@inproceedings{rakotonirina-etal-2026-correct,
title = "Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning",
author = {Rakotonirina, Nathana{\"e}l Carraz and
Pang, Ren and
Anna John, Neha and
Bohlke-Schneider, Michael and
Hardalov, Momchil},
editor = "Liakata, Maria and
Moreira, Viviane P. and
Zhang, Jiajun and
Jurgens, David",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {ACL} 2026",
month = jul,
year = "2026",
address = "San Diego, California, United States",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-acl.622/",
pages = "12798--12810",
ISBN = "979-8-89176-395-1",
abstract = "The reasoning capabilities of large language models (LLMs) have improved substantially through increased test-time computation, typically in the form of intermediate tokens known as chain-of-thought (CoT). However, CoT often becomes unnecessarily long, increasing computation costs without improving accuracy and sometimes even degrading performance, a phenomenon known as ``\textit{overthinking}''. We propose a multi-stage efficient reasoning method that combines supervised fine-tuning{---}via rejection sampling or reasoning trace reformatting{---}with reinforcement learning using an adaptive length penalty. We introduce a lightweight reward function that penalizes tokens generated after the first correct answer, encouraging the model to perform self-verification only when beneficial. We conduct a holistic evaluation across seven diverse reasoning tasks, analyzing the accuracy{--}response length trade-off. Our approach reduces response length by an average of 28{\%} for 8B models and 40{\%} for 32B models, while incurring only minor performance drops of 1.6 and 2.5 points, respectively. Despite its conceptual simplicity, it achieves a better trade-off than more complex state-of-the-art efficient reasoning methods, scoring 76.6 on the area under the Overthinking-Adjusted Accuracy curve ($\text{AUC}_{\text{OAA}}$){---}5 points above the base model and 2.5 points above the second-best approach."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rakotonirina-etal-2026-correct">
<titleInfo>
<title>Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nathanaël</namePart>
<namePart type="given">Carraz</namePart>
<namePart type="family">Rakotonirina</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ren</namePart>
<namePart type="family">Pang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Neha</namePart>
<namePart type="family">Anna John</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michael</namePart>
<namePart type="family">Bohlke-Schneider</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Momchil</namePart>
<namePart type="family">Hardalov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-07</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: ACL 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="family">Liakata</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Viviane</namePart>
<namePart type="given">P</namePart>
<namePart type="family">Moreira</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiajun</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Jurgens</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">San Diego, California, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-395-1</identifier>
</relatedItem>
<abstract>The reasoning capabilities of large language models (LLMs) have improved substantially through increased test-time computation, typically in the form of intermediate tokens known as chain-of-thought (CoT). However, CoT often becomes unnecessarily long, increasing computation costs without improving accuracy and sometimes even degrading performance, a phenomenon known as “overthinking”. We propose a multi-stage efficient reasoning method that combines supervised fine-tuning—via rejection sampling or reasoning trace reformatting—with reinforcement learning using an adaptive length penalty. We introduce a lightweight reward function that penalizes tokens generated after the first correct answer, encouraging the model to perform self-verification only when beneficial. We conduct a holistic evaluation across seven diverse reasoning tasks, analyzing the accuracy–response length trade-off. Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models, while incurring only minor performance drops of 1.6 and 2.5 points, respectively. Despite its conceptual simplicity, it achieves a better trade-off than more complex state-of-the-art efficient reasoning methods, scoring 76.6 on the area under the Overthinking-Adjusted Accuracy curve (\textAUC_\textOAA)—5 points above the base model and 2.5 points above the second-best approach.</abstract>
<identifier type="citekey">rakotonirina-etal-2026-correct</identifier>
<location>
<url>https://aclanthology.org/2026.findings-acl.622/</url>
</location>
<part>
<date>2026-07</date>
<extent unit="page">
<start>12798</start>
<end>12810</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning
%A Rakotonirina, Nathanaël Carraz
%A Pang, Ren
%A Anna John, Neha
%A Bohlke-Schneider, Michael
%A Hardalov, Momchil
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Findings of the Association for Computational Linguistics: ACL 2026
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-395-1
%F rakotonirina-etal-2026-correct
%X The reasoning capabilities of large language models (LLMs) have improved substantially through increased test-time computation, typically in the form of intermediate tokens known as chain-of-thought (CoT). However, CoT often becomes unnecessarily long, increasing computation costs without improving accuracy and sometimes even degrading performance, a phenomenon known as “overthinking”. We propose a multi-stage efficient reasoning method that combines supervised fine-tuning—via rejection sampling or reasoning trace reformatting—with reinforcement learning using an adaptive length penalty. We introduce a lightweight reward function that penalizes tokens generated after the first correct answer, encouraging the model to perform self-verification only when beneficial. We conduct a holistic evaluation across seven diverse reasoning tasks, analyzing the accuracy–response length trade-off. Our approach reduces response length by an average of 28% for 8B models and 40% for 32B models, while incurring only minor performance drops of 1.6 and 2.5 points, respectively. Despite its conceptual simplicity, it achieves a better trade-off than more complex state-of-the-art efficient reasoning methods, scoring 76.6 on the area under the Overthinking-Adjusted Accuracy curve (\textAUC_\textOAA)—5 points above the base model and 2.5 points above the second-best approach.
%U https://aclanthology.org/2026.findings-acl.622/
%P 12798-12810
Markdown (Informal)
[Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning](https://aclanthology.org/2026.findings-acl.622/) (Rakotonirina et al., Findings 2026)
ACL