Selected work

Representative Papers

A Hybrid Vision Transformer Approach for Mathematical Expression Recognition

Nov 30, 2022International Conference on Digital Image Computing: Techniques and Applications

This work proposes a hybrid Vision Transformer-based sequence-to-sequence approach to address the challenges in mathematical expression recognition arising from the two-dimensional layout and scale variations of symbols. The encoder integrates 2D positional encoding to effectively capture spatial dependencies among symbols and innovatively employs the ViT’s [CLS] token as the initial embedding for the decoder. To mitigate issues of under- or over-parsing during decoding, a coverage attention mechanism is introduced. Evaluated on the IM2LATEX-100K dataset, the proposed method achieves a BLEU score of 89.94, outperforming current state-of-the-art approaches and significantly improving the accuracy of formula recognition.

2 citationsRead paper

Reservoir of Importance: Learning Semi-Structured Sparsity with Differentiable Subset Sampling

Aug 24, 2026

本文提出了一种轻量级的半结构化剪枝框架RoI，通过可微子集采样学习稀疏掩码，减少参数和内存开销，实现高效的大语言模型加速。

0 citationsRead paper

Semi-Dense Matching Uncertainty Is Not Just Local Confidence

Aug 09, 2026

This work addresses the challenge of accurately quantifying uncertainty in existing semi-dense matching methods, which often overlook catastrophic failures during the coarse matching stage, leading to biased geometric estimates. To remedy this, the authors propose a lightweight post-processing framework that explicitly incorporates coarse matching failures into uncertainty modeling. Specifically, they introduce a two-component calibrated Laplacian mixture model with only nine learnable parameters to capture the long-tailed distribution arising from both local refinement noise and coarse matching outliers. Additionally, a CoRe (Coarse-and-Refined) geometric refitting module is designed to leverage Bayesian posterior probabilities for generating soft correspondence weights, enabling reweighted optimization. The method consistently improves downstream geometric accuracy across diverse pretrained matchers and robust estimators while incurring minimal computational overhead.

0 citationsRead paper

SpecRoll: Fast-Slow Verifier-Feedback Adaptation for Speculative Reinforcement Learning Rollouts

Aug 05, 2026

This work addresses the inefficiency of autoregressive rollout generation in reinforcement learning post-training and the inability of existing speculative decoding methods to adapt to continuously evolving policies. To overcome these limitations, we propose SpecRoll—a dual-timescale adaptive speculative rollout engine that significantly accelerates generation while strictly preserving the target policy’s sampling distribution. SpecRoll employs a lightweight future-token head for parallel proposal generation and integrates a backpropagation-free Reflex module for local hidden-state correction along trajectories. It further incorporates concurrency-aware sparse-tree verification, exact target validation, adaptive fast-slow path routing, and delayed validator feedback. Evaluated across models ranging from 1.5B to 14B parameters and three mathematical reasoning benchmarks, SpecRoll achieves 1.26–2.15× faster rollout generation and 1.21–2.04× end-to-end speedup, consistently outperforming FastGRPO.

0 citationsRead paper

Lightweight 3D Object Detection via Mamba-Based Knowledge Distillation

Aug 04, 2026

This work addresses the trade-off between accuracy and efficiency in LiDAR-based 3D object detection for autonomous driving and robotic navigation by proposing a knowledge distillation framework built upon the Mamba architecture. The approach introduces a multi-branch teacher backbone and a box-aware, voxel-level feature transfer mechanism that enables selective feature alignment in voxel space, effectively distilling rich semantic knowledge from the teacher model into a lightweight student network. By focusing feature imitation on task-relevant regions, the method substantially reduces computational overhead while achieving detection accuracy comparable to state-of-the-art approaches on both public benchmarks and real-world datasets, thereby offering a favorable balance between deployment efficiency and performance.

0 citationsRead paper

Recent publications

Latest Papers

Reservoir of Importance: Learning Semi-Structured Sparsity with Differentiable Subset Sampling

Aug 24, 2026

本文提出了一种轻量级的半结构化剪枝框架RoI，通过可微子集采样学习稀疏掩码，减少参数和内存开销，实现高效的大语言模型加速。

0 citationsRead paper

Semi-Dense Matching Uncertainty Is Not Just Local Confidence

Aug 09, 2026

This work addresses the challenge of accurately quantifying uncertainty in existing semi-dense matching methods, which often overlook catastrophic failures during the coarse matching stage, leading to biased geometric estimates. To remedy this, the authors propose a lightweight post-processing framework that explicitly incorporates coarse matching failures into uncertainty modeling. Specifically, they introduce a two-component calibrated Laplacian mixture model with only nine learnable parameters to capture the long-tailed distribution arising from both local refinement noise and coarse matching outliers. Additionally, a CoRe (Coarse-and-Refined) geometric refitting module is designed to leverage Bayesian posterior probabilities for generating soft correspondence weights, enabling reweighted optimization. The method consistently improves downstream geometric accuracy across diverse pretrained matchers and robust estimators while incurring minimal computational overhead.

0 citationsRead paper

SpecRoll: Fast-Slow Verifier-Feedback Adaptation for Speculative Reinforcement Learning Rollouts

Aug 05, 2026

This work addresses the inefficiency of autoregressive rollout generation in reinforcement learning post-training and the inability of existing speculative decoding methods to adapt to continuously evolving policies. To overcome these limitations, we propose SpecRoll—a dual-timescale adaptive speculative rollout engine that significantly accelerates generation while strictly preserving the target policy’s sampling distribution. SpecRoll employs a lightweight future-token head for parallel proposal generation and integrates a backpropagation-free Reflex module for local hidden-state correction along trajectories. It further incorporates concurrency-aware sparse-tree verification, exact target validation, adaptive fast-slow path routing, and delayed validator feedback. Evaluated across models ranging from 1.5B to 14B parameters and three mathematical reasoning benchmarks, SpecRoll achieves 1.26–2.15× faster rollout generation and 1.21–2.04× end-to-end speedup, consistently outperforming FastGRPO.

0 citationsRead paper

Lightweight 3D Object Detection via Mamba-Based Knowledge Distillation

Aug 04, 2026

This work addresses the trade-off between accuracy and efficiency in LiDAR-based 3D object detection for autonomous driving and robotic navigation by proposing a knowledge distillation framework built upon the Mamba architecture. The approach introduces a multi-branch teacher backbone and a box-aware, voxel-level feature transfer mechanism that enables selective feature alignment in voxel space, effectively distilling rich semantic knowledge from the teacher model into a lightweight student network. By focusing feature imitation on task-relevant regions, the method substantially reduces computational overhead while achieving detection accuracy comparable to state-of-the-art approaches on both public benchmarks and real-world datasets, thereby offering a favorable balance between deployment efficiency and performance.

0 citationsRead paper

Max-Min Secrecy Rate Optimization for Secure ISAC Networks: Global Optimization and Low-Complexity Algorithm

Jun 11, 2026

This work addresses the challenge of securing communication in integrated sensing and communication (ISAC) systems against potential eavesdropping by multiple untrusted sensing users. It introduces, for the first time, a max-min secrecy rate fairness criterion to jointly optimize security and sensing performance under constraints on transmit power and beam pattern matching error. The resulting highly non-convex optimization problem is tackled via two proposed solution strategies: a globally optimal algorithm based on semidefinite relaxation combined with branch-and-bound, which guarantees convergence to the optimum within a prescribed accuracy, and a low-complexity suboptimal algorithm leveraging successive convex approximation that achieves near-optimal secrecy performance with significantly reduced computational overhead. The latter approach balances theoretical rigor with practical deployability, offering an efficient trade-off between performance and complexity.

0 citationsRead paper