Boxun Xu
Boxun Xu

Boxun Xu

My research aims to advance efficient and scalable AI through algorithmic innovation in efficient modeling and the design of high-performance machine learning systems. I currently focus on training and inference for multimodal content generation, addressing computational and memory bottlenecks to enable high-quality, long-horizon generation in real time.

I earned my Ph.D. at UC Santa Barbara, where I was fortunate to work with Prof. Peng Li (IEEE Fellow). Before UCSB, I completed my M.S. at the University of Michigan, Ann Arbor, where I had the pleasure of working with Prof. David Blaauw and Prof. Dennis Sylvester. I received my B.S. from the University of Electronic Science and Technology of China.

My work received consecutive William J. McCalla Best Paper Award nominations at ICCAD 2024 and 2025. I was awarded the Radhakrishnan Nagarajan Family Fellowship at UC Santa Barbara and won the Second Place at the DAC 2026 Ph.D. Forum.

Email Google Scholar GitHub LinkedIn Twitter ORCID

Experience

Meta AI

2024/6–2024/9

Research Intern

Explored Multi-teacher knowledge distillation for multi-modal and multi-task foundation models supporting integrity detection across Facebook, Instagram, and WhatsApp.

Selected Publications

All publications

I have published in leading conferences and journals in machine learning, computer architecture, and design automation, including ISCA, NeurIPS, AAAI, CVPR, ICCV, ICCAD, TCAD, and JSSC.

Efficient Generative Modeling

NeurIPS 2026
Self-Forcing
BF16
VAR-Q INT4
67.49 GB16.95 GB (↓74.88%)
InfinityStar 720p
BF16
VAR-Q INT4
62.72 GB16.00 GB (↓74.49%)

VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation

Boxun Xu*, Jiaji Lu*, Yuxuan Yin, Zihu Wang, Ziyue Liu, Yu Wang, Zirui Liu, Peng Li

Conference on Neural Information Processing Systems (NeurIPS), 2026

Large KV caches limit the scalability of autoregressive image and video generation. First tuning-free KV-cache quantization for visual autoregressive generation, evaluated across nine image and video models.

[Paper][Code][Project]
AAAI 2026
AMS-KV original mechanism: cache-efficient and cache-demanding layers keep different token budgets.

AMS-KV: Adaptive KV Caching in Multi-Scale Visual Autoregressive Transformers

Boxun Xu, Yu Wang, Zihu Wang, Peng Li

AAAI Conference on Artificial Intelligence (AAAI), 2026

Different layers and image scales need different cache budgets to preserve generation quality. First adaptive KV-caching framework tailored to multi-scale visual autoregressive transformers.

[Paper]

AI and System Co-design

ISCA 2025 Bishop heterogeneous accelerator overview DRAMGLBWorkloadSchedulerMemory busCtrl busCtrl busMemory busCtrl busNeuromorphic TransformerAcceleratorDenseTTB CoreTTBRouterSparseTTB CoreSeq. SpikeGenerationArrayReconfigurableSpikingAttention ArrayMemory busMemory bus
Dense Array: four by four spatiotemporal systolic processing array Dense ArraySpike TT-BundleBufferWeight BufferOutput BufferPEPEPEPEPEPEPEPEPEPEPEPEPEPEPEPESpatiotemporalSystolic Array Attention Array: reconfigurable three by three spiking processing array Attention ArrayK/V Spiking TT-bundleBufferQ Spiking TT-bundle BufferY Output BufferPEPEPEPEPEPEPEPEPE

Bishop: Sparsified Bundling Spiking Transformers on Heterogeneous Cores with Error-Constrained Pruning

Boxun Xu, Yuxuan Yin, Vikram Iyer, Peng Li

International Symposium on Computer Architecture (ISCA), 2025

Spiking transformers produce workloads with different activity densities across tokens and time. First heterogeneous HW/SW co-design framework for brain-inspired neuromorphic transformers.

[Paper][Talk]
ICCAD 2025
Original Figure 1(b): four spiking expert cores and a router mapped to a 3D IC.
Original Figure 1(b): four spiking attention cores and a router mapped to a 3D IC.

3D Acceleration for Mixture-of-Experts and Multi-Head Attention Spiking Transformers with Dynamic Head Pruning

Boxun Xu, Junyoung Hwang, Pruek Vanna-iampikul, Yuxuan Yin, Sung Kyu Lim, Peng Li

ACM/IEEE International Conference on Computer-Aided Design (ICCAD), 2025

William J. McCalla Best Paper Award Nomination

Mixture-of-experts expands model capacity through conditional computation, creating a need for parallel, distributed hardware. First 3D distributed and parallel accelerator for spiking MoE and multi-head attention.

[Paper]
ICCAD 2024
Spiking attention dataflow across stacked memory and compute planes.

Spiking Transformer Hardware Accelerators in 3D Integration

Boxun Xu, Junyoung Hwang, Pruek Vanna-iampikul, Sung-Kyu Lim, Peng Li

ACM/IEEE International Conference on Computer-Aided Design (ICCAD), 2024

William J. McCalla Best Paper Award Nomination

Spiking transformers combine attention with binary spike operations, but require dedicated hardware to realize their efficiency. First dedicated 3D accelerator for brain-inspired neuromorphic transformers.

[Paper]

Honors & Awards

  • 2026Second Place, DAC Ph.D. Forum
  • 2026Radhakrishnan Nagarajan Family Fellowship, UC Santa Barbara
  • 2026Gold Reviewer, ICML
  • 2025 William J. McCalla Best Paper Award Nomination, ICCAD
  • 2024 William J. McCalla Best Paper Award Nomination, ICCAD
  • 2023DAC Young Fellow
  • 2023ECE Outstanding Teaching Assistant Award, UC Santa Barbara

News

  1. VAR-Q was accepted at NeurIPS 2026.ProjectPaper

  2. Defended my Ph.D. dissertation, “Co-Architecting Models and Systems for Efficient Perception and Generation,” at UC Santa Barbara.

  3. Received Second Place at the DAC 2026 Ph.D. Forum and a travel grant from ACM SIGDA / IEEE CEDA.

  4. Recognized as an ICML 2026 Gold Reviewer.

  5. Released the Sparse Forcing preprint with Meta Superintelligence Labs.ProjectPaper

  6. Received the Radhakrishnan Nagarajan Family Fellowship at UC Santa Barbara.

  7. Completed my internship and student research role at Meta Superintelligence Labs, working on efficient video generation.

  8. VEGAS was accepted at CVPR 2026 Findings.

  9. AMS-KV and Khan-GCL were accepted at AAAI 2026.

  10. Received a second consecutive William J. McCalla Best Paper Award nomination at ICCAD 2025 for the 3D MoE spiking transformer paper.

  11. Joined Meta in Seattle as a research intern, working on efficient video generation.

  12. The 3D MoE spiking transformer paper was accepted at ICCAD 2025.

  13. The Vmin prediction paper was accepted at ITC 2025.

  14. The spiking vision transformer quantization paper was accepted at ASAP 2025.

  15. Bishop was accepted at ISCA 2025.

  16. SpikeX was accepted as a long paper in TCAD.

  17. Received a William J. McCalla Best Paper Award nomination at ICCAD 2024 for the 3D spiking transformer paper.

  18. Two papers were accepted at ICCAD 2024, on 3D spiking transformers and LLM-guided analog design.

  19. Joined Meta AI as a research intern, working on multimodal, multitask knowledge distillation.

  20. AIMMI was accepted in JSSC.

  21. Selected as a DAC Young Fellow.

  22. Received the ECE Outstanding Teaching Assistant Award at UC Santa Barbara.

  23. Started my Ph.D. at UC Santa Barbara, advised by Prof. Peng Li.

  24. The 10 TOPS/W multimodal SoC paper was accepted at the VLSI Symposium 2022.

Service

Conference Reviewer & Program Committee
NeurIPS, ICML, ICLR, AAAI.
Journal Reviewer
IEEE TPAMI, IEEE TCAD, IEEE TCAS-AI, IEEE TCAS-II, ACM TODAES, ACM TACO, IEEE TC, Pattern Recognition, Neural Networks, Neurocomputing, Integration, JSA.