Selected work

Representative Papers

Rethinking Graph Generalization through the Lens of Sharpness-Aware Minimization

Feb 09, 2026

This work addresses the vulnerability of graph neural networks (GNNs) to minimal shift flipping (MSF)—a phenomenon where minor out-of-distribution perturbations induce misclassification under distributional shifts. From the perspective of Sharpness-Aware Minimization (SAM), the study establishes, for the first time, a theoretical connection between the local robustness radius and generalization error in graph learning, and proposes an energy-based function as a computable proxy for this radius. Building upon this insight, the authors introduce E2A, an energy-driven generative augmentation framework that leverages the energy landscape to guide the generation of pseudo out-of-distribution samples, thereby enhancing model robustness. Extensive experiments demonstrate that E2A significantly outperforms existing methods across multiple benchmarks, effectively mitigating the MSF issue and consistently improving the out-of-distribution generalization capability of GNNs.

0 citationsRead paper

Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization

Oct 23, 2025

Graph Neural Networks (GNNs) face a critical bottleneck in out-of-distribution generalization: identifying causal subgraphs requires environment annotations, and existing Invariant Risk Minimization (IRM) approaches rely on costly environment partitioning or heuristic splitting. This paper proposes an IRM-free framework that, for the first time, establishes a quantitative relationship between representation norm and distributional shift, thereby defining a novel distribution-invariance criterion independent of environment labels. Building upon this, we design a norm-guided invariant distribution objective that jointly leverages theoretical analysis and self-supervision to discover causal subgraphs. Our method achieves significant improvements over state-of-the-art methods on two mainstream graph benchmarks, demonstrating superior cross-environment stability and structural generalization robustness. The core contribution lies in eliminating the need for environment annotations and unifying the characterization of distribution-invariant graph representations both theoretically and algorithmically.

0 citationsRead paper

Is Model Editing Built on Sand? Revealing Its Illusory Success and Fragile Foundation

Oct 01, 2025

This work exposes a reliability crisis in large language model (LLM) knowledge editing: prevailing methods frequently exploit spurious input-output shortcuts rather than performing genuine semantic knowledge updates, yielding “illusory success.” To address this, we propose the first systematic evaluation framework targeting editing authenticity. Our framework innovatively incorporates negation-based counterexamples and counterfactual reasoning tests, jointly assessing semantic consistency and robustness to rigorously verify whether knowledge has been substantively revised. Empirical results demonstrate that state-of-the-art editing methods suffer sharp performance degradation on simple negation queries—confirming their widespread reliance on superficial correlations rather than robust, semantically grounded knowledge integration. This study challenges foundational assumptions underlying current editing paradigms and establishes a critical evaluation benchmark and theoretical foundation for developing trustworthy, interpretable model editing techniques.

0 citationsRead paper

TABLET: Table Structure Recognition using Encoder-only Transformers

Jun 08, 2025

To address the instability in bounding-box prediction and the trade-off between accuracy and efficiency in large-scale, dense table structure recognition, this paper proposes TABLET—a pure-encoder Transformer architecture. It introduces a novel Split-Merge two-stage paradigm: the Split stage formulates row/column segmentation as a sequence labeling task, while the Merge stage casts cell merging as a grid-cell classification task—fully eliminating bounding-box regression. TABLET comprises dual encoders for splitting and a single encoder for merging, requiring no detection head, thereby achieving high accuracy, low latency, and strong robustness. On FinTabNet and PubTabNet, it significantly outperforms state-of-the-art methods in both accuracy and inference speed, with substantially reduced computational overhead. Its lightweight design and end-to-end differentiability enable industrial-scale deployment for real-world table understanding tasks.

0 citationsRead paper

Learning Optimal Multimodal Information Bottleneck Representations

May 26, 2025

Existing multimodal information bottleneck (MIB) methods suffer from two critical limitations: (1) employing a fixed, empirically chosen regularization weight lacking theoretical justification, and (2) ignoring the imbalance in task-relevant information across modalities, thereby hindering optimal MIB learning. To address these, we propose OMIB—a theoretically grounded framework that (i) derives the first analytically tractable feasibility bound for the regularization weight; (ii) introduces a modality-adaptive dynamic weighting mechanism to explicitly model and balance task-relevant information across modalities; and (iii) establishes a rigorous variational information-theoretic foundation ensuring provably convergent optimization to the optimal MIB solution. OMIB integrates variational information bottleneck optimization, modality-specific regularization, and principled information-theoretic modeling. We validate its theoretical properties on synthetic data and demonstrate significant improvements over state-of-the-art methods across multiple downstream multimodal tasks.

0 citationsRead paper

Recent publications

Latest Papers

Rethinking Graph Generalization through the Lens of Sharpness-Aware Minimization

Feb 09, 2026

This work addresses the vulnerability of graph neural networks (GNNs) to minimal shift flipping (MSF)—a phenomenon where minor out-of-distribution perturbations induce misclassification under distributional shifts. From the perspective of Sharpness-Aware Minimization (SAM), the study establishes, for the first time, a theoretical connection between the local robustness radius and generalization error in graph learning, and proposes an energy-based function as a computable proxy for this radius. Building upon this insight, the authors introduce E2A, an energy-driven generative augmentation framework that leverages the energy landscape to guide the generation of pseudo out-of-distribution samples, thereby enhancing model robustness. Extensive experiments demonstrate that E2A significantly outperforms existing methods across multiple benchmarks, effectively mitigating the MSF issue and consistently improving the out-of-distribution generalization capability of GNNs.

0 citationsRead paper

Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization

Oct 23, 2025

Graph Neural Networks (GNNs) face a critical bottleneck in out-of-distribution generalization: identifying causal subgraphs requires environment annotations, and existing Invariant Risk Minimization (IRM) approaches rely on costly environment partitioning or heuristic splitting. This paper proposes an IRM-free framework that, for the first time, establishes a quantitative relationship between representation norm and distributional shift, thereby defining a novel distribution-invariance criterion independent of environment labels. Building upon this, we design a norm-guided invariant distribution objective that jointly leverages theoretical analysis and self-supervision to discover causal subgraphs. Our method achieves significant improvements over state-of-the-art methods on two mainstream graph benchmarks, demonstrating superior cross-environment stability and structural generalization robustness. The core contribution lies in eliminating the need for environment annotations and unifying the characterization of distribution-invariant graph representations both theoretically and algorithmically.

0 citationsRead paper

Is Model Editing Built on Sand? Revealing Its Illusory Success and Fragile Foundation

Oct 01, 2025

This work exposes a reliability crisis in large language model (LLM) knowledge editing: prevailing methods frequently exploit spurious input-output shortcuts rather than performing genuine semantic knowledge updates, yielding “illusory success.” To address this, we propose the first systematic evaluation framework targeting editing authenticity. Our framework innovatively incorporates negation-based counterexamples and counterfactual reasoning tests, jointly assessing semantic consistency and robustness to rigorously verify whether knowledge has been substantively revised. Empirical results demonstrate that state-of-the-art editing methods suffer sharp performance degradation on simple negation queries—confirming their widespread reliance on superficial correlations rather than robust, semantically grounded knowledge integration. This study challenges foundational assumptions underlying current editing paradigms and establishes a critical evaluation benchmark and theoretical foundation for developing trustworthy, interpretable model editing techniques.

0 citationsRead paper

TABLET: Table Structure Recognition using Encoder-only Transformers

Jun 08, 2025

To address the instability in bounding-box prediction and the trade-off between accuracy and efficiency in large-scale, dense table structure recognition, this paper proposes TABLET—a pure-encoder Transformer architecture. It introduces a novel Split-Merge two-stage paradigm: the Split stage formulates row/column segmentation as a sequence labeling task, while the Merge stage casts cell merging as a grid-cell classification task—fully eliminating bounding-box regression. TABLET comprises dual encoders for splitting and a single encoder for merging, requiring no detection head, thereby achieving high accuracy, low latency, and strong robustness. On FinTabNet and PubTabNet, it significantly outperforms state-of-the-art methods in both accuracy and inference speed, with substantially reduced computational overhead. Its lightweight design and end-to-end differentiability enable industrial-scale deployment for real-world table understanding tasks.

0 citationsRead paper

Learning Optimal Multimodal Information Bottleneck Representations

May 26, 2025

Existing multimodal information bottleneck (MIB) methods suffer from two critical limitations: (1) employing a fixed, empirically chosen regularization weight lacking theoretical justification, and (2) ignoring the imbalance in task-relevant information across modalities, thereby hindering optimal MIB learning. To address these, we propose OMIB—a theoretically grounded framework that (i) derives the first analytically tractable feasibility bound for the regularization weight; (ii) introduces a modality-adaptive dynamic weighting mechanism to explicitly model and balance task-relevant information across modalities; and (iii) establishes a rigorous variational information-theoretic foundation ensuring provably convergent optimization to the optimal MIB solution. OMIB integrates variational information bottleneck optimization, modality-specific regularization, and principled information-theoretic modeling. We validate its theoretical properties on synthetic data and demonstrate significant improvements over state-of-the-art methods across multiple downstream multimodal tasks.

0 citationsRead paper