News

Recruiting, projects, publications, and milestones — newest first.

• June 2026Pinned

  Recruiting research interns. I'm helping a group I work with at LMU Munich find research interns. The group works on reinforcement learning and LLM / web agents (a recent paper accepted at ICLR 2026) and is continuing RL work on vertical agent applications. Remote; ample compute provided for RL / LLM work.

  Looking for: students able to commit full-time (≥40 h/week) or part-time (≥20 h/week), with substantial hands-on RL experience — you know the theory — and able to build agent systems independently. Strong contributors typically earn co-authorship on resulting work.

  Start: rolling — begin anytime; budget ~1 week to ramp up. Interested? Email me at leo@lsamc.website with your CV and a brief description of a relevant project you've built (a GitHub repo link is welcome) — RL / agent ideally, though strong project experience in other areas is also welcome. Happy to chat first.

• August 2026

  📄 Our paper QuoteBench: How Matched Scores Can Hide Command-Path Failures is now on arXiv (with Yao Zhang, Volker Tresp, and Yuanyuan Yang) — showing that matched execution scores for LLM coding agents can hide failures on the command path, so evals must report the generation contract, execution path, and validator, not just the score. Project site.

• July 2026

  🔨 Introducing Human-Crafted Software — a badge and spec for projects you designed and wrote yourself, with no AI generating or editing the committed source code. Proudly human-made — a small counterpoint to vibe-coding.

• July 2026

  🧹 New skill: code-humanizer — an agent skill that strips AI-generated structural slop from a codebase (duplicated helpers, single-implementation abstractions, swallowed exceptions, narrating comments — 16 patterns, behavior-preserving). GitHub · technical details.

• July 2026

  🧪 QuoteBench is live — a focused benchmark for shell quoting and escaping in LLM agents. Its 56 execution-verified tasks and three tool contracts separate raw shell skill from failures introduced by JSON and wrapped agent harnesses. GitHub.

• July 2026

  📘 The discrete mathematics textbook 《离散数学：思维与挑战》 has been published. I served as a proofreader for the two chapters on group theory and combinatorics.

• June 2026

  🌉 Say hello to CaveBridge! A natural-language interface that brings the 1977 Colossal Cave Adventure to modern players — an LLM front-end over the original deterministic engine, so the world model never drifts. GitHub · details.

• June 2026

  🎓 I graduated! Earned my Bachelor of Science in Information and Computing Sciences (Strengthening Basic Disciplines Plan of Computer Science) from Nanjing University.

• June 2026

  🎉 Celebrating that VocoType, my offline voice-to-text Chinese input method for Linux, has reached 160 stars on GitHub!

• April 2026

  🎓 Excited to announce I'll be joining Stony Brook University's ECE department as a PhD student in Computer Engineering.

• April 2026

  🔬 Say hello to LangMatch — an empirical study of whether a denser prompt language (Classical Chinese) buys a better quality–cost frontier for LLMs, under fixed tasks and model backends. details · GitHub.

• December 2025

  🎙️ Introducing VocoType-linux — an offline Chinese speech-to-text input method with full Linux support (IBus & Fcitx5), running 100% locally. GitHub · details.