Saved in:
Bibliographic Details
Main Authors: Yang, Jian, Kou, Zhizhuo, Tian, Yao, Zhang, Hao, Chen, Han, Han, Sirui, Guo, Yike
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.28053
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911723668111360
author Yang, Jian
Kou, Zhizhuo
Tian, Yao
Zhang, Hao
Chen, Han
Han, Sirui
Guo, Yike
author_facet Yang, Jian
Kou, Zhizhuo
Tian, Yao
Zhang, Hao
Chen, Han
Han, Sirui
Guo, Yike
contents Test-time training (TTT) adapts an LLM during generation by reading and updating request-owned state, such as fast weights, low-rank deltas, or streaming learner state. This breaks batched LLM serving, which assumes shared static weights: serial execution is correct but slow, while naive batching can corrupt request state. We formulate this problem as read-write TTT serving and present RW-TTT , which tags each decode step with its owner, version, and READ/WRITE effect, batches only compatible phases, and commits updates only to the owner. On one GPU with eight fast-weight InPlace-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31x over sequential serving and 3.44x over per-stream replicas under the same memory budget. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.
format Preprint
id arxiv_https___arxiv_org_abs_2605_28053
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle RW-TTT: Batched Serving for Request-Owned Test-Time Training State
Yang, Jian
Kou, Zhizhuo
Tian, Yao
Zhang, Hao
Chen, Han
Han, Sirui
Guo, Yike
Machine Learning
Test-time training (TTT) adapts an LLM during generation by reading and updating request-owned state, such as fast weights, low-rank deltas, or streaming learner state. This breaks batched LLM serving, which assumes shared static weights: serial execution is correct but slow, while naive batching can corrupt request state. We formulate this problem as read-write TTT serving and present RW-TTT , which tags each decode step with its owner, version, and READ/WRITE effect, batches only compatible phases, and commits updates only to the owner. On one GPU with eight fast-weight InPlace-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31x over sequential serving and 3.44x over per-stream replicas under the same memory budget. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.
title RW-TTT: Batched Serving for Request-Owned Test-Time Training State
topic Machine Learning
url https://arxiv.org/abs/2605.28053