Haoyu Lei雷浩瑀

11 papers · * Equal contribution

1. Jingwei Zhang, Haoyu Lei, Zijin Feng, Jiacheng Sun, Farzan Farnia

   Exploring More to Solve More: Boosting Diversity in Text Diffusion Models via Entropy-Based Guidance

   ICML 2026
   [PAPER]
   Summary & figure

   We propose a training-free Semantic-Aware Kernel Entropy (SAKE) guidance method for diffusion LLMs to enhance the reasoning ability.

   
2. Haoyu Lei*, Mohammad Jalali*, Chin Wa Lau, Farzan Farnia

   On the Fragility of AI-Based Channel Decoders under Small Channel Perturbations

   ISIT 2026
   [PAPER]
   Summary & figure

   We theoretically and numerically study the lens of robustness to distributional shifts at the channel output.

   
3. Haoyu Lei, Chin Wa Lau, Kaiwen Zhou, Nian Guo, Farzan Farnia

   Syndrome-Flow Consistency Model Achieves One-step Denoising Error Correction Codes

   AAAI 2026 ML4Wireless (Best Paper Award)
   [PAPER]
   Summary & figure

   We propose a highly-efficient 1-NFE consistency flow model for decoding tasks in Error Correction Codes with superior performance.

   
4. Haoyu Lei, Kaiwen Zhou, Yinchuan Li, Zhitang Chen, Farzan Farnia

   Boosting Generalization in Diffusion-Based Neural Combinatorial Solver via Inference Time Adaptation

   AAAI 2026
   [PAPER]
   Summary & figure

   We propose a general energy-guided sampling framework during inference time that enhances both the cross-scale and cross-problem generalization capabilities of diffusion-based NCO solvers without requiring additional training.

   
5. Mohammad Jalali*, Haoyu Lei*, Amin Gohari, Farzan Farnia (*Equal Contribution)

   SPARKE: Scalable Prompt-Aware Diversity and Novelty Guidance in Diffusion Models via RKE Scores

   NeurIPS 2025
   [PAPER] [CODE]
   Summary & figure

   We propose the Scalable Prompt-Aware Renyi Kernel Entropy Diversity Guidance (SPARKE) method to improve prompt-aware diversity in text-to-image diffusion models, allowing large-scale generation settings with linear inference computational overhead.

   
6. Mateo Clemente, Leo Brunswic, Rui Heng Yang, Xuan Zhao, Yasser Khalil, Haoyu Lei, Amir Rasouli, Yinchuan Li

   Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising

   NeurIPS 2025
   [PAPER]
   Summary & figure

   We propose an efficient two-step sampling strategy for diffusion policy models for robotic manipulation satisfying physical constraints.

   
7. Garry Yang, Zizhe Chen, Man Hon Wong, Haoyu Lei, Yongqiang Chen, Zhenguo Li, Kaiwen Zhou, James Cheng

   MESH - Understanding Videos Like Human: Measuring Hallucinations in Large Video Models

   ACMMM 2025
   [PAPER] [CODE]
   Summary & figure

   We introduce MESH, a benchmark designed to evaluate hallucinations in LVMs systematically. MESH uses a Question-Answering framework with binary and multi-choice formats, incorporating target and trap instances.

   
8. Haoyu Lei, Shizhan Gong, Qi Dou, Farzan Farnia

   pFedFair: Towards Optimal Group Fairness-Accuracy Trade-off in Heterogeneous Federated Learning

   Preprint, 2025
   [PAPER]
   Summary & figure

   We propose the pFedFair to address personalized federated learning under heterogeneous feature distribution, while applying image embeddings to improve the performance of group-fairness learning methods on computer vision datasets.

   
9. Shizhan Gong, Haoyu Lei, Qi Dou, Farzan Farnia

   Boosting the visual interpretability of CLIP via adversarial fine-tuning

   ICLR 2025
   [PAPER] [CODE]
   Summary & figure

   We propose an unsupervised adversarial fine-tuning (AFT) with norm-regularization to enhance the visual interpretability of CLIP.

   
10. Haoyu Lei, Amin Gohari, Farzan Farnia

    On the Inductive Biases of Demographic Parity-based Fair Learning Algorithms

    UAI 2024
    [PAPER] [CODE]
    Summary & figure

    We theoretically show that Inductive Biases widely exist in DP-based fair supervised learning algorithms, and further propose a distributionally robust optimization framework to mitigate the biases.

    
11. Xiaolu Wang, Chaorui Yao, Haoyu Lei, Anthony Man-Cho So

    An Efficient Alternating Direction Method For Graph Learning From Smooth Signals

    ICASSP 2021
    [PAPER]
    Summary & figure

    We cast the graph learning formulation as a nonsmooth, strictly convex optimization problem and develop an efficient alternating direction method of multipliers (ADMM) to solve it.

    

Education

Research & Professional Experience

Languages: English (Advanced), Cantonese (Intermediate), Putonghua (Native)

Programming Skills: Python (Main), Pytorch (Main), Matlab, R, C, C++, C#, Julia, LaTeX

hylei22@cse.cuhk.edu.hk

Hong Kong SAR, China

Feel free to email me about internships or research collaborations.