Math major & coding enthusiast. Focusing on TCS (theoretical computer science), AI-assisted formal mathematics, and scientific computing. Some projects are for fun, others are research — this page covers both.
Current work — Archon4CR
LLM-agent-driven Lean 4 formalization of computational complexity reductions.
- Framework integration — adapted the Archon multi-agent theorem-proving framework to drive a Lean 4 reduction library beyond its original mathlib target. Shipped source fixes including a write-first protocol that stops prover agents from "figuring it out but never writing it down". Full documentation and test reports in Archon4CR/docs.
- Benchmark — curated a 64-problem NP-completeness reduction set, extracted and cleaned from university course materials.
- Evaluation — ran a 21-iteration automated proving campaign: 3 reductions
formally verified, 0
sorry, axiom-clean. The test report includes an honest analysis of where formal theorems remain weaker than their natural-language claims.
-
gyre-tracking — multi-target tracking on an analytic double-gyre flow. Core: a from-scratch Hungarian (JV) solver whose every matching ships with a verified dual-optimality certificate; around it, a two-pass track lifecycle and a fragment stitcher that reuses the same certified solver one level up. The report measures the certified-vs-greedy gap over 2880 runs and documents the failure modes found on the way — including two sweeps that produced self-consistent garbage until unit tests caught what large-scale experiments could not.
-
TheWitnessSolver — solver and generator for The Witness puzzles; constraint solving with backtracking search.
-
Functionals-py — a functional-programming toolkit for Python: currying, lazy evaluation, function composition and decorators.
Python · Lean 4 (beginner) · C# · TypeScript · Rust · C++ · LaTeX
OS: Windows 11 & WSL2 Ubuntu
Contact: zz_404@mail.ustc.edu.cn
Posts: posts.md

