Multimodal models and agents, and how to train, measure, and ship them.
Hi, I'm Sheng-Tse, and you can also call me Hunter. I'm a master's student at Carnegie Mellon University, where I do research with Prof. Bhiksha Raj and Massa Baali at the Language Technologies Institute. I primarily work on open source, agents, and applied research, and this site is mostly a place to keep track of what I'm building and learning along the way.
My research focuses on multimodal models and agents, with a particular emphasis on agentic reinforcement learning, efficient LLM systems, and making models explainable and rigorously evaluated.
My fix to FlashInfer’s Mamba state-update kernels was merged, so head dims that aren’t multiples of 64 now compile and give correct results.
Released Argonauts Code, one orchestrator for every coding agent: Claude Code, Codex and others build in parallel and review each other’s work.
Released the code and data for AcoustiClaim. The paper is under review at ICASSP 2027.
Joined Smile Line AI as a Forward Deployed Engineer.
Released jupynvim, Jupyter notebooks and remote kernels in Neovim.
@misc{lin2026acousticlaim,
title = {AcoustiClaim: A Numeric Claim Benchmark with Instrument Ground Truth},
author = {Lin, Sheng-Tse and Zhai, Siyuan and Kuo, Chien-Liang and Baali, Massa and Raj, Bhiksha},
year = {2026},
eprint = {2609.30483},
archivePrefix = {arXiv},
primaryClass = {cs.SD},
note = {Under review at ICASSP 2027}
}* equal contribution
A terminal app that runs the coding agents you already use, from Claude Code and Codex to open-weight models, as one crew. They build in parallel and review each other's work, and nothing merges without you or the rules you set.
Jupyter notebooks and remote kernels in Neovim, built on a Rust backend that speaks the Jupyter wire protocol directly.
Thanks for scrolling all the way down. Say hi anytime.