B.Eng. candidate in Artificial Intelligence, Wuhan University (2023–2027)
I build measurement tools — benchmarks that can tell a model which did the task from one that happened to guess right.
jiahuili@whu.edu.cn · Wuhan University · Wuhan, China
Benchmark validity for multimodal models. I co-designed a VLM benchmark for
instrument reading and built its data and evaluation pipeline: 1,821 images, 7
subsets, 32 instrument types, every value and box annotated by a human. Two
results shaped how I work. Agent-generated ground truth inflates reading accuracy
(22.4% vs 14.5%) and flattens prompt effects (+0.11 pp vs +4.12 pp); a
same-model label-and-evaluate loop inflates it by 75%. And the metric the
project had planned to headline turned out to be an arithmetic identity — it
equals exactly half the whole-image reading accuracy. What survived is ec_rate,
the share of correct answers whose localization was wrong: 22.8% (50/219).
Self-evolving data engines. A five-stage annotation engine for animal pose (YOLO-World → SAM2.1 → ViTPose++ → CoTracker3 → quad-triage). The design choice that mattered was making tracking abstain rather than propagate unreliable keypoints: Gold clips 56 → 71, worst failure class 36 → 24.
Label quality in remote sensing. Morphological filters for OSM-derived road labels. At an equal budget of 17,802 samples, morphology-based selection raised recall from 72.18% to 80.15% over random sampling. Second author.
| Repository | What it is |
|---|---|
tape |
Time-travel debugger for AI agent sessions — reads the JSONL your agent already writes, zero dependencies |
agentic-risk-scan |
Dependency-free scanner for agent workflows, MCP servers and instruction-file risks |
model-performance-test |
Frontend benchmark for LLMs: one prompt, one 100-point rubric, several models |
m2html |
Markdown → HTML in Rust, GFM and syntax highlighting, with CLI and Python bindings |
xv6-riscv |
xv6 / RISC-V labs — OS internals in C |
mogwai |
A terminal pet whose brain is a git repo; memory decays, forgetting is reversible |
Python · PyTorch · C / C++ · Rust · JavaScript · Linux · Slurm · LaTeX
Research code lives in private repositories while the work is in progress — happy to walk through any of it on request.