Canyu Zhang

📍 New York, NY

Hi! I am a senior at NYU Shanghai studying Computer Science (GPA: 3.93/4.0). I’m currently studying away at NYU in New York. I’m especially excited about the vision of enabling AI systems to better understand and interact with the real world, connecting advances in machine learning with applications in perception, reasoning, and intelligent agents.

I am currently working with two advisors. With Shengjie Wang, my work explores methods for improving the efficiency and scalability of large AI systems, including projects on KV cache recomputation (InfoFlow KV^*, ICML 2026) and submodular context compression (Sub-CP). With Mengye Ren, I am exploring how video generative models can serve as a form of visual memory.

My current interests include:

Efficient LLM inference — KV cache management, attention approximation, and speculative decoding for scalable long-context reasoning

Multimodal & Generative Models — exploring how vision-language models and video generative models can support richer perception, reasoning, and visual memory

Embodied AI & Intelligent Agents — connecting perception, reasoning, and action in interactive environments

I’m always excited to explore new ideas across these areas. I am currently looking for research internships and PhD opportunities starting in 2027. Feel free to reach out!

publications

ICML
InfoFlow KV: Information-Flow-Aware KV Recomputation for Long Context

Xin Teng^*, Canyu Zhang^*, Shaoyi Zheng^*, and 3 more authors

arXiv preprint arXiv:2603.05353, 2026

*Co-first author. Accepted at ICML 2026.

arXiv

Retrieval-augmented generation for long-context QA is bottlenecked by inference-time prefilling over large retrieved contexts. We propose InfoFlow KV, casting selective KV recomputation as an information flow problem. An attention-norm signal from the query identifies tokens that are both semantically relevant and structurally positioned to propagate information. We further introduce an information-flow-guided chunk reordering strategy, achieving consistent gains over prior methods on LLM and VLM benchmarks.
@article{teng2026infoflow, title = {InfoFlow {KV}: Information-Flow-Aware {KV} Recomputation for Long Context}, author = {Teng, Xin and Zhang, Canyu and Zheng, Shaoyi and Zhuo, Danyang and Zhou, Tianyi and Wang, Shengjie}, journal = {arXiv preprint arXiv:2603.05353}, year = {2026}, note = {*Co-first author. Accepted at ICML 2026.}, }
Preprint
Submodular Context Partitioning and Compression for In-Context Learning

Shaoyi Zheng, Canyu Zhang, Tianyi Zhou, and 1 more author

arXiv preprint arXiv:2510.05130, 2025

arXiv

We propose Sub-CP, a block-aware context selection framework for in-context learning that leverages submodular objectives to control block diversity, balancing global diversity and local coherence. Extensive experiments show Sub-CP consistently improves performance across model scales, achieving up to +7.1% average F1 over baseline partition methods and +15.4% F1 on TREC.
@article{zheng2025subcp, title = {Submodular Context Partitioning and Compression for In-Context Learning}, author = {Zheng, Shaoyi and Zhang, Canyu and Zhou, Tianyi and Wang, Shengjie}, journal = {arXiv preprint arXiv:2510.05130}, year = {2025}, }
Preprint
SAGE: Benchmarking and Improving Retrieval for Deep Research Agents

Tiansheng Hu, Yilun Zhao, Canyu Zhang, and 2 more authors

arXiv preprint arXiv:2602.05975, 2026

Under review at EACL 2027.

arXiv

We introduce SAGE, a benchmark for scientific literature retrieval with 1,200 queries across four domains and a 200K-paper corpus. We evaluate six deep research agents and find BM25 outperforms LLM-based retrievers by 30%. We propose a corpus-level test-time scaling framework that augments documents with metadata and keywords, yielding +8% and +2% gains on short-form and open-ended questions.
@article{hu2026sage, title = {{SAGE}: Benchmarking and Improving Retrieval for Deep Research Agents}, author = {Hu, Tiansheng and Zhao, Yilun and Zhang, Canyu and Cohan, Arman and Zhao, Chen}, journal = {arXiv preprint arXiv:2602.05975}, year = {2026}, note = {Under review at EACL 2027.}, }