@inproceedings{datta-etal-2026-llms,
title = "{LLM}s Decide Early and Explain Later: You Can Stop Them Early",
author = "Datta, Ayan and
Zhao, Zhixue and
Marreddy, Mounika and
Verma, Bhuvanesh and
Mamidi, Radhika and
Mehler, Alexander",
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.2412/",
pages = "51796--51822",
ISBN = "979-8-89176-462-0",
abstract = "Large Language Models often achieve strong performance by generating long intermediate chain-of-thought (CoT) reasoning. However, it remains unclear when a model{'}s final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using *forced answer completion*, which elicits a model{'}s intermediate prediction at partial reasoning prefixes. Across a range of models and datasets, we find that the final answer is fixed from the earliest stages of generation in most cases, with **only 32{\%} of answers changing after CoT**. Moreover, once the final answer switch does occur, models continue to generate substantial reasoning, producing **760 additional tokens per query** on average and accounting for around **40{\%} of the total reasoning**. Motivated by these findings, we explore early-stopping strategies that halt generation once the answer has stabilized. We show that simple probe-based stopping can reduce reasoning token usage by **500 tokens per query** while incurring only a **2{\%} drop in accuracy**. Together, our results indicate that a large portion of chain-of-thought generation functions as **post hoc explanation** and can be reduced with minimal impact on performance."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="datta-etal-2026-llms">
<titleInfo>
<title>LLMs Decide Early and Explain Later: You Can Stop Them Early</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ayan</namePart>
<namePart type="family">Datta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhixue</namePart>
<namePart type="family">Zhao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mounika</namePart>
<namePart type="family">Marreddy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhuvanesh</namePart>
<namePart type="family">Verma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Radhika</namePart>
<namePart type="family">Mamidi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alexander</namePart>
<namePart type="family">Mehler</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>Large Language Models often achieve strong performance by generating long intermediate chain-of-thought (CoT) reasoning. However, it remains unclear when a model’s final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using *forced answer completion*, which elicits a model’s intermediate prediction at partial reasoning prefixes. Across a range of models and datasets, we find that the final answer is fixed from the earliest stages of generation in most cases, with **only 32% of answers changing after CoT**. Moreover, once the final answer switch does occur, models continue to generate substantial reasoning, producing **760 additional tokens per query** on average and accounting for around **40% of the total reasoning**. Motivated by these findings, we explore early-stopping strategies that halt generation once the answer has stabilized. We show that simple probe-based stopping can reduce reasoning token usage by **500 tokens per query** while incurring only a **2% drop in accuracy**. Together, our results indicate that a large portion of chain-of-thought generation functions as **post hoc explanation** and can be reduced with minimal impact on performance.</abstract>
<identifier type="citekey">datta-etal-2026-llms</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2412/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51796</start>
<end>51822</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T LLMs Decide Early and Explain Later: You Can Stop Them Early
%A Datta, Ayan
%A Zhao, Zhixue
%A Marreddy, Mounika
%A Verma, Bhuvanesh
%A Mamidi, Radhika
%A Mehler, Alexander
%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 datta-etal-2026-llms
%X Large Language Models often achieve strong performance by generating long intermediate chain-of-thought (CoT) reasoning. However, it remains unclear when a model’s final answer is actually determined during generation. If the answer is already fixed at an intermediate stage, subsequent reasoning tokens may constitute post-decision explanation, increasing inference cost and latency without improving correctness. We study the evolution of predicted answers over reasoning steps using *forced answer completion*, which elicits a model’s intermediate prediction at partial reasoning prefixes. Across a range of models and datasets, we find that the final answer is fixed from the earliest stages of generation in most cases, with **only 32% of answers changing after CoT**. Moreover, once the final answer switch does occur, models continue to generate substantial reasoning, producing **760 additional tokens per query** on average and accounting for around **40% of the total reasoning**. Motivated by these findings, we explore early-stopping strategies that halt generation once the answer has stabilized. We show that simple probe-based stopping can reduce reasoning token usage by **500 tokens per query** while incurring only a **2% drop in accuracy**. Together, our results indicate that a large portion of chain-of-thought generation functions as **post hoc explanation** and can be reduced with minimal impact on performance.
%U https://aclanthology.org/2026.findings-emnlp.2412/
%P 51796-51822
Markdown (Informal)
[LLMs Decide Early and Explain Later: You Can Stop Them Early](https://aclanthology.org/2026.findings-emnlp.2412/) (Datta et al., Findings 2026)
ACL
- Ayan Datta, Zhixue Zhao, Mounika Marreddy, Bhuvanesh Verma, Radhika Mamidi, and Alexander Mehler. 2026. LLMs Decide Early and Explain Later: You Can Stop Them Early. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 51796–51822, Budapest, Hungary. Association for Computational Linguistics.