii) a multimodal registration and topology-constrained segmentation network; and (iii) topology-sensitive metrics including branch F1 and topo-Dice. Contribution/Results: This benchmark has attracted >140 teams across four continents. State-of-the-art models achieve ≈90% Dice on most arterial branches, while exposing persistent topological matching bottlenecks—particularly for communicating arteries and anatomical variants.

24 citations2 influentialRead paper

CapTrack: Multifaceted Evaluation of Forgetting in LLM Post-Training

Feb 19, 2026

This work addresses the systematic capability degradation observed in large language models during post-training, which extends well beyond conventional notions of "knowledge forgetting." The authors propose CapTrack, a framework that redefines forgetting as systematic behavioral drift and introduces the first capability-centric, multidimensional evaluation system. Combining behavioral taxonomies with capability-specific metrics, they conduct large-scale experiments across models up to 80B parameters, spanning multiple algorithms, domains, and model families. Their analysis reveals that forgetting substantially impairs robustness and default behaviors; instruction tuning induces the strongest drift, whereas preference optimization is comparatively conservative and partially reversible. Notably, different model families exhibit markedly distinct forgetting patterns, indicating that no universal mitigation strategy currently exists.

1 citationsRead paper

Towards Identifiability of Interventional Stochastic Differential Equations

May 21, 2025

This work addresses the structural identifiability of parameters in stochastic differential equation (SDE) models under multiple interventions—i.e., whether SDE parameters can be uniquely recovered from samples of post-intervention stationary distributions. Theoretically, we establish the first uniqueness guarantee for SDE parameter recovery under multi-intervention settings; for linear SDEs, we derive a tight lower bound on the minimum number of required interventions; for weak-noise nonlinear SDEs, we obtain an upper bound on identifiability. Methodologically, we propose a parametric framework featuring learnable activation functions, integrating intervention modeling, stationary distribution analysis, and weak-noise asymptotic theory. Experiments on synthetic data demonstrate that our approach accurately recovers ground-truth parameters, and the theory-guided learnable architecture significantly improves both estimation accuracy and robustness.

1 citationsRead paper

NervePool: A Simplicial Pooling Layer

May 10, 2023arXiv.org

Existing graph pooling methods fail to preserve higher-order combinatorial and topological consistency when applied to simplicial complexes—topological data structures capable of encoding high-order relational information. Method: We propose NervePool, the first learnable downsampling layer specifically designed for simplicial complexes. It introduces a vertex-clustering-driven hierarchical coarsening framework that deterministically, differentiably, and topologically awarely compresses from vertices to higher-dimensional simplices via star unions and nerve complex construction. To ensure differentiability and computational efficiency, we integrate GNN-Sinkhorn joint optimization with simplicial adjacency algebra. Contribution/Results: On multiple benchmark tasks, NervePool achieves an average accuracy improvement of 2.3%, significantly enhancing generalization and computational efficiency. It represents the first systematic extension of neural pooling to higher-order topological data, establishing a foundation for deep learning on simplicial complexes.

1 citationsRead paper

You Only Flow Once: Calibrated and Real-Time Radar Pose Estimation with Multi-Hypothesis Normalizing Flows

Aug 10, 2026

Millimeter-wave radar point clouds are inherently sparse and noisy, leading to ambiguities in human pose estimation that deterministic methods struggle to capture. To address this challenge, this work proposes the Multi-Hypothesis Normalizing Flow Pose Generator (MH-NFPG), which introduces conditional normalizing flows into radar-based pose estimation for the first time. By integrating a spatio-temporal Transformer backbone with a Laplace base distribution, MH-NFPG generates diverse, multimodal pose hypotheses in parallel through a single forward pass. Evaluated on three benchmarks—MM-Fi, mmRadPose, and mRI—the method reduces calibration error by up to 85%, achieves over 20× faster inference, and demonstrates significantly improved pose accuracy on two datasets while matching state-of-the-art performance on the third. Moreover, it maintains reliable coverage under cross-environment settings.

0 citationsRead paper

Recent publications

Latest Papers

You Only Flow Once: Calibrated and Real-Time Radar Pose Estimation with Multi-Hypothesis Normalizing Flows

Aug 10, 2026

Millimeter-wave radar point clouds are inherently sparse and noisy, leading to ambiguities in human pose estimation that deterministic methods struggle to capture. To address this challenge, this work proposes the Multi-Hypothesis Normalizing Flow Pose Generator (MH-NFPG), which introduces conditional normalizing flows into radar-based pose estimation for the first time. By integrating a spatio-temporal Transformer backbone with a Laplace base distribution, MH-NFPG generates diverse, multimodal pose hypotheses in parallel through a single forward pass. Evaluated on three benchmarks—MM-Fi, mmRadPose, and mRI—the method reduces calibration error by up to 85%, achieves over 20× faster inference, and demonstrates significantly improved pose accuracy on two datasets while matching state-of-the-art performance on the third. Moreover, it maintains reliable coverage under cross-environment settings.

0 citationsRead paper

CUPA-T2*: Covariance-Aware Uncertainty Propagation and Alignment for T2* Mapping in Accelerated MRI

Aug 09, 2026

This work addresses the degradation in T2* quantification accuracy in accelerated MRI caused by artifacts and noise from undersampled reconstruction, compounded by the lack of effective modeling of uncertainty propagation to downstream fitting. The study introduces the first explicit covariance-aware uncertainty propagation framework from reconstruction to T2* fitting in accelerated T2* mapping. It estimates voxel-wise multi-echo reconstruction uncertainty via Monte Carlo Dropout, propagates this uncertainty to T2* fitting using a covariance-aware sampling strategy, and aligns predicted variances with reconstruction uncertainties through a heteroscedastic MLP and a correlation-based regularization term. Experiments demonstrate that the proposed method significantly improves T2* fitting performance in white matter under high acceleration in brain MRI, enhances uncertainty consistency, and yields interpretable voxel-level uncertainty maps.

0 citationsRead paper

Learning Biomechanically Plausible Human Motion from Sparse Radar Point Clouds

Aug 04, 2026

This work addresses the frequent neglect of anatomical constraints in existing radar-based human pose estimation methods, which often yields biomechanically implausible motions. To enhance physical plausibility and subject-specific adaptation, the authors propose an end-to-end differentiable framework that integrates a biomechanical skeleton model and differentiable forward kinematics into radar point cloud processing. The approach incorporates kinematic supervision, individualized geometric parameter fitting, temporal pose prediction, and contact classification loss. Evaluated on rehabilitation movements from 11 subjects, the system achieves an average joint position error of 6.46 cm, a mean joint angle error of 8.08°, a contact classification F1 score of 0.935, and a limb scaling error of 3.4%, significantly outperforming current state-of-the-art methods.

0 citationsRead paper

Variational Inference Using a Differentiable Multigrid Linear Solver

Aug 01, 2026

This work addresses the computational challenge of gradient evaluation in high-dimensional forward models arising in PDE-constrained inverse problems by introducing a Differentiable Multigrid Solver (DMGS). For the first time, DMGS enables compatibility between exact adjoint operations and automatic differentiation frameworks. The method explicitly derives the full multigrid hierarchy’s adjoint operators for steady-state diffusion-absorption problems and integrates them into JAX, efficiently supporting both Jacobian-vector and vector-Jacobian products. Applied to a 3D tissue diffuse radiative transfer inverse problem, DMGS achieves accurate reconstruction of effective radiation sources (χ² = 1.1) with reduced peak memory usage and controllable backward-pass overhead, while demonstrating strong generalization across 32 validation cases.

0 citationsRead paper

Dense Temporal Contrast Synthesis via Conditioned Latent Transport

Jul 31, 2026

This work addresses the limitations of conventional dynamic contrast-enhanced MRI (DCE-MRI), which relies on gadolinium-based contrast agents and suffers from contraindications, prolonged scan times, and environmental toxicity, while existing synthesis methods struggle to simultaneously preserve spatial fidelity and temporal continuity. The authors propose a conditional latent space transport framework that leverages anatomical priors to anchor latent trajectories and incorporates continuous-time embeddings to generate high-fidelity, patient-specific DCE-MRI sequences at arbitrary time points in a single forward pass. This approach represents the first non-iterative, temporally continuous method for DCE-MRI synthesis. In multi-center clinical validation, it significantly outperforms current techniques, improving tumor segmentation Dice scores by 22.4%, reducing boundary errors by over 39%, and achieving radiologist approval for clinical decision-making in 70% of cases.

0 citationsRead paper