XQXi QinLet's talk

Research & open-source engineering

From first
principles
to working code.

I'm Xi Qin, at Wuhan University. I work on learning theory, reproducible AI evaluation, and the software that puts ideas to work.

Explore the research Find me on GitHub Latest paper · WSDM 2027 Findings
FIELD NOTES / 2026FIG. 01
A geometric study of points, a simplex, and their meanx₁x₂x₃x₄μPOINTS → STRUCTURE → PROOF
Nullius in verba.

Ideas worth proving.
Systems worth building.

04Research papers
1 accepted · 3 preprints67Public PRs merged
27Repositories contributed to
04Competition awards
01

Selected research

Questions, made precise.

Sharp mathematical answers and evaluations you can inspect.

AcceptedWSDM 2027 Findings

Failure Steps Are Not Repair Intents: Measuring Actionable Ambiguity in Multi-Agent System Diagnosis

Do different blamed execution steps imply different repairs?

We study expert-recommended actions from MP-Bench through a prediction-blind, human-audited model-assisted analysis of 96 multi-agent traces and 1,436 actions.

In the primary analysis, 85.1% of traces contain fewer distinct repair intents than blamed steps, while top repair-set alignment reaches 95.7%. The findings show how step-level evaluation can overstate actionable diversity while often preserving the first repair priority.

First author: Xi QinWSDM 2027 Findings · Accepted

Latest preprint · Distribution testing · 23 October 2026

Instance-Optimal Identity Testing at the Original Total-Variation Radius: Resolving Canonne's 2024 Open Question 1

How many samples are needed to test whether an unknown distribution equals a specified reference—at the total-variation radius actually given? I give an explicit characterization, up to universal constants, for every finite or countable reference distribution at the original radius.

The preprint resolves Canonne's COLT 2024 Open Question 1 in the countable-alphabet model, with matching upper and lower bounds and a polynomial-time factor-two approximation for explicit rational input.

Xi Qin · Wuhan UniversityRead the paper LaTeX source DOI 10.5281/zenodo.23584756
PreprintLearning theory · October 2026

Exact Three-Example Selection in Every Dimension

How much can you lose by keeping just three examples?

For mean regression, this paper determines the exact worst-case risk inflation of selecting three support points, with repetition allowed, and predicting their equally weighted mean—in every dimension.

It resolves the n = 3 column of the data-selection question posed at COLT 2025, and characterizes how the extremal geometry changes with dimension.

Xi Qin · Wuhan University 03 October 2026
Read the paper Source & verification
DOI 10.5281/zenodo.23119254

AI evaluation · arXiv:2608.19269 · v3

What Does an Evaluation License?

A Commit-Bound Census of Claim-Relative Inference in Inspect Evals

A framework for connecting evaluation artifacts to the claims they support. The census records 124 eligible units: 14 deterministic analyses and 110 typed stops that identify missing evidence.

Xi Qin · Wuhan UniversityRead on arXiv Explore the code

Research updated .

02

Open-source contributions

Built together. Merged upstream.

From query engines to developer tools and agent security.

14 contributions merged since 2 October.

Browse all 67
01 / Query enginesMerged

apache / datafusion #25765

Give lazy generators a serialization hook

Let LazyMemoryExec generators encode themselves while preserving the existing single-generator wire format, extension fallback, and repeatable execution after a round trip.

Legacy wire compatibility · round-trip execution tests
02 / Developer toolsMerged

microsoft / vscode-eslint #2227

Keep linting inside the workspace

Added an opt-in setting to skip local files outside the workspace, while retaining standalone-file, untitled, and virtual-document behavior.

30 client/server tests · workspace boundary regressions
03 / Agent securityMerged

lacs-project / sysknife #573

Redact credentials before tools reach the model

Screen read-tool results and retained provider context for plaintext credentials, preserving useful log records and ordinary host information.

Recording-provider regression · 1,949 Linux tests
04 / Query enginesMerged

apache / datafusion #25150

Skip hashing for fully ordered group keys

Added a columnar path for supported, fully ordered multi-column aggregations. Adjacent equal keys share a group without a hash table, reducing grouping work and retained state.

Differential tests · cross-batch coverage · reproducible benchmarks
05 / Scientific PythonMerged

pydata / xarray #11552

Bring apply_ufunc to DataTree

Extended array-function dispatch to complete DataTree structures, preserving child datasets, inherited coordinates, core dimensions, and multiple outputs.

323 targeted tests passed · tree-structure regressions
06 / Numerical methodsMerged

awslabs / stickler #284

Replace sklearn metrics with NumPy

Implemented ROC AUC, Rand index, V-measure, and Kendall tau-b in NumPy, removing scikit-learn from the core dependencies while checking results against reference implementations.

130 affected/oracle tests · 1,947 portable tests passed
07 / Data systemsMerged

AndreaBozzo / dataprof #655

Read and profile only selected columns

Applied column projection across Rust and Python profiling APIs, including streaming and Parquet paths. Physical Parquet projection avoids reading unselected column chunks.

Rust workspace tests · 353 Python tests · byte-range regressions
08 / GeometryMerged

3b1b / manim #2501

Fix batched 2D vector rotation

Added planar rotation for batches of 2D vectors, restoring the VectorField curl example while preserving the existing rotation behavior for 3D vectors.

Four 2D/3D regressions · real VectorField reproduction

Also just merged: SysKnife #612 — credential detection after escaped boundaries

Public merged PRs outside the QinXi-ai personal namespace. Snapshot: .

03

Selected projects

Tools for the work ahead.

Research software, developer tooling, and small devices with useful jobs.

01

Pocket MOSS Image

Low-memory MOSS-Image inference: the original 9B text encoder with packed INT6 weights and a streamed, fused INT8 DiT. No retraining or replacement encoder.

Python / PyTorch02

evalspec-audit

Reproducible audits of how evaluation choices affect scores, rankings, and the claims a benchmark supports.

Python03

FastRead

A paper-reading workspace with page-linked evidence, research reports, notes, and conversations grounded in the original PDF.

Python / TypeScript04

MoonCat · 月薪喵

An openvela wearable recovery coach combining an LVGL interface, scheduled reminders, and on-device INT8 intent routing.

C / openvela05

moonguard

Public API compatibility checks and semantic-versioning safeguards for MoonBit libraries.

MoonBit06

AstroTwin · public demo

A browser-based astronomy research demo with TESS evidence snapshots, guided analyses, and optional Qwen-assisted research planning.

JavaScript07

openvela-smarthome-agent

A smart-home agent with simulated devices, a DeepSeek tool-calling loop, and a reproducible offline mode on the openvela emulator.

C / openvela
04

Selected recognition

A few milestones.

2026 · Competitions & awards / 竞赛与获奖

First Prize一等奖

CCF Open Source Innovation Competition

CCF 开源创新大赛

openvela Smart Band Track / openvela 智能手环赛道

About the competition / 赛事介绍

Wearable application development on openvela, covering embedded systems, interaction design, and integration. Organized by the China Computer Federation (CCF); the 2026 competition received 586 valid entries and awarded 34 first prizes across all tracks.

基于 openvela 开发智能手环应用，涵盖嵌入式开发、交互设计与系统集成。中国计算机学会（CCF）主办；2026 年全赛共收到 586 份有效作品，评出 34 项一等奖（所有赛道合计）。

Official competition details
Third Prize三等奖

CCF Open Source Innovation Competition

CCF 开源创新大赛

openvela Smart Home Track / openvela 智能家居赛道

About the competition / 赛事介绍

Develop smart-home AI agents that connect natural-language requests with device tools and control workflows. This CCF competition track tests integration across language models, embedded software, and open-source systems.

围绕 openvela 构建智能家居 AI Agent，将自然语言请求接入设备工具与控制流程。作为 CCF 开源竞赛赛道，重点考察大模型、嵌入式软件与开源系统的综合实现能力。

Official competition details
First Prize一等奖

CCF Quantum Computing Programming Challenge

CCF 量子计算编程挑战赛

Origin Quantum Cup / 本源杯

About the competition / 赛事介绍

A CCF-organized quantum programming challenge using Origin Quantum’s QPanda3 and VQNet tools. Tasks span circuit design and algorithm implementation, testing the ability to turn quantum theory into working code.

由 CCF 主办的量子计算编程赛事，使用本源量子的 QPanda3、VQNet 等工具，涵盖量子线路设计与算法实现，考察将量子理论转化为可运行程序的专业能力。

Official competition details
Second Prize二等奖

National Undergraduate Electrical Mathematical Modeling Competition

全国大学生电工数学建模竞赛

CSEE Cup · “Diangong Cup” / 中国电机工程学会杯 · 电工杯

About the competition / 赛事介绍

A national competition organized by the Chinese Society for Electrical Engineering. Teams of three have 72 hours to model an applied problem and submit a paper, testing mathematical reasoning, computation, and technical writing.

中国电机工程学会主办的全国性竞赛。三人组队，在 72 小时内完成实际问题的建模、求解与论文，综合考察数学分析、计算实现及技术表达能力。

Official competition details