@inproceedings{wen-etal-2026-parathinker,
title = "{P}ara{T}hinker: Native Parallel Reasoning for Scaling {LLM} Test-Time Compute",
author = "Wen, Hao and
Su, Yifan and
Zhang, Feifei and
Wu, Xinrui and
Sun, Yi and
Liu, Yunxin and
Liu, Yunhao and
Zhang, Ya-Qin and
Li, Yuanchun",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1469/",
pages = "31265--31285",
ISBN = "979-8-89176-462-0",
abstract = "Test-time scaling improves the reasoning capability of large language models by generating longer sequential chains of thought. While effective, this approach reaches a bottleneck as computation increases, resulting in a ``test-time scaling plateau'' where further computation provides only marginal gains. We argue that this ceiling reflects a limitation of the scaling strategy rather than model capability, a phenomenon we term Tunnel Vision, where imperfect early steps can lock the model into a suboptimal reasoning path. Motivated by this observation, we propose ParaThinker, an end-to-end parallel reasoning framework that trains LLMs to generate multiple reasoning paths in parallel and aggregate them into a final answer within a single architecture. By exploring different lines of thought, ParaThinker sidesteps the Tunnel Vision issue and unlocks the model{'}s reasoning potential. On challenging reasoning benchmarks, ParaThinker achieves accuracy improvements (e.g., 6.5{\%}{--}20.7{\%} on AIME-24 with a 1.5B model) over sequential-reasoning LLMs under the same budget of decoding tokens. Under matched decoding token budgets and with KV-cache reuse, ParaThinker achieves 38.7{\%}{--}66.8{\%} lower inference latency than sequential decoding via batched parallel generation. The code is available at https://github.com/MobileLLM/ParaThinker."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wen-etal-2026-parathinker">
<titleInfo>
<title>ParaThinker: Native Parallel Reasoning for Scaling LLM Test-Time Compute</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hao</namePart>
<namePart type="family">Wen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yifan</namePart>
<namePart type="family">Su</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Feifei</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xinrui</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Sun</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunxin</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yunhao</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ya-Qin</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuanchun</namePart>
<namePart type="family">Li</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Test-time scaling improves the reasoning capability of large language models by generating longer sequential chains of thought. While effective, this approach reaches a bottleneck as computation increases, resulting in a “test-time scaling plateau” where further computation provides only marginal gains. We argue that this ceiling reflects a limitation of the scaling strategy rather than model capability, a phenomenon we term Tunnel Vision, where imperfect early steps can lock the model into a suboptimal reasoning path. Motivated by this observation, we propose ParaThinker, an end-to-end parallel reasoning framework that trains LLMs to generate multiple reasoning paths in parallel and aggregate them into a final answer within a single architecture. By exploring different lines of thought, ParaThinker sidesteps the Tunnel Vision issue and unlocks the model’s reasoning potential. On challenging reasoning benchmarks, ParaThinker achieves accuracy improvements (e.g., 6.5%–20.7% on AIME-24 with a 1.5B model) over sequential-reasoning LLMs under the same budget of decoding tokens. Under matched decoding token budgets and with KV-cache reuse, ParaThinker achieves 38.7%–66.8% lower inference latency than sequential decoding via batched parallel generation. The code is available at https://github.com/MobileLLM/ParaThinker.</abstract>
<identifier type="citekey">wen-etal-2026-parathinker</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1469/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>31265</start>
<end>31285</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ParaThinker: Native Parallel Reasoning for Scaling LLM Test-Time Compute
%A Wen, Hao
%A Su, Yifan
%A Zhang, Feifei
%A Wu, Xinrui
%A Sun, Yi
%A Liu, Yunxin
%A Liu, Yunhao
%A Zhang, Ya-Qin
%A Li, Yuanchun
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F wen-etal-2026-parathinker
%X Test-time scaling improves the reasoning capability of large language models by generating longer sequential chains of thought. While effective, this approach reaches a bottleneck as computation increases, resulting in a “test-time scaling plateau” where further computation provides only marginal gains. We argue that this ceiling reflects a limitation of the scaling strategy rather than model capability, a phenomenon we term Tunnel Vision, where imperfect early steps can lock the model into a suboptimal reasoning path. Motivated by this observation, we propose ParaThinker, an end-to-end parallel reasoning framework that trains LLMs to generate multiple reasoning paths in parallel and aggregate them into a final answer within a single architecture. By exploring different lines of thought, ParaThinker sidesteps the Tunnel Vision issue and unlocks the model’s reasoning potential. On challenging reasoning benchmarks, ParaThinker achieves accuracy improvements (e.g., 6.5%–20.7% on AIME-24 with a 1.5B model) over sequential-reasoning LLMs under the same budget of decoding tokens. Under matched decoding token budgets and with KV-cache reuse, ParaThinker achieves 38.7%–66.8% lower inference latency than sequential decoding via batched parallel generation. The code is available at https://github.com/MobileLLM/ParaThinker.
%U https://aclanthology.org/2026.findings-emnlp.1469/
%P 31265-31285
Markdown (Informal)
[ParaThinker: Native Parallel Reasoning for Scaling LLM Test-Time Compute](https://aclanthology.org/2026.findings-emnlp.1469/) (Wen et al., Findings 2026)
ACL
- Hao Wen, Yifan Su, Feifei Zhang, Xinrui Wu, Yi Sun, Yunxin Liu, Yunhao Liu, Ya-Qin Zhang, and Yuanchun Li. 2026. ParaThinker: Native Parallel Reasoning for Scaling LLM Test-Time Compute. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 31265–31285, Budapest, Hungary. Association for Computational Linguistics.