Chuyue Wang

Can a model referee a paper it has never seen? I built the post-cutoff benchmark that tries to answer it — 400 papers, close to 20,000 evaluations — and the causal studies that pull apart reasoning, memorization, and role-play. On anonymized papers it ranks citations correctly 82.5% of the time. The uncomfortable finding is elsewhere: what a model remembers about an author can outweigh what the paper in front of it actually says.

• Python
• LLM Eval
• Benchmarks

Co-authored Traj2Mem, a procedural memory that mines an agent's past trajectories for the strategies that actually worked, then replays them when a new request is underspecified. No retraining. I owned the trajectory-processing, retrieval, and evaluation pipelines that keep the cross-model ablations reproducible. The results have since gone into the institute's technology-transfer base, which works closely with ByteDance, OPPO and other companies.

• Python
• Agent Memory
• Retrieval

Agent tests were flaky, because the payloads were encrypted and no two runs were alike. I shipped the Case Test Service — record once, replay deterministically, inject fixtures at runtime — and the 191 assertion pipelines that decide whether a replay actually passed. I also built the parallel evaluation harness on Redis and Postgres, and a context assembler that merges ranked vector search, SQLite FTS5, and live OS signals fast enough to sit in the request path.

• Python
• Redis
• PostgreSQL
• WebArena

TreeSparseAttention parses a prompt into a semantic tree, then pages the KV cache by what the query actually needs. I wrote the scorer by hand in CUDA/C++ on cuBLASLt FP8, where it runs 20.9× faster at batch 128 and latency stays nearly flat as batch size climbs, which is where serving cost actually goes. Profiling Qwen3-VL on B200 and RTX 5060 Ti cut decode latency 31% and held 1.67× SGLang throughput at batch 4.

Before that, FullStackArena: 1,000+ browser-agent tasks across fintech, ride-sharing, and mapping apps, scored by LLM judges and checked against real DOM and API state. Under review at NeurIPS 2026.

• CUDA/C++
• FP8
• SGLang
• FlashInfer

Eight hundred years of Song-dynasty writing, 2,700+ works, catalogued by hand one at a time. I built a Flask and PostgreSQL platform that pulls entities out with NLP and scores its own confidence before anything is committed, so a human only reviews what the model is unsure of, and about 70% of the hand cataloguing goes away. The Greater Philadelphia Asian Studies Consortium picked it as one of two projects from over 100 submissions.

• Flask
• PostgreSQL
• NLP

A networking agent that lives inside iMessage, with no app to install. LangGraph runs the orchestration; matching runs on hybrid vector and graph search over people and what they have worked on. I built the backend for the case where things go wrong: durable queues, isolated failures, backoff retries, and rate limits on FastAPI, Redis, and Postgres.

• LangGraph
• FastAPI
• Redis

Implemented Gaussian Heath–Jarrow–Morton models for interest-rate swaption pricing, cutting pricing error about 15%. Then rewrote the exotic structured-equity valuation path the desk ran every day, taking a run from two days down to four hours.

• Python
• Numerical Methods