Selected work

Representative Papers

mHC-GNN: Manifold-Constrained Hyper-Connections for Graph Neural Networks

Jan 05, 2026arXiv.org

This work addresses the limitations of deep graph neural networks (GNNs), which suffer from over-smoothing and are constrained in expressive power by the 1-Weisfeiler-Lehman (1-WL) test. To overcome these issues, the authors propose a manifold-constrained hyperconnection mechanism that constructs multiple parallel representation streams and employs Sinkhorn–Knopp normalization to constrain the stream mixing matrix to the Birkhoff polytope. This approach effectively mitigates over-smoothing and surpasses the 1-WL expressiveness barrier. Notably, it is the first to integrate manifold-constrained hyperconnections into GNNs, achieving consistent performance gains across ten benchmark datasets and four mainstream GNN architectures. Remarkably, the model maintains over 74% accuracy even at a depth of 128 layers, representing an improvement of more than 50 percentage points over standard GNNs.

1 citationsRead paper

Query-Side Attacks on GNN-Based KGQA: Tracing Failures from Entity Linking to Answer Generation

Aug 26, 2026

研究通过引入两种针对知识图谱的对抗性扰动方法，分析了基于GNN的知识图谱问答系统在面临输入问题扰动时各阶段的失效原因，指出子图构建是主要脆弱环节。

0 citationsRead paper

Free energy landscape of Dense Associative Memory

Jul 21, 2026

This work investigates the free energy landscape and memory retrieval mechanisms of high-order dense associative memory models. Leveraging large deviation theory and statistical physics, it constructs a free energy functional tailored to polynomial interactions and Log-Sum-Exponential (LSE) activation, enabling a rigorous analysis of temperature-dependent behavior and ground state energy in the finite pattern regime. The study establishes, for the first time, the exact full-retrieval phase transition threshold for LSE-based models, elucidates the critical role of initial conditions in memory recovery within high-order networks, and develops a general analytical framework extensible to complex associative memory architectures. This framework not only reproduces classical results from the Hopfield model but also systematically extends the theoretical foundations of dense associative memory.

0 citationsRead paper

Attacking Graph Foundation Models Through Their Shared Representation

Jul 20, 2026

Graph foundation models rely on alignment layers to achieve cross-domain generalization, yet their security remains unexplored. This work is the first to expose the vulnerability of alignment layers as an independent attack surface, demonstrating that model predictions can be disrupted during inference through targeted perturbations in the representation space and feasible input manipulations—such as modifications to edges, node features, or textual attributes. We introduce a carrier gain metric grounded in the decoder’s local Lipschitz sensitivity and reveal that OpenGraph is particularly susceptible due to its spectral tokenizer design. Experiments show that representation-space attacks are broadly effective across six prominent models, with three suffering over 50% accuracy degradation under realistic attack conditions; notably, OpenGraph exhibits significant failure with only one-fifth of the typical attack budget.

0 citationsRead paper

Directed Distance Fields for Constant-Time Ray Queries on Gaussian Splatting

May 30, 2026

This work addresses the limitation of 3D Gaussian Splatting, which supports only primary ray rendering and struggles to efficiently handle secondary ray queries required for shadows, ambient occlusion, and global illumination. To overcome this, the authors propose distilling a pre-trained 3D Gaussian Splatting scene into a lightweight directed distance field (DDF), enabling constant-time distance and hit queries for arbitrary rays. The method employs a mesh-free, end-to-end training pipeline with exact distance supervision to recover fine geometric details. The resulting DDF achieves 26–72× faster query speeds than sphere tracing, with memory and computational costs independent of scene complexity. Evaluated on 142 objects and real-world scenes, the approach produces high-quality secondary ray effects, achieving signal-to-noise ratios of 30.3 dB for shadows and 21.3 dB for ambient occlusion.

0 citationsRead paper

Recent publications

Latest Papers

Query-Side Attacks on GNN-Based KGQA: Tracing Failures from Entity Linking to Answer Generation

Aug 26, 2026

研究通过引入两种针对知识图谱的对抗性扰动方法，分析了基于GNN的知识图谱问答系统在面临输入问题扰动时各阶段的失效原因，指出子图构建是主要脆弱环节。

0 citationsRead paper

Free energy landscape of Dense Associative Memory

Jul 21, 2026

This work investigates the free energy landscape and memory retrieval mechanisms of high-order dense associative memory models. Leveraging large deviation theory and statistical physics, it constructs a free energy functional tailored to polynomial interactions and Log-Sum-Exponential (LSE) activation, enabling a rigorous analysis of temperature-dependent behavior and ground state energy in the finite pattern regime. The study establishes, for the first time, the exact full-retrieval phase transition threshold for LSE-based models, elucidates the critical role of initial conditions in memory recovery within high-order networks, and develops a general analytical framework extensible to complex associative memory architectures. This framework not only reproduces classical results from the Hopfield model but also systematically extends the theoretical foundations of dense associative memory.

0 citationsRead paper

Attacking Graph Foundation Models Through Their Shared Representation

Jul 20, 2026

Graph foundation models rely on alignment layers to achieve cross-domain generalization, yet their security remains unexplored. This work is the first to expose the vulnerability of alignment layers as an independent attack surface, demonstrating that model predictions can be disrupted during inference through targeted perturbations in the representation space and feasible input manipulations—such as modifications to edges, node features, or textual attributes. We introduce a carrier gain metric grounded in the decoder’s local Lipschitz sensitivity and reveal that OpenGraph is particularly susceptible due to its spectral tokenizer design. Experiments show that representation-space attacks are broadly effective across six prominent models, with three suffering over 50% accuracy degradation under realistic attack conditions; notably, OpenGraph exhibits significant failure with only one-fifth of the typical attack budget.

0 citationsRead paper

Directed Distance Fields for Constant-Time Ray Queries on Gaussian Splatting

May 30, 2026

This work addresses the limitation of 3D Gaussian Splatting, which supports only primary ray rendering and struggles to efficiently handle secondary ray queries required for shadows, ambient occlusion, and global illumination. To overcome this, the authors propose distilling a pre-trained 3D Gaussian Splatting scene into a lightweight directed distance field (DDF), enabling constant-time distance and hit queries for arbitrary rays. The method employs a mesh-free, end-to-end training pipeline with exact distance supervision to recover fine geometric details. The resulting DDF achieves 26–72× faster query speeds than sphere tracing, with memory and computational costs independent of scene complexity. Evaluated on 142 objects and real-world scenes, the approach produces high-quality secondary ray effects, achieving signal-to-noise ratios of 30.3 dB for shadows and 21.3 dB for ambient occlusion.

0 citationsRead paper

A Deterministic Separation Lemma

May 27, 2026

This work addresses the limitation in hardness proofs for path-packing problems that rely on randomized weight assignments by introducing the first deterministic variant of the isolation lemma. Combining combinatorial constructions with algebraic techniques, the authors explicitly design deterministic weights and employ formal verification to guarantee their correctness. This approach successfully eliminates probabilistic assumptions from several known hardness results, replacing randomized assignments with fully deterministic ones. Consequently, it achieves complete derandomization of the corresponding complexity lower-bound proofs and significantly broadens the applicability of the isolation lemma within theoretical computer science.

0 citationsRead paper