@article{hosseini-etal-2026-cg,
title = "{CG}-{TTRL}: Context-Guided Test-Time Reinforcement Learning for On-Device Large Language Models",
author = "Hosseini, Peyman and
Bohdal, Ondrej and
Ceritli, Taha and
Castro, Ignacio and
Purver, Matthew and
Ozay, Mete and
Michieli, Umberto",
journal = "Transactions of the Association for Computational Linguistics",
volume = "14",
year = "2026",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2026.tacl-1.85/",
doi = "10.1162/tacl.a.783",
pages = "1883--1898",
abstract = "Test-Time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, In-Context Learning has been widely explored at inference time to enhance model performance without weight updates. However, TTRL{'}s two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose Context-Guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7{\%} relative accuracy improvement over TTRL), while boosting efficiency by obtaining strong performance after only a few steps of Test-Time Training (e.g. 8{\%} relative improvement rather than 1{\%} over TTRL after 3 steps)."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hosseini-etal-2026-cg">
<titleInfo>
<title>CG-TTRL: Context-Guided Test-Time Reinforcement Learning for On-Device Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Peyman</namePart>
<namePart type="family">Hosseini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ondrej</namePart>
<namePart type="family">Bohdal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Taha</namePart>
<namePart type="family">Ceritli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ignacio</namePart>
<namePart type="family">Castro</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Matthew</namePart>
<namePart type="family">Purver</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mete</namePart>
<namePart type="family">Ozay</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Umberto</namePart>
<namePart type="family">Michieli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<genre authority="bibutilsgt">journal article</genre>
<relatedItem type="host">
<titleInfo>
<title>Transactions of the Association for Computational Linguistics</title>
</titleInfo>
<originInfo>
<issuance>continuing</issuance>
<publisher>MIT Press</publisher>
<place>
<placeTerm type="text">Cambridge, MA</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">periodical</genre>
<genre authority="bibutilsgt">academic journal</genre>
</relatedItem>
<abstract>Test-Time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, In-Context Learning has been widely explored at inference time to enhance model performance without weight updates. However, TTRL’s two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose Context-Guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7% relative accuracy improvement over TTRL), while boosting efficiency by obtaining strong performance after only a few steps of Test-Time Training (e.g. 8% relative improvement rather than 1% over TTRL after 3 steps).</abstract>
<identifier type="citekey">hosseini-etal-2026-cg</identifier>
<identifier type="doi">10.1162/tacl.a.783</identifier>
<location>
<url>https://aclanthology.org/2026.tacl-1.85/</url>
</location>
<part>
<date>2026</date>
<detail type="volume"><number>14</number></detail>
<extent unit="page">
<start>1883</start>
<end>1898</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Journal Article
%T CG-TTRL: Context-Guided Test-Time Reinforcement Learning for On-Device Large Language Models
%A Hosseini, Peyman
%A Bohdal, Ondrej
%A Ceritli, Taha
%A Castro, Ignacio
%A Purver, Matthew
%A Ozay, Mete
%A Michieli, Umberto
%J Transactions of the Association for Computational Linguistics
%D 2026
%V 14
%I MIT Press
%C Cambridge, MA
%F hosseini-etal-2026-cg
%X Test-Time Reinforcement Learning (TTRL) has shown promise in adapting foundation models for complex tasks at test time, resulting in large performance improvements. TTRL leverages an elegant two-phase sampling strategy: first, multi-sampling derives a pseudo-label via majority voting, while subsequent downsampling and reward-based fine-tuning encourages the model to explore and learn diverse valid solutions, with the pseudo-label modulating the reward signal. Meanwhile, In-Context Learning has been widely explored at inference time to enhance model performance without weight updates. However, TTRL’s two-phase sampling strategy under-utilizes contextual guidance, which can potentially improve pseudo-label accuracy in the initial exploitation phase while regulating exploration in the second. To address this, we propose Context-Guided TTRL (CG-TTRL), integrating context dynamically into both sampling phases and propose a method for efficient context selection for on-device applications. Our evaluations on mathematical and scientific QA benchmarks show CG-TTRL outperforms TTRL (e.g. additional 7% relative accuracy improvement over TTRL), while boosting efficiency by obtaining strong performance after only a few steps of Test-Time Training (e.g. 8% relative improvement rather than 1% over TTRL after 3 steps).
%R 10.1162/tacl.a.783
%U https://aclanthology.org/2026.tacl-1.85/
%U https://doi.org/10.1162/tacl.a.783
%P 1883-1898
Markdown (Informal)
[CG-TTRL: Context-Guided Test-Time Reinforcement Learning for On-Device Large Language Models](https://aclanthology.org/2026.tacl-1.85/) (Hosseini et al., TACL 2026)
ACL