Selected work

Representative Papers

Combinatorial Optimization Augmented Machine Learning

Jan 15, 2026

This work addresses the central challenge of integrating predictive models with combinatorial optimization to enable data-driven intelligent decision-making while preserving solution feasibility. It proposes a unified framework that embeds combinatorial optimization solvers directly into machine learning pipelines, systematically combining empirical risk minimization, imitation learning, and reinforcement learning. A key component of the framework is a feasibility-preserving mechanism designed to operate effectively in both static and dynamic settings. Beyond algorithmic integration, the study establishes a comprehensive problem taxonomy and algorithmic paradigm for this research direction, and provides a systematic review of theoretical foundations and applications in domains such as scheduling and routing, thereby offering a clear roadmap for future research.

1 citationsRead paper

BOP-Distrib: Revisiting 6D Pose Estimation Benchmarks for Better Evaluation under Visual Ambiguities

Aug 30, 2024

Current 6D pose estimation benchmarks oversimplify visual ambiguities—such as symmetry and occlusion—as global object symmetries, neglecting image-level, viewpoint-dependent visibility variations, thereby misrepresenting real-world pose uncertainty. To address this, we propose a novel image-level pose distribution evaluation paradigm: (1) the first automatic pose distribution annotation method grounded in single-image surface visibility; (2) BOP-Dist, the first pose distribution benchmark tailored to realistic images; and (3) a symmetry-aware sampling strategy coupled with a distribution-aware accuracy/recall evaluation framework. After re-annotating all BOP datasets with pose distributions, we observe substantial corrections to the performance ranking of state-of-the-art single-solution methods—revealing their rankings to be highly sensitive to annotation granularity. This work establishes a physically interpretable, reproducible, and quantitative evaluation standard for multi-solution pose estimation.

1 citationsRead paper

DecompRL: Solving Harder Problems by Learning Modular Code Generation

Jul 02, 2026

This work addresses the challenge that large language models struggle to solve extremely difficult programming problems due to the near-zero probability of sampling correct solutions through conventional methods or reinforcement learning (RL). To overcome this, we propose DecompRL, the first RL approach tailored for modular code generation. DecompRL decomposes complex problems into independently solvable subfunctions, learns to generate composable hierarchical code structures, and leverages verifiable reward signals for training. This framework enables exponential combinatorial exploration of candidate solutions while allowing efficient CPU-based evaluation of composed programs, drastically reducing GPU inference costs. Experiments on LiveCodeBench and CodeContests demonstrate that DecompRL substantially outperforms both standard and diversity-enhanced RL baselines, successfully solving problems intractable to prior methods while reducing GPU token consumption by approximately 50×.

0 citationsRead paper

UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation

Jun 22, 2026

This work addresses the challenge of limited generalizability of Vision Transformers (ViTs) in Earth observation due to disparities in sensor modalities, spatial scales, and resolutions. To overcome this, the authors propose UniverSat, a ViT architecture built upon a universal patch encoder that maps heterogeneous remote sensing data—spanning arbitrary spatial, spectral, and temporal resolutions, as well as optical and non-optical modalities—into a unified embedding space through shared-weight projections. Coupled with self-supervised pretraining, UniverSat enables unified representation learning across diverse multimodal inputs without relying on fixed patch projections, a constraint inherent in conventional ViTs. The model achieves state-of-the-art performance on multiple benchmarks, including GeoBench, PANGEABench, and SpectralEarth, consistently outperforming existing methods in both classification and segmentation tasks.

0 citationsRead paper

FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity

Jun 17, 2026

Reconstructing complete 3D objects from sparse, casually captured images remains challenging, as generative models often suffer from synthetic biases while optimization-based methods struggle to inpaint unobserved regions. This work proposes a training-free dual-space guidance mechanism that reformulates the reconstruction task as a guided inverse problem within a flow-matching framework. By steering ODE trajectories, the method efficiently leverages generative priors to complete unseen regions while strictly preserving consistency with observed data. It further refines geometry and appearance using 3D Gaussian Splatting. To our knowledge, this is the first approach within the flow-matching paradigm to simultaneously achieve high geometric completeness and strong observational fidelity, significantly outperforming both generative and optimization-based baselines on both synthetic and real-world datasets.

0 citationsRead paper

Recent publications

Latest Papers

DecompRL: Solving Harder Problems by Learning Modular Code Generation

Jul 02, 2026

This work addresses the challenge that large language models struggle to solve extremely difficult programming problems due to the near-zero probability of sampling correct solutions through conventional methods or reinforcement learning (RL). To overcome this, we propose DecompRL, the first RL approach tailored for modular code generation. DecompRL decomposes complex problems into independently solvable subfunctions, learns to generate composable hierarchical code structures, and leverages verifiable reward signals for training. This framework enables exponential combinatorial exploration of candidate solutions while allowing efficient CPU-based evaluation of composed programs, drastically reducing GPU inference costs. Experiments on LiveCodeBench and CodeContests demonstrate that DecompRL substantially outperforms both standard and diversity-enhanced RL baselines, successfully solving problems intractable to prior methods while reducing GPU token consumption by approximately 50×.

0 citationsRead paper

UniverSat: Resolution- and Modality-Agnostic Transformers for Earth Observation

Jun 22, 2026

This work addresses the challenge of limited generalizability of Vision Transformers (ViTs) in Earth observation due to disparities in sensor modalities, spatial scales, and resolutions. To overcome this, the authors propose UniverSat, a ViT architecture built upon a universal patch encoder that maps heterogeneous remote sensing data—spanning arbitrary spatial, spectral, and temporal resolutions, as well as optical and non-optical modalities—into a unified embedding space through shared-weight projections. Coupled with self-supervised pretraining, UniverSat enables unified representation learning across diverse multimodal inputs without relying on fixed patch projections, a constraint inherent in conventional ViTs. The model achieves state-of-the-art performance on multiple benchmarks, including GeoBench, PANGEABench, and SpectralEarth, consistently outperforming existing methods in both classification and segmentation tasks.

0 citationsRead paper

FlowObject: Flow Steering for Bridging Generative Priors and Reconstruction Fidelity

Jun 17, 2026

Reconstructing complete 3D objects from sparse, casually captured images remains challenging, as generative models often suffer from synthetic biases while optimization-based methods struggle to inpaint unobserved regions. This work proposes a training-free dual-space guidance mechanism that reformulates the reconstruction task as a guided inverse problem within a flow-matching framework. By steering ODE trajectories, the method efficiently leverages generative priors to complete unseen regions while strictly preserving consistency with observed data. It further refines geometry and appearance using 3D Gaussian Splatting. To our knowledge, this is the first approach within the flow-matching paradigm to simultaneously achieve high geometric completeness and strong observational fidelity, significantly outperforming both generative and optimization-based baselines on both synthetic and real-world datasets.

0 citationsRead paper

Pano3D: Unified 3D Reconstruction and Panoptic Segmentation

Jun 12, 2026

This work addresses the challenge of achieving high-fidelity 3D geometric reconstruction and robust semantic understanding without requiring camera parameters. To this end, it proposes the first unified framework that jointly trains 3D feedforward reconstruction and 3D panoptic segmentation. The approach incorporates geometric priors during feature initialization and enables end-to-end co-optimization through combined geometric and semantic losses. A set-based mask decoder is employed, compatible with both online and full pairwise attention architectures. The method achieves state-of-the-art performance in 3D panoptic segmentation on ScanNet, ScanNet200, and ScanNet++. Ablation studies confirm that joint training mutually enhances both reconstruction accuracy and semantic understanding.

0 citationsRead paper

Fast Cubical Persistent Homology on 2D and 3D Images via Union-Find, Pruning, and Lookup Tables

Jun 03, 2026

This work addresses the computational inefficiency of V-filtration–based cubical persistent homology in 2D/3D images by introducing a highly efficient algorithm. By incorporating union-find data structures and duality principles—both novel to cubical persistent homology—and accelerating the merge process via an edge-pruning strategy, the method substantially reduces both time and memory overhead. Furthermore, it leverages lookup tables to precompute local cubical configurations, enhancing computational speed without compromising correctness. The proposed approach achieves state-of-the-art performance, offering the fastest computation and lowest memory footprint currently available for V-filtration cubical persistent homology, thereby enabling scalable topological analysis of large-scale image data.

0 citationsRead paper