@inproceedings{shelat-etal-2026-chain,
title = "Chain-of-Thought Hurts When the Answer Speaks for Itself",
author = "Shelat, Shlok and
Salvi, Shrey Vimalbhai and
Gaur, Manas and
Roy, Souvik",
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.1277/",
doi = "10.18653/2026.findings-emnlp.1277",
pages = "27047--27071",
ISBN = "979-8-89176-462-0",
abstract = "Chain-of-thought (CoT) supervision has become a default assumption in fine-tuning large language models for reasoning tasks, with the intuition being that exposing models to intermediate reasoning steps leads to deeper learning than training on answers alone. We challenge this assumption. We introduce answer backtrackability, a measurable quantity capturing the degree to which a question can be reconstructed from its answer, and argue it is a key determinant of whether CoT supervision helps or hurts during fine-tuning. To test this, we construct two novel benchmarks derived from automata theory (NFA and DFA), where answers are transition tables from which the question is largely recoverable, and fine-tune three LLMs using LoRA on CoT and No-CoT versions of five datasets: GSM8K, MathInstruct, MATH, and our two formal-language benchmarks. On mathematical datasets, CoT consistently outperforms No-CoT. Within MathInstruct, Program-of-Thought (PoT) problems, where executable code answers are high-backtrackability, favour No-CoT, while CoT-format problems favour CoT. On our benchmarks, the pattern reverses: No-CoT outperforms CoT across all models and difficulty tiers. When the answer encodes sufficient structure to reconstruct the reasoning process, intermediate supervision adds noise rather than signal."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shelat-etal-2026-chain">
<titleInfo>
<title>Chain-of-Thought Hurts When the Answer Speaks for Itself</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shlok</namePart>
<namePart type="family">Shelat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shrey</namePart>
<namePart type="given">Vimalbhai</namePart>
<namePart type="family">Salvi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Manas</namePart>
<namePart type="family">Gaur</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Souvik</namePart>
<namePart type="family">Roy</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>Chain-of-thought (CoT) supervision has become a default assumption in fine-tuning large language models for reasoning tasks, with the intuition being that exposing models to intermediate reasoning steps leads to deeper learning than training on answers alone. We challenge this assumption. We introduce answer backtrackability, a measurable quantity capturing the degree to which a question can be reconstructed from its answer, and argue it is a key determinant of whether CoT supervision helps or hurts during fine-tuning. To test this, we construct two novel benchmarks derived from automata theory (NFA and DFA), where answers are transition tables from which the question is largely recoverable, and fine-tune three LLMs using LoRA on CoT and No-CoT versions of five datasets: GSM8K, MathInstruct, MATH, and our two formal-language benchmarks. On mathematical datasets, CoT consistently outperforms No-CoT. Within MathInstruct, Program-of-Thought (PoT) problems, where executable code answers are high-backtrackability, favour No-CoT, while CoT-format problems favour CoT. On our benchmarks, the pattern reverses: No-CoT outperforms CoT across all models and difficulty tiers. When the answer encodes sufficient structure to reconstruct the reasoning process, intermediate supervision adds noise rather than signal.</abstract>
<identifier type="citekey">shelat-etal-2026-chain</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1277</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1277/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>27047</start>
<end>27071</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Chain-of-Thought Hurts When the Answer Speaks for Itself
%A Shelat, Shlok
%A Salvi, Shrey Vimalbhai
%A Gaur, Manas
%A Roy, Souvik
%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 shelat-etal-2026-chain
%X Chain-of-thought (CoT) supervision has become a default assumption in fine-tuning large language models for reasoning tasks, with the intuition being that exposing models to intermediate reasoning steps leads to deeper learning than training on answers alone. We challenge this assumption. We introduce answer backtrackability, a measurable quantity capturing the degree to which a question can be reconstructed from its answer, and argue it is a key determinant of whether CoT supervision helps or hurts during fine-tuning. To test this, we construct two novel benchmarks derived from automata theory (NFA and DFA), where answers are transition tables from which the question is largely recoverable, and fine-tune three LLMs using LoRA on CoT and No-CoT versions of five datasets: GSM8K, MathInstruct, MATH, and our two formal-language benchmarks. On mathematical datasets, CoT consistently outperforms No-CoT. Within MathInstruct, Program-of-Thought (PoT) problems, where executable code answers are high-backtrackability, favour No-CoT, while CoT-format problems favour CoT. On our benchmarks, the pattern reverses: No-CoT outperforms CoT across all models and difficulty tiers. When the answer encodes sufficient structure to reconstruct the reasoning process, intermediate supervision adds noise rather than signal.
%R 10.18653/2026.findings-emnlp.1277
%U https://aclanthology.org/2026.findings-emnlp.1277/
%U https://doi.org/10.18653/2026.findings-emnlp.1277
%P 27047-27071
Markdown (Informal)
[Chain-of-Thought Hurts When the Answer Speaks for Itself](https://aclanthology.org/2026.findings-emnlp.1277/) (Shelat et al., Findings 2026)
ACL
- Shlok Shelat, Shrey Vimalbhai Salvi, Manas Gaur, and Souvik Roy. 2026. Chain-of-Thought Hurts When the Answer Speaks for Itself. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 27047–27071, Budapest, Hungary. Association for Computational Linguistics.