Selected work

Representative Papers

Manifold GCN: Diffusion-based Convolutional Neural Network for Manifold-valued Graphs

Jan 25, 2024arXiv.org

Modeling graph-structured data residing on Riemannian manifolds poses challenges in preserving intrinsic geometric structure while ensuring equivariance under both node permutations and manifold isometries. Method: We propose the first equivariant graph neural network layer that jointly incorporates manifold diffusion modeling and nonlinear equivariant mapping in tangent spaces. The layer defines graph convolution via the manifold diffusion equation and constructs an equivariant multilayer perceptron in the tangent space at each node, enabling native support for arbitrary graph topologies and sizes. Contribution/Results: Our layer rigorously satisfies equivariance under node permutations and Riemannian isometries, and uniformly accommodates diverse Riemannian manifolds—including spheres, hyperbolic spaces, and triangulated surfaces—by embedding strong geometric inductive biases. Experiments on synthetic manifold graph datasets and a real-world Alzheimer’s disease classification task using right hippocampal triangular meshes demonstrate performance competitive with or superior to state-of-the-art specialized methods, alongside significantly improved generalization.

3 citationsRead paper

Embedding Model-form Uncertainty in Probabilistic Calibration of Digital Twins for Bridges

Sep 09, 2026

本文提出一种通过随机参数嵌入显式表示模型形式不确定性(MFU)的校准框架，提高桥梁数字孪生预测的可靠性和可信度。

0 citationsRead paper

Race, Exchange, Improve: Finding high-quality MIP solutions quickly

Sep 05, 2026

Mixed-integer programming (MIP) is a cornerstone in applied optimization, both in industry and academia. Recently, there has been increased attention to finding strong primal solutions quickly. This is reflected, for example, in the development of the NVIDIA cuOpt solver and, most recently, in the new MIPFEAS benchmark, which has a tight time limit of 600 seconds and evaluates solvers based on how quickly they find high-quality primal solutions. This article introduces a MIP portfolio parallelization scheme, focusing on efficiently exchanging information between its workers. We present two implementations of this scheme: one built directly into the open-source MIP solver SCIP, and an external one, which we call ReXi. ReXi is currently the fastest non-commercial solver in the MIPFEAS benchmark, followed by the SCIP-integrated implementation. Moreover, we present new versions of both implementations that considerably outperform their predecessors on the MIPFEAS benchmark.

0 citationsRead paper

Beyond Parameter Space: NTK-Guided Personalized Aggregation for Robust Federated Learning

Aug 12, 2026

This work addresses the performance degradation in federated learning under data heterogeneity, where conventional aggregation based on parameter-space similarity often fails. The authors propose LIGHTYEAR, a novel framework that introduces Neural Tangent Kernels (NTK) into federated learning to measure consistency between model updates and each client’s target-domain predictive behavior in function space. By leveraging a peer-to-peer topology and enabling clients to select beneficial updates using their private validation data, LIGHTYEAR replaces parameter-space similarity with function-space alignment, thereby enabling personalized, target-aware aggregation. Experimental results across five datasets demonstrate that LIGHTYEAR significantly outperforms nine state-of-the-art baselines, achieving substantial gains in both model performance and robustness in heterogeneous settings.

0 citationsRead paper

When Does LLM Orchestration Pay Off? A Controlled Evaluation of Accuracy, Cost, and Task Difficulty

Aug 01, 2026

This study investigates whether the additional inference cost of large language model (LLM) orchestration is justified and whether its effectiveness varies with task difficulty. Under a unified evaluation framework that controls for optimization budget and stratifies tasks by difficulty, the authors systematically assess orchestration strategies—including Self-Refine, Best-of-N, and Debate—across programming, mathematics, and chess domains in terms of accuracy–cost trade-offs. Results show that orchestration improves accuracy by up to 4.6 percentage points on average but incurs a 2–4× increase in token consumption. Crucially, the gains depend strongly on the base model’s capability rather than task difficulty, challenging the common assumption that more elaborate reasoning structures inherently yield better performance.

0 citationsRead paper

Recent publications

Latest Papers

Embedding Model-form Uncertainty in Probabilistic Calibration of Digital Twins for Bridges

Sep 09, 2026

本文提出一种通过随机参数嵌入显式表示模型形式不确定性(MFU)的校准框架，提高桥梁数字孪生预测的可靠性和可信度。

0 citationsRead paper

Race, Exchange, Improve: Finding high-quality MIP solutions quickly

Sep 05, 2026

Mixed-integer programming (MIP) is a cornerstone in applied optimization, both in industry and academia. Recently, there has been increased attention to finding strong primal solutions quickly. This is reflected, for example, in the development of the NVIDIA cuOpt solver and, most recently, in the new MIPFEAS benchmark, which has a tight time limit of 600 seconds and evaluates solvers based on how quickly they find high-quality primal solutions. This article introduces a MIP portfolio parallelization scheme, focusing on efficiently exchanging information between its workers. We present two implementations of this scheme: one built directly into the open-source MIP solver SCIP, and an external one, which we call ReXi. ReXi is currently the fastest non-commercial solver in the MIPFEAS benchmark, followed by the SCIP-integrated implementation. Moreover, we present new versions of both implementations that considerably outperform their predecessors on the MIPFEAS benchmark.

0 citationsRead paper

Beyond Parameter Space: NTK-Guided Personalized Aggregation for Robust Federated Learning

Aug 12, 2026

This work addresses the performance degradation in federated learning under data heterogeneity, where conventional aggregation based on parameter-space similarity often fails. The authors propose LIGHTYEAR, a novel framework that introduces Neural Tangent Kernels (NTK) into federated learning to measure consistency between model updates and each client’s target-domain predictive behavior in function space. By leveraging a peer-to-peer topology and enabling clients to select beneficial updates using their private validation data, LIGHTYEAR replaces parameter-space similarity with function-space alignment, thereby enabling personalized, target-aware aggregation. Experimental results across five datasets demonstrate that LIGHTYEAR significantly outperforms nine state-of-the-art baselines, achieving substantial gains in both model performance and robustness in heterogeneous settings.

0 citationsRead paper

When Does LLM Orchestration Pay Off? A Controlled Evaluation of Accuracy, Cost, and Task Difficulty

Aug 01, 2026

This study investigates whether the additional inference cost of large language model (LLM) orchestration is justified and whether its effectiveness varies with task difficulty. Under a unified evaluation framework that controls for optimization budget and stratifies tasks by difficulty, the authors systematically assess orchestration strategies—including Self-Refine, Best-of-N, and Debate—across programming, mathematics, and chess domains in terms of accuracy–cost trade-offs. Results show that orchestration improves accuracy by up to 4.6 percentage points on average but incurs a 2–4× increase in token consumption. Crucially, the gains depend strongly on the base model’s capability rather than task difficulty, challenging the common assumption that more elaborate reasoning structures inherently yield better performance.

0 citationsRead paper

Closed-Loop Generative Selection: Convergence, Memory, and Noisy Oracles

Jul 24, 2026

Closed-loop generative selection algorithms are widely used in computational drug discovery, yet they lack convergence guarantees due to the violation of the Markov property caused by model retraining at each iteration. This work restores the Markov structure by expanding the state space and provides, for the first time, an almost sure convergence proof along with a decomposition of runtime complexity for such algorithms. We further reveal a dual role of memory depth in search efficiency: excessive memory depth can paradoxically slow convergence. By integrating absorbing-state analysis, exit-time estimation, and robust multi-objective optimization, we propose a practical strategy that minimizes the number of evaluation calls. Theoretical predictions are corroborated through reproducible experiments, demonstrating consistent performance under both light- and heavy-tailed noise conditions.

0 citationsRead paper