Chuyue Wang
LLM inference & agents. Penn '29.
I make LLM inference faster at Penn's Distributed Systems Lab, and test whether models can be trusted to review scientific papers at Wharton. What breaks along the way gets fixed upstream, in SGLang and vLLM.
steven’s thinking
I work on the parts of an agent that have to hold up when nobody is checking. Sparse attention that cuts KV-cache reads without trading away accuracy. A benchmark that asks whether a model can referee a paper it has never seen. A memory layer that lets an agent reuse what it worked out last week.
The hard part is never the idea. It is the latency budget, the stale state, and the instruction that only half specifies the job.
Most of it runs on SGLang and vLLM. When they break under me I fix them there rather than around them, because the next person to hit it should find it already gone. Inference is where an idea meets the hardware, and I intend to stay there.
Internships and research, most recent first. Open a row for the full account.
Internships
Software Engineering Intern
Kosa turns a company's inbox into the deals it is about. I built the layer that commits each suggestion to the right tenant exactly once, and shipped MiniKosa, a macOS copilot with on-device transcription.
- TypeScript
- Electron
- Whisper
- FTS5
Research Engineering Intern
Co-authored Traj2Mem, a procedural memory that lets an agent reuse strategies from its past trajectories without retraining, and owned its retrieval and evaluation pipelines. Now in the institute's technology-transfer base, which works with ByteDance, OPPO and other companies.
- Python
- Agent Memory
- Retrieval
Agent Engineering Intern
Built the record-and-replay harness that made the Engine agents' tests deterministic, and the assertion pipelines that decide whether a replay passed.
- Python
- Redis
- PostgreSQL
- WebArena
Founding Engineer
A networking agent that lives inside iMessage. I built the backend: LangGraph orchestration, hybrid vector-and-graph matching, durable queues and retries.
- LangGraph
- FastAPI
- Redis
Derivatives Valuation Intern
Gaussian Heath–Jarrow–Morton models for swaption pricing, and a rewrite of the desk's daily exotic-equity valuation that cut a run from two days to four hours.
- Python
- Numerical Methods
Research
Research Engineer, advised by Prof. Bingxin Zhao
Built the post-cutoff benchmark that asks whether a model can referee a paper it has never seen. The finding: what a model remembers about an author can outweigh the paper in front of it.
- Python
- LLM Eval
- Benchmarks
Research Assistant, advised by Prof. Vincent Liu
TreeSparseAttention: hand-written CUDA/C++ scoring kernels on FP8 that page the KV cache by what a query needs, served through SGLang.
Co-developed FullStackArena, a browser-agent benchmark across six stateful apps. Under review at NeurIPS 2026.
- CUDA/C++
- FP8
- SGLang
- FlashInfer
Research Software Engineer, advised by Prof. Hsiao-Wen Cheng
Built and deployed yijianzhi.net, a trilingual database of 2,745 Song-dynasty stories used by 120+ scholars. Selected for the 2026 GPASC Undergraduate Summer Grant, one of two projects from 100+ submissions.
- Flask
- PostgreSQL
- NLP
Under submission. Titles are withheld where double-blind review applies.
A mysterious paper on Agent Memory Retrieval
Chuyue Wang, …
Title and co-authors withheld under double-blind review.
A mysterious paper on 3D Gaussian Splatting and Reinforcement Learning
…, Chuyue Wang*, …
*Equal contribution. Title and co-authors withheld under double-blind review.
FullStackArena: Evaluating Browser Agents in Dynamic and Temporally Evolving Full-Stack Environments
Jiaheng Lu, Xian Wang, Shrishti Roy, Colin Zhao, Sidharth Sankhe, Chuyue Wang, Xuting Liu, Spyros Pavlatos, Linh Phan, Vincent Liu
Open the paper (PDF)Nights and weekends, my own.
Visit SILKern. (opens in new tab)
Sparse-index localization kernels for context-parallel decode — deterministic, allocation-free, CUDA-graph-safe.
Hand-written CUDA kernels that localize sparse KV indices under context-parallel decode — deterministic, allocation-free, and CUDA-graph-safe, so they can sit inside a captured decode graph without ever touching the allocator.
- CUDA
- Python
Visit Township (opens in new tab)
AI residents deliberate civic questions in a living pixel town — a replayable, scenario-first simulation engine.
Visit AutoPaperLab (opens in new tab)
Turns research briefs into auditable paper packages — verified citations, reproducible experiments.
A research-to-paper pipeline built on one rule: every citation, number, and result has to trace back to something real, or the draft does not ship.
- Python
- LLM Agents
Fixes sent upstream to the inference stacks I build on.
SGLangVisit SGLang on GitHub (opens in new tab)
Contributor to sgl-eval, the project's accuracy evaluation harness
sgl-eval is how SGLang checks that a served model still scores what it should. My patches fix chat-template preset precedence, pin MMLU to a verified archive, and reject duplicate example IDs before a run starts. Two fixes to SGLang itself, on the FlashInfer CUDA-graph decode path and the Responses API, are in review.
- sgl-eval2 merged
- Python
- Eval Harness
vLLMVisit vLLM on GitHub (opens in new tab)
Contributor to vLLM, its FlashAttention fork, vllm-metal and agentic-api
In vLLM itself, the attention backend's sync with upstream FlashAttention, paired with bringing the project's own FlashAttention fork up to Dao-AILab's ce088ab9: 33 upstream commits landed through one pull request. In vllm-metal, the Apple Silicon backend: paged attention for EXAONE 4.0 and StableLM, BF16 LoRA adapters loaded natively in MLX, token logprobs, MLX ring bootstrap, and Whisper chunking. In agentic-api, the stateful layer for agent apps: Responses reasoning and text configuration, streamed-reasoning reconciliation, relayed call IDs, and proxy-header handling.
- vllm1 merged
- flash-attention33 merged to flash-attention (opens in new tab)
- vllm-metal6 merged
- agentic-api7 merged
- Python
- CUDA
- Rust
- MLX
Everything else.
Chuyue Wang
Philadelphia · wangcy07@sas.upenn.edu
- GPASC Undergraduate Summer Grant — one of two projects selected from 100+ submissions2026
- 1st Place, Democratic Governance — Anthropic × University of Pennsylvania AI Hackathon2026
- 1st Place — Avalanche Team 1 × Foundry Start-Up in a Weekend, Hybrid Intelligence Track2025
- Finalist (Top 11 of 967) — High School Mathematical Contest in Modeling, COMAP2024
- National Gold Medal — Physics Bowl, AAPT2023