<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="eliasz-etal-2026-reasoning">
    <titleInfo>
        <title>Reasoning Beyond English: Separating Capability and Language in Cross-Lingual Reasoning Model Adaptation</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Jan</namePart>
        <namePart type="family">Eliasz</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Mikołaj</namePart>
        <namePart type="family">Langner</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Dzmitry</namePart>
        <namePart type="family">Pihulski</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Konrad</namePart>
        <namePart type="family">Wojtasik</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Przemyslaw</namePart>
        <namePart type="family">Kazienko</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jan</namePart>
        <namePart type="family">Kocon</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Teddy</namePart>
        <namePart type="family">Ferdinan</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
    </relatedItem>
    <abstract>Many top-performing reasoning language models (RLMs) exhibit English-centric reasoning behavior, generating English chains of thought even when the prompt and final answer are in another language. Beyond its impact on multilingual reasoning performance, this also blocks downstream uses of the reasoning trace, such as debugging model errors and auditing decisions. We studied different approaches to RLM language adaptation by adapting Qwen3-4B-Thinking-2507 into Polish. We find that the accuracy collapse of SFT decomposes additively into termination and quality failure modes. On the other hand, KL-anchored fine-tuning mitigates these failures but exhibits a sharp plateau-then-cliff structure under reverse divergence: language match collapses from 0.89 to 0.37 as α doubles, with no comparable structure under forward divergence. However, a second training stage leveraging GRPO helps close the remaining gap and recover the language match even from configurations past the cliff, revealing that the two stages divide labor: the supervised stage controls reasoning quality while the RL stage closes residual language match gaps. These findings suggest that, in a two-stage RLM post-training, the supervised phase can be considerably less constrained than current practice assumes.</abstract>
    <identifier type="citekey">eliasz-etal-2026-reasoning</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2481/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>56324</start>
            <end>56344</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Reasoning Beyond English: Separating Capability and Language in Cross-Lingual Reasoning Model Adaptation
%A Eliasz, Jan
%A Langner, Mikołaj
%A Pihulski, Dzmitry
%A Wojtasik, Konrad
%A Kazienko, Przemyslaw
%A Kocon, Jan
%A Ferdinan, Teddy
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F eliasz-etal-2026-reasoning
%X Many top-performing reasoning language models (RLMs) exhibit English-centric reasoning behavior, generating English chains of thought even when the prompt and final answer are in another language. Beyond its impact on multilingual reasoning performance, this also blocks downstream uses of the reasoning trace, such as debugging model errors and auditing decisions. We studied different approaches to RLM language adaptation by adapting Qwen3-4B-Thinking-2507 into Polish. We find that the accuracy collapse of SFT decomposes additively into termination and quality failure modes. On the other hand, KL-anchored fine-tuning mitigates these failures but exhibits a sharp plateau-then-cliff structure under reverse divergence: language match collapses from 0.89 to 0.37 as α doubles, with no comparable structure under forward divergence. However, a second training stage leveraging GRPO helps close the remaining gap and recover the language match even from configurations past the cliff, revealing that the two stages divide labor: the supervised stage controls reasoning quality while the RL stage closes residual language match gaps. These findings suggest that, in a two-stage RLM post-training, the supervised phase can be considerably less constrained than current practice assumes.
%U https://aclanthology.org/2026.emnlp-main.2481/
%P 56324-56344

Markdown (Informal)

[Reasoning Beyond English: Separating Capability and Language in Cross-Lingual Reasoning Model Adaptation](https://aclanthology.org/2026.emnlp-main.2481/) (Eliasz et al., EMNLP 2026)

• Reasoning Beyond English: Separating Capability and Language in Cross-Lingual Reasoning Model Adaptation (Eliasz et al., EMNLP 2026)

ACL

• Jan Eliasz, Mikołaj Langner, Dzmitry Pihulski, Konrad Wojtasik, Przemyslaw Kazienko, Jan Kocon, and Teddy Ferdinan. 2026. Reasoning Beyond English: Separating Capability and Language in Cross-Lingual Reasoning Model Adaptation. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 56324–56344, Budapest, Hungary. Association for Computational Linguistics.