RW-TTT: Batched Serving for Request-Owned Test-Time Training State

Jian Yang; Zhizhuo Kou; Chen Han; Hao Zhang; Yao Tian; Sirui Han; Yike Guo

RW-TTT: Batched Serving for Request-Owned Test-Time Training State

Jian Yang, Zhizhuo Kou, Chen Han, Hao Zhang, Yao Tian, Sirui Han, Yike Guo

Correct Metadata for

Use this form to create a GitHub issue with structured data describing the correction. You will need a GitHub account. Once you create that issue, the correction will be reviewed by a staff member.

⚠️ Mobile Users: Submitting this form to create a new issue will only work with github.com, not the GitHub Mobile app.

Important: The Anthology treat PDFs as authoritative. Please use this form only to correct data that is out of line with the PDF. See our corrections guidelines if you need to change the PDF.

Title Adjust the title. Retain tags such as <fixed-case>.

Authors Adjust author names and order to match the PDF.

Abstract Correct abstract if needed. Retain XML formatting tags such as <tex-math>. You may use ... for bold, ... for italic, ... for underline, <sc>...</sc> for small-caps, <tt>...<tt> for typewriter text, <url>...</url> for URLs, <a href=...> for hyperlinks, and <par/> for paragraph breaks.

Verification against PDF Ensure that the new title/authors match the snapshot below. (If there is no snapshot or it is too small, consult the PDF.)

Authors concatenated from the text boxes above:

ALL author names match the snapshot above—including middle initials, hyphens, and accents.

Abstract

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 In-Place-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31× over sequential serving and 3.44× over memory-matched independent replicas. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.

Anthology ID:: 2026.emnlp-main.232
Volume:: Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
Month:: October
Year:: 2026
Address:: Budapest, Hungary
Editors:: Hung-Yi Lee, Sunipa Dev, Jessy Li, Desmond Elliott
Venue:: EMNLP
SIG:: SIGDAT
Publisher:: Association for Computational Linguistics
Note:
Pages:: 5328–5338
Language:
URL:: https://aclanthology.org/2026.emnlp-main.232/
DOI:
Bibkey:
Cite (ACL):: Jian Yang, Zhizhuo Kou, Chen Han, Hao Zhang, Yao Tian, Sirui Han, and Yike Guo. 2026. RW-TTT: Batched Serving for Request-Owned Test-Time Training State. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 5328–5338, Budapest, Hungary. Association for Computational Linguistics.
Cite (Informal):: RW-TTT: Batched Serving for Request-Owned Test-Time Training State (Yang et al., EMNLP 2026)
Copy Citation:
PDF:: https://aclanthology.org/2026.emnlp-main.232.pdf

PDF Cite Search Fix data

Export citation

BibTeX
MODS XML
Endnote
Preformatted

@inproceedings{yang-etal-2026-rw,
    title = "{RW}-{TTT}: Batched Serving for Request-Owned Test-Time Training State",
    author = "Yang, Jian  and
      Kou, Zhizhuo  and
      Han, Chen  and
      Zhang, Hao  and
      Tian, Yao  and
      Han, Sirui  and
      Guo, Yike",
    editor = "Lee, Hung-Yi  and
      Dev, Sunipa  and
      Li, Jessy  and
      Elliott, Desmond",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.232/",
    pages = "5328--5338",
    ISBN = "979-8-89176-459-0",
    abstract = "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 \textit{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 In-Place-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31$\times$ over sequential serving and 3.44$\times$ over memory-matched independent replicas. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks."
}

Download as File

<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="yang-etal-2026-rw">
    <titleInfo>
        <title>RW-TTT: Batched Serving for Request-Owned Test-Time Training State</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Jian</namePart>
        <namePart type="family">Yang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Zhizhuo</namePart>
        <namePart type="family">Kou</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Chen</namePart>
        <namePart type="family">Han</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Hao</namePart>
        <namePart type="family">Zhang</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yao</namePart>
        <namePart type="family">Tian</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Sirui</namePart>
        <namePart type="family">Han</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yike</namePart>
        <namePart type="family">Guo</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-10</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Hung-Yi</namePart>
            <namePart type="family">Lee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sunipa</namePart>
            <namePart type="family">Dev</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jessy</namePart>
            <namePart type="family">Li</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Desmond</namePart>
            <namePart type="family">Elliott</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Budapest, Hungary</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>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 In-Place-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31\times over sequential serving and 3.44\times over memory-matched independent replicas. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.</abstract>
    <identifier type="citekey">yang-etal-2026-rw</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.232/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>5328</start>
            <end>5338</end>
        </extent>
    </part>
</mods>
</modsCollection>

Download as File

%0 Conference Proceedings
%T RW-TTT: Batched Serving for Request-Owned Test-Time Training State
%A Yang, Jian
%A Kou, Zhizhuo
%A Han, Chen
%A Zhang, Hao
%A Tian, Yao
%A Han, Sirui
%A Guo, Yike
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F yang-etal-2026-rw
%X 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 In-Place-TTT streams, RW-TTT reaches 274.61 aggregate tok/s, 9.31\times over sequential serving and 3.44\times over memory-matched independent replicas. It preserves behavior on RULER, a long-context benchmark, and passes owner/version checks.
%U https://aclanthology.org/2026.emnlp-main.232/
%P 5328-5338

Download as File

Markdown (Informal)

[RW-TTT: Batched Serving for Request-Owned Test-Time Training State](https://aclanthology.org/2026.emnlp-main.232/) (Yang et al., EMNLP 2026)

RW-TTT: Batched Serving for Request-Owned Test-Time Training State (Yang et al., EMNLP 2026)

ACL

Jian Yang, Zhizhuo Kou, Chen Han, Hao Zhang, Yao Tian, Sirui Han, and Yike Guo. 2026. RW-TTT: Batched Serving for Request-Owned Test-Time Training State. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 5328–5338, Budapest, Hungary. Association for Computational Linguistics.