Personal website  ·  Latest paper  ·  Merged pull requests  ·  Email

_{Research updated 4 November 2026 · Public merged PRs outside my personal namespace, verified 20 October 2026}

WSDM 2027 Findings · Accepted · First author: Xi Qin

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.

Latest preprint · Distribution testing · 23 October 2026

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.

Read the paper ↗ · LaTeX source ↓ · DOI: 10.5281/zenodo.23584756

Preprint · Learning theory · 3 October 2026

How much can you lose by keeping just three examples? I determine the exact worst-case risk inflation for mean regression when selecting three support points, with repetition allowed, and predicting their equally weighted mean—in every dimension.

The result resolves the n = 3 column of the data-selection question posed at COLT 2025, with sharp constants and a characterization of the extremal geometry.

Read the paper ↗ · Source & verification scripts ↓ · DOI: 10.5281/zenodo.23119254

Preprint · AI evaluation · arXiv:2608.19269 · v3

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.

Read on arXiv ↗ · Code ↗

14 PRs merged since 2 October. Selected work across query engines, developer tools, scientific computing, and agent security.

Also just merged: SysKnife #612 — detect credentials after escaped boundaries · 19 October.

Browse all 67 public merged PRs →

Pocket MOSS Image · PYTHON / PYTORCH
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.

evalspec-audit · PYTHON
Reproducible audits of how evaluation choices affect scores, rankings, and the claims a benchmark supports.

FastRead · PYTHON / TS
A paper-reading workspace with page-linked evidence, reports, notes, and conversations grounded in the original PDF.

MoonCat · 月薪喵 · C / OPENVELA
An openvela wearable recovery coach: LVGL interface, scheduled reminders, and on-device INT8 intent routing.

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

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

openvela-smarthome-agent · C / OPENVELA
A smart-home agent with simulated devices, a DeepSeek tool-calling loop, and a reproducible offline mode.

Competition background & details / 赛事介绍 →

Works in Rust · Python · TypeScript · C · MoonBit · openvela / LVGL

Nihil ultra testimonium. — Nothing beyond the evidence.

_{Nullius in verba.  ·  qinxi-ai.github.io}