Xiaoling Zhou | 周小灵
Email: endlesstilly@gmail.com · GitHub · CV
I am an undergraduate in Computer Science and Technology at Zhejiang University’s Chu Kochen Honors College, with an expected graduation date of June 2027.
I will pursue a master’s degree at Zhejiang University’s College of Computer Science and Technology, in the CAD&CG Lab, advised by Prof. Chunhua Shen and Prof. Xuan Shen.
My research focuses on efficient deep learning, with an emphasis on efficient attention, KV-cache optimization, and mixed-precision computation. I work on algorithms and CUDA kernels to reduce inference latency and memory usage while maintaining model quality.
Last updated: November 5, 2026
News
- DraftKV: an ongoing project on efficient long-term memory for video generation.
- Draft Attention v2 and MosaicAttention are under review.
- Mind Dreamer was accepted to ICML 2026. Paper
Education
Zhejiang University, Chu Kochen Honors College
B.Eng. in Computer Science and Technology
GPA: 3.97/4.00
Zhejiang University, College of Computer Science and Technology
Master’s degree
CAD&CG Lab
Advisors:
Chunhua Shen & Xuan Shen
Publications & Manuscripts
2026
-
DraftKV: Efficient Long-Term Memory via Attention-Guided KV Retrieval for Video GenerationOct 2026Project in progress; First Author
-
Draft Attention v2: Fast Video Diffusion with Low-Resolution-Guided Mixed-Precision AttentionSep 2026Under review
-
MosaicAttention: Hierarchical Block-Sparse Attention for Multi-View Geometry TransformersSep 2026Under review
-
Mind Dreamer: Untethering Imagination via Active Causal Intervention on Latent ManifoldsIn International Conference on Machine Learning (ICML), May 2026ICML 2026
Research Experience
DraftKV
Attention-Guided KV Retrieval for Long-Horizon Video Generation
Retrieves local and salient long-range memories from full-precision CPU-resident KV, using low-resolution attention scores to balance long-term memory and inference cost.
- 35.4% lower peak GPU memory on SolarWM-H3 33B, with improved generation quality and long-history consistency.
- 1.21× end-to-end speedup for 60-second LongLive 1.0 generation; 1.91× / 2.51× Fetch+Attention speedups on RTX 4090 / 5090.
Methods & contributions
- Extended DraftAttention2’s low-resolution importance scoring to historical KV retrieval.
- Designed Streaming FlashAttention to overlap host-to-device KV transfer with attention computation.
DraftAttention2
Low-Resolution-Guided Mixed-Precision Attention for Video Generation
Uses low-resolution attention importance estimates to route visual interactions to 16-bit, 8/4-bit, or sparse computation, preserving critical information in few-step diffusion.
- Up to 2.51× / 2.60× DiT inference speedups on HunyuanVideo-13B / MiniMax-H3.
- 1.87× speedup in 8-step video generation while maintaining near-dense attention quality.
Methods & contributions
- Combined average/max dual-path pooling with an attention-output error bound to study mixed precision versus pure sparsity.
- Designed a fused CUDA kernel for quantization, data reordering, and cross-precision Online Softmax, reducing intermediate writes and redundant memory traffic.
MosaicAttention
Sparse Attention for Multi-View Geometry Transformers
Combines 2D average/max pooling and hierarchical sparse routing to retain important cross-view geometry while reducing redundant global attention.
- Up to 3.12× end-to-end inference speedup at 128 input views, while matching or improving accuracy on CO3Dv2 and other multi-view geometry tasks.
Methods & contributions
- Allocated computation across layers while preserving 2D image structure and critical cross-view interactions.
- Incorporated mixed-precision execution to further improve inference efficiency.
Mind Dreamer
Untethering World-Model Imagination via Active Intervention
Addresses Historical Tethering in model-based reinforcement learning by synthesizing high-value latent states as new imagination starting points beyond observed trajectories.
- 1.67× average and up to 8.8× sample-efficiency gains over DreamerV3 on DeepMind Control Suite, with higher final returns on multiple sparse-reward tasks.
Methods & contributions
- Identified how imagination initialized only from observed states limits exploration of unvisited, high-value regions.
- Jointly scored task value and uncertainty to guide active exploration in latent space.
Beyond Token Supervision
Semantic Anchors for Latent Reasoning in LLMs
Uses sparse autoencoders to extract semantic features from an explicit chain-of-thought teacher and supervise intermediate latent reasoning states.
- 11.7% GSM8K accuracy improvement over the COCONUT latent-reasoning baseline, reducing semantic drift.
Methods & contributions
- Identified semantic drift caused by the absence of explicit intermediate supervision in continuous latent reasoning.
- Aligned intermediate latent states with semantic anchors extracted from correct explicit reasoning trajectories.
Awards
-
National Scholarship
-
Zhejiang University First-Class Scholarship
-
Chu Kochen Honors College Leading Scholarship
-
Zhejiang Provincial Undergraduate Physics Competition
Skills
- Programming
- Python, C/C++, CUDA
- Frameworks & Tools
- PyTorch, MuJoCo, Git, LaTeX
Additional Project
ClawKeeper
Security management and work monitoring for AI agents, combining a security control plane, runtime watchdog, and structured operator feedback.
Project details
Risk-aware tool interception and auditing; timeout, stall, loop, and repeated-output detection; runtime health probes and operator notifications.
