PithTrain: A Compact and Agent-Native MoE Training System

Ruihang Lai, Hao Kang, Haozhan Tang, Akaash R. Parthasarathy, Zichun Yu, Junru Shao, Todd C. Mowry, Chenyan Xiong, Tianqi Chen · 2026

Read PithTrain: A Compact and Agent-Native MoE Training System by Ruihang Lai, Hao Kang, Haozhan Tang, Akaash R. Parthasarathy, Zichun Yu, Junru Shao, Todd C. Mowry, Chenyan Xiong, Tianqi Chen (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-31463.txt.

Description

Lai and colleagues build PithTrain, a compact training framework for mixture-of-experts models that is made for AI coding agents to grasp, run, and extend. Production frameworks took years to build, yet changing them for new designs stays costly, and speed tests hide the effort agents spend. PithTrain matches production throughput, and on the new ATE-Bench it cuts agent turns by up to 62 percent and active GPU time by up to 64 percent.

Catalog record

CREATORRuihang Lai, Hao Kang, Haozhan Tang, Akaash R. Parthasarathy, Zichun Yu, Junru Shao, Todd C. Mowry, Chenyan Xiong, Tianqi Chen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60187
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.31463

Browse the full library · About the Tower of Babel