Yuan Si

Yuan Si

AI4SE Researcher · University of Waterloo

Math / Data Science & Combinatorics Optimization & Pure Math

About

I am a research assistant at the University of Waterloo, working on AI for software engineering: I combine large language models, multimodal signals (video, execution traces, code), and formal methods to debug, tutor, evaluate, and verify block-based programs — 9 first-authored papers, including VisionScratch (published at FSE 2026) and ScratchEval (accepted at ISSTA 2026), with recent work on behavioral equivalence, schedule robustness, certificate-carrying transformation, and robustness of programming-by-example. On the side, I work on analytic number theory and probability (4 preprints): the Collatz affine random model is under review at Forum of Mathematics, Sigma, and the first of a three-paper series on rational-distance problems — Guy's four-corner problem (D19) — is under review at the Journal of Number Theory. Previously worked at Microsoft Azure & AI Research.

Research Interests

LLM-based Program Repair Multimodal Debugging AI for Computing Education Evaluation of LLMs Program Analysis & Verification Analytic Number Theory Probability Theory

Publications

Conference Papers
VisionScratch: LLM-Based Automated Feedback Generation using Code-Produced Videos for Scratch Programs
Yuan Si, Daming Li, Hanyuan Shi, Jialu Zhang
Proceedings of the ACM on Software Engineering, Vol. 3 (FSE 2026), pp. 3534–3557 Published
ScratchEval: A Multimodal Evaluation Framework for LLMs in Block-Based Programming
Yuan Si, Simeng Han, Daming Li, Hanyuan Shi, Jialu Zhang
ACM SIGSOFT International Symposium on Software Testing and Analysis (ISSTA) 2026 Accepted
Journal Submissions
A Microcanonical Phase Transition for the Collatz Affine Random Model
Yuan Si
Submitted to Forum of Mathematics, Sigma, 2026 — sharp resonant phase transition at the entropy line for Tao's Syracuse affine model, reducing hard-frequency mixing to a primitive ternary Bernoulli-bridge transform Under Review
Mixed Parity, Diagonal Denominator, and the Pell-Chord Genus-Five Obstruction for the Four-Corner Rational Distance Problem
Yuan Si
Submitted to Journal of Number Theory, 2026 — unconditional necessary conditions on Guy's problem D19 (Pillai's unit-square four-distance) and a reduction of the residual obstruction to a non-fixed Pell-chord involution on a family of arithmetic-genus-five curves Under Review
Preprints
Fixed-Set Robustness in Programming by Example: Example Corruption and Semantic Partition Recovery
Yuan Si, Jialu Zhang
Preprint, 2026
Checked Program Recovery from Execution Video: A Sound Oracle for Untrusted Generators
Yuan Si, Jialu Zhang
Preprint, 2026
SchedCheck: Schedule-Robustness Analysis for Event-Driven Block Programs
Yuan Si, Jialu Zhang
Preprint, 2026
Certificate-Carrying Transformation of Event-Driven Block Programs
Yuan Si, Jialu Zhang
Preprint, 2026 — soundness proof mechanized in Lean; validated on 300 real Scratch projects
ScratchLens: Lens-Parametric Behavioral Equivalence for Scratch Programs
Yuan Si, Jialu Zhang
Preprint, 2026
EcoScratch: Cost-Effective Multimodal Repair for Scratch Using Execution Feedback
Yuan Si, Ming Wang, Daming Li, Hanyuan Shi, Jialu Zhang
Preprint, 2026 Under Review
Stitch: Step-by-step LLM Guided Tutoring for Scratch
Yuan Si, Kyle Qi, Daming Li, Hanyuan Shi, Jialu Zhang
Preprint, 2025 Under Review
Elliptic Decomposition of the Pell-Chord Genus-Five Obstruction for the Four-Corner Rational Distance Problem
Yuan Si
Preprint, 2026 — structural follow-up to Paper I: decomposes the residual genus-five curve into full-2-torsion elliptic pieces via 2-isogeny and Kani–Rosen Jacobian factorization; reduces the four-corner problem to a two-variable Pythagorean-slope exclusion
An Elementary Projection Obstruction for Rational Distances to Regular Polygons
Yuan Si
Preprint, 2026 — an elementary proof that a regular n-gon (n ≥ 5, n ≠ 6) with rational side length admits no point at rational distance from all vertices, via projection identities, the vertex zero-sum relation, and Niven's theorem
Research Reports & Other
Multiplayer Rock-Paper-Scissors: Nash Equilibria via Linear Programming
Yuan Si
Research Report, 2025
Tesla Charging Station Optimization via Independent Dominating Sets
Yuan Si
Research Report, 2025
Public Goods Game: Cooperation Dynamics and Intervention Analysis
Yuan Si
Research Report, 2025
Textbook
A Gentle Introduction to Optimization
Yuan Si
2024 — Adopted as required reading in 3 university courses

Experience

University Researcher — AI in Multidimensional Input

July 2025 – Present
University of Waterloo
  • Led 9 first-authored papers on LLM-driven debugging, tutoring, evaluation, and formal verification for Scratch (VisionScratch published at FSE 2026, ScratchEval accepted at ISSTA 2026).
  • Built multimodal repair systems fusing gameplay video, execution feedback, and project JSON to localize bugs and synthesize fixes via LLM-guided loops.
  • Developed an interactive tutoring system (Stitch) and a 100-project executable benchmark (ScratchEval) for evaluating LLM repair quality on block-based code.

University Researcher — Game Theory & Graph Theory

Dec. 2024 – March 2025
University of Waterloo, Combinatorics & Optimization Department
  • Derived Nash equilibria for multiplayer Rock-Paper-Scissors via linear programming. [report]
  • Modeled optimal Tesla charging station placement using independent dominating sets. [report]
  • Analyzed cooperation dynamics in Public Goods Games. [report]
  • Investigated strategic dynamics in graph-based games (Cops and Robbers).

Microsoft Researcher — AI Trends

May 2024 – Aug. 2024
Microsoft — Azure & AI, Research
  • Investigated AI scribe technologies for healthcare; synthesized findings into research recommendations for physician workflow automation.
  • Designed and evaluated improvements to an insurance chatbot through systematic analysis of user interaction data.

Projects

ResearchOS

Local-first research decision and execution system. Turns a brief into runnable experiments, evidence-checked claims, and a packaged manuscript draft. Provider-agnostic LLM adapters (OpenAI / Anthropic / mock), dual-agent code worker, HITL approval gates, encrypted secret store.

Python · FastAPI · React · TypeScript · LLM · AGPL-3.0

WizardingWorld

Harry Potter Hogwarts experience mod for Terraria via tModLoader. 590 C# files, 12 multi-phase bosses, Pensieve memory replay framework, three-language localization. Built with Claude Code + Codex in a dual-agent dev workflow.

C# · .NET 8 · tModLoader · MIT

DevToolkit

43 zero-dependency single-file Python CLI tools across 9 categories: web, data, process, MCP, security, scaffolding. Copy any file and run.

Python · CLI · MCP