Bocheng Li  [GitHub] [Google Scholar]

Ph.D. Student, School of Computer Science

State Key Laboratory of Cognitive Intelligence

University of Science and Technology of China (USTC)

Email: bcli (at) mail.ustc.edu.cn

About Me

I am a 3rd year Ph.D. student at University of Science and Technology of China (USTC), advised by Prof. Linli Xu. I received my B.Eng in computer science from the same institution in 2023. I play with generative models, and design mechanisms and representations that make them faster, more stable, and more controllable across text, vision, and action modalities.

Aside from my work, I spent a wonderful year with my friends at USTCLUG.

Publications
(* = equal contribution)

On more efficient/unified generative mechanisms:

• Unifying Continuous and Discrete Text Diffusion with Non-simultaneous Diffusion Processes [arXiv] [ACL Anthology] [Code]
  Bocheng Li*, Zhujin Gao* and Linli Xu
  In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), Oral Presentation
  Hybrid text diffusion model that unifies continuous and discrete diffusion with fine-grained noising control, context-aware reverse process and optimized noising schedule. Outperforms several discrete/continuous diffusion & NAR baselines on conditional text generation.

• Few-shot Temporal Pruning Accelerates Diffusion Models for Text Generation  [ACL Anthology] [Code]
  Bocheng Li, Zhujin Gao, Yongxin Zhu, Kun Yin, Haoyu Cao, Deqiang Jiang and Linli Xu
  In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (COLING 2024)
  Super straightforward and simple Bayesian optimization to accelerate Multinomial Diffusion, Absorbing Diffusion, and DiffuSeq by up to 400x in less than 1 minute.

• FastHybrid: Accelerating Hybrid Autoregressive Image Generation with Lookahead and Guided Decoding
  Zhengguo Jiang, Fang Zhang, Yongxiang Hua, Bocheng Li, Wentao Zhang and Linli Xu
  In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2026 (CVPR 2026)
  Hybrid autoregressive–diffusion image generation with lookahead semantic refinement and guided diffusion sampling, achieving high-fidelity synthesis with substantially fewer denoising steps.

On more robust/learnable representation learning:

• Addressing Representation Collapse in Vector Quantized Models with One Linear Layer [arXiv] [Code] [Post on kexue.fm] [@lucidrain's VQ Repo]
  Yongxin Zhu, Bocheng Li, Yifei Xin and Linli Xu
  In Proceedings of the International Conference on Computer Vision 2025 (ICCV 2025)
  A method which reparameterizes the code vectors through a linear transformation layer based on a learnable latent basis to resolve the representation collapse problem in VQ models. Adopted by Kelix @ Kuaishou and many other research projects.

• Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective [arXiv] [Code]
  Yongxin Zhu, Bocheng Li, Hang Zhang, Xin Li, Linli Xu and Lidong Bing
  In Proceedings of the Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024)
  An auto-regressive generative model performing next-token prediction in an abstract latent space derived from self-supervised learning (SSL) models, SOTA performance for AR image generation & understanding on ImageNet. Among the earliest works of vision tokenizers for unified generative models.

Other super interesting things we've explored:

• Learning Native Continuation for Action Chunking Flow Policies [arXiv]
  Yufeng Liu, Hang Yu, Juntu Zhao, Bocheng Li, Di Zhang, Mingzhu Li, Wenxuan Wu, Yingdong Hu, Junyuan Xie, Junliang Guo, Dequan Wang and Yang Gao
  In Proceedings of Robotics: Science and Systems 2026 (RSS 2026)
  A training-time continuation method for action-chunked VLA policies that aligns training and inference dynamics to produce intrinsically smooth, temporally consistent actions, outperforming RTC/Train-time RTC(which is the one deployed on Pi0.6 @ Physical Intelligence) in both smoothness and efficiency.

• Input Domain Aware MoE: Decoupling Routing Decisions from Task Optimization in Mixture of Experts [arXiv]
  Yongxiang Hua, Haoyu Cao, Zhou Tao, Bocheng Li, Zihao Wu, Chaohu Liu and Linli Xu
  In Proceedings of the 33rd ACM International Conference on Multimedia (ACM MM 2025)
  An MoE routing framework using a probabilistic mixture model to partition the input space, achieving superior expert specialization, load balancing, and task performance over standard sMoE approaches.

Experience

• Research Intern, Foundation Models @ Spirit AI Robotics, November 2025 - February 2026
  Train foundation VLA models at scale. Spirit-v1.5 beats Pi 0.5 on Table30 Benchmark.

• Research Intern, Tencent YouTu Research, March 2023 - October 2023
  Focused on Generative Models and MLLMs for VQA.

• President of USTC Linux User Group, May 2022 - May 2023
  USTCLUG is the 5-star student club at USTC in 2023, the highest honor bestowed upon student organizations.

Awards

• National Scholarship, October 2025

• USTC-Suzhou Industrial Park Scholarship, November 2024

• Excellence Award (Advanced Game Theory Algorithm, Top 10%), Tencent AIArena National Open Competition for Artificial Intelligence, December 2023

• USTC Outstanding Thesis Award (Bachelor Thesis; 83 out of 1860), June 2023

Teaching Assistant

• Basics of Artificial Intelligence, 2023, 2024 Spring (USTC)

• Computer Systems: A Programmer's Perspective[CS:APP], 2022 Spring (USTC)

Services

• Reviewer: ACL ARR, NeurIPS 2024/2025, ICLR 2025/2026, AAAI 2026, CVPR 2026