Selected work

Representative Papers

Emergent Multi-View Geometry Through Self-Distillation

Sep 30, 2026

This study addresses the challenge of disentangling geometry from appearance in multi-view visual representation learning without explicit 3D supervision. We propose Poincaré3, a self-supervised framework that departs from conventional RGB reconstruction paradigms by leveraging masked patch modeling and image-level self-distillation to achieve geometry–appearance disentanglement. Specifically, Poincaré3 introduces a Poincaré adapter that enables a teacher model to observe additional views, facilitating the training of multi-view geometric representations from scratch. Experimental results demonstrate that Poincaré3 significantly outperforms baselines such as DINOv3 across correspondence estimation, camera pose prediction, and 3D reconstruction tasks, while encoding camera motion more precisely. Ultimately, this work establishes a new paradigm for unsupervised multi-view understanding.

0 citationsRead paper

Shrinkage through multiple identifiability

Apr 20, 2026

This study addresses the challenge of robustly combining multiple estimators to infer a common causal parameter when the underlying functionals are only partially identified or arise from non-nested settings. The authors propose an empirical Bayes framework that aggregates asymptotically linear estimators via posterior means, ensuring consistency under two distinct non-nested scenarios: exact identification and zero-mean bias. Innovatively integrating the bias structures of multiple functionals with Bayesian shrinkage, the approach distinguishes between different identification mechanisms and constructs both frequentist confidence intervals and Bayesian predictive intervals accordingly. Theoretical results establish the estimator’s consistency and asymptotic efficiency, while practical implementation leverages sandwich variance estimation, subsampling, and mixture distribution modeling to effectively synthesize evidence from observational data and randomized trials.

0 citationsRead paper

On the Interplay of Cube Learning and Dependency Schemes in QCDCL Proof Systems

Oct 07, 2025

This work addresses the insufficient synergy between cube learning and dependency schemes in QCDCL proof systems. We formally define an enhanced QCDCL system that natively integrates cube learning with full dependency schemes—specifically Dstd and Drrs—for the first time. We establish sufficient conditions ensuring soundness and completeness, and prove theoretically that, under relaxed decision orders, both Dstd and Drrs provably reduce refutation length for false QBFs. Empirical evaluation demonstrates significant improvements in variable propagation and decision efficiency, leading to shorter proofs and accelerated QBF solving. The core innovation lies in establishing a tight coupling mechanism between cube learning and dependency schemes, yielding a new paradigm for quantified Boolean satisfiability that unifies theoretical rigor with practical efficiency.

0 citationsRead paper

Application of discrete Ricci curvature in pruning randomly wired neural networks: A case study with chest x-ray classification of COVID-19

Aug 30, 2025

This study addresses model compression for Random Wired Neural Networks (RWNNs) in COVID-19 chest X-ray classification. We propose a discrete Ricci curvature-based edge pruning method, introducing Forman-Ricci curvature—computationally more efficient than Ollivier-Ricci curvature—into neural network pruning for the first time, while preserving pruning efficacy. The method is further enhanced by integrating edge betweenness centrality and systematically evaluated across Erdős–Rényi, Watts–Strogatz, and Barabási–Albert topologies to demonstrate robustness. Experiments show that our approach significantly reduces parameter count and theoretical computational cost—yielding higher speedup—while maintaining classification accuracy, sensitivity, and specificity. Moreover, it achieves a superior trade-off between preserving modular structure and sustaining global information propagation efficiency.

0 citationsRead paper

Predictive posteriors under hidden confounding

Jul 07, 2025

To address the failure of cross-domain predictive generalization under distributional shift caused by latent confounding, this paper proposes the first Bayesian Generative Invariance (BGI) framework. Unlike frequentist generative invariance (GI) methods—which cannot identify causal structures and lack uncertainty quantification—BGI endows the number of observable environments with an asymptotic prior interpretation, enabling identifiable causal discovery and well-calibrated out-of-distribution prediction without hyperparameter tuning or prior knowledge of distribution shifts. The method integrates multi-environment joint Bayesian modeling with the principle of generative invariance, supporting robust high-dimensional inference. Experiments demonstrate that BGI achieves empirical coverage rates close to nominal levels in low- to medium-dimensional settings, significantly outperforming conservative regularization and existing frequentist approaches.

0 citationsRead paper

Recent publications

Latest Papers

Emergent Multi-View Geometry Through Self-Distillation

Sep 30, 2026

This study addresses the challenge of disentangling geometry from appearance in multi-view visual representation learning without explicit 3D supervision. We propose Poincaré3, a self-supervised framework that departs from conventional RGB reconstruction paradigms by leveraging masked patch modeling and image-level self-distillation to achieve geometry–appearance disentanglement. Specifically, Poincaré3 introduces a Poincaré adapter that enables a teacher model to observe additional views, facilitating the training of multi-view geometric representations from scratch. Experimental results demonstrate that Poincaré3 significantly outperforms baselines such as DINOv3 across correspondence estimation, camera pose prediction, and 3D reconstruction tasks, while encoding camera motion more precisely. Ultimately, this work establishes a new paradigm for unsupervised multi-view understanding.

0 citationsRead paper

Shrinkage through multiple identifiability

Apr 20, 2026

This study addresses the challenge of robustly combining multiple estimators to infer a common causal parameter when the underlying functionals are only partially identified or arise from non-nested settings. The authors propose an empirical Bayes framework that aggregates asymptotically linear estimators via posterior means, ensuring consistency under two distinct non-nested scenarios: exact identification and zero-mean bias. Innovatively integrating the bias structures of multiple functionals with Bayesian shrinkage, the approach distinguishes between different identification mechanisms and constructs both frequentist confidence intervals and Bayesian predictive intervals accordingly. Theoretical results establish the estimator’s consistency and asymptotic efficiency, while practical implementation leverages sandwich variance estimation, subsampling, and mixture distribution modeling to effectively synthesize evidence from observational data and randomized trials.

0 citationsRead paper

On the Interplay of Cube Learning and Dependency Schemes in QCDCL Proof Systems

Oct 07, 2025

This work addresses the insufficient synergy between cube learning and dependency schemes in QCDCL proof systems. We formally define an enhanced QCDCL system that natively integrates cube learning with full dependency schemes—specifically Dstd and Drrs—for the first time. We establish sufficient conditions ensuring soundness and completeness, and prove theoretically that, under relaxed decision orders, both Dstd and Drrs provably reduce refutation length for false QBFs. Empirical evaluation demonstrates significant improvements in variable propagation and decision efficiency, leading to shorter proofs and accelerated QBF solving. The core innovation lies in establishing a tight coupling mechanism between cube learning and dependency schemes, yielding a new paradigm for quantified Boolean satisfiability that unifies theoretical rigor with practical efficiency.

0 citationsRead paper

Application of discrete Ricci curvature in pruning randomly wired neural networks: A case study with chest x-ray classification of COVID-19

Aug 30, 2025

This study addresses model compression for Random Wired Neural Networks (RWNNs) in COVID-19 chest X-ray classification. We propose a discrete Ricci curvature-based edge pruning method, introducing Forman-Ricci curvature—computationally more efficient than Ollivier-Ricci curvature—into neural network pruning for the first time, while preserving pruning efficacy. The method is further enhanced by integrating edge betweenness centrality and systematically evaluated across Erdős–Rényi, Watts–Strogatz, and Barabási–Albert topologies to demonstrate robustness. Experiments show that our approach significantly reduces parameter count and theoretical computational cost—yielding higher speedup—while maintaining classification accuracy, sensitivity, and specificity. Moreover, it achieves a superior trade-off between preserving modular structure and sustaining global information propagation efficiency.

0 citationsRead paper

Predictive posteriors under hidden confounding

Jul 07, 2025

To address the failure of cross-domain predictive generalization under distributional shift caused by latent confounding, this paper proposes the first Bayesian Generative Invariance (BGI) framework. Unlike frequentist generative invariance (GI) methods—which cannot identify causal structures and lack uncertainty quantification—BGI endows the number of observable environments with an asymptotic prior interpretation, enabling identifiable causal discovery and well-calibrated out-of-distribution prediction without hyperparameter tuning or prior knowledge of distribution shifts. The method integrates multi-environment joint Bayesian modeling with the principle of generative invariance, supporting robust high-dimensional inference. Experiments demonstrate that BGI achieves empirical coverage rates close to nominal levels in low- to medium-dimensional settings, significantly outperforming conservative regularization and existing frequentist approaches.

0 citationsRead paper