<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bhattacharyya-wein-2026-reason">
    <titleInfo>
        <title>For What Reason? Interpreting Models’ Encoding of Causation and Antithesis</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Abhidip</namePart>
        <namePart type="family">Bhattacharyya</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shira</namePart>
        <namePart type="family">Wein</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-08</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Jinho</namePart>
            <namePart type="given">D</namePart>
            <namePart type="family">Choi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Yun-Nung</namePart>
            <namePart type="family">Chen</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Kotaro</namePart>
            <namePart type="family">Funakoshi</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ali</namePart>
            <namePart type="family">Emami</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Atlanta, Georgia, USA</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
    </relatedItem>
    <abstract>Discourse relations provide document structure, critical to language understanding and enabling language model performance and ethicality. In this work, we investigate how instruction-tuned Transformer models (LLaMA and Mistral) encode discourse relations, with a focus on the somewhat opposed relations on causation and antithesis. Framing the task as a next-token prediction task and applying a suite of interpretability techniques to test model internals, our findings show that certain early layers make predictive decisions at mid-sequence tokens, while some mid-level layers finalize their decisions closer to the last token. Most of the remaining layers primarily propagate earlier decisions rather than actively influencing them. Additionally, we observe that some layers exhibit a preference for one answer over alternatives, suggesting asymmetric representation of discourse-based reasoning.</abstract>
    <identifier type="citekey">bhattacharyya-wein-2026-reason</identifier>
    <location>
        <url>https://aclanthology.org/2026.sigdial-1.17/</url>
    </location>
    <part>
        <date>2026-08</date>
        <extent unit="page">
            <start>237</start>
            <end>253</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T For What Reason? Interpreting Models’ Encoding of Causation and Antithesis
%A Bhattacharyya, Abhidip
%A Wein, Shira
%Y Choi, Jinho D.
%Y Chen, Yun-Nung
%Y Funakoshi, Kotaro
%Y Emami, Ali
%S Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue
%D 2026
%8 August
%I Association for Computational Linguistics
%C Atlanta, Georgia, USA
%F bhattacharyya-wein-2026-reason
%X Discourse relations provide document structure, critical to language understanding and enabling language model performance and ethicality. In this work, we investigate how instruction-tuned Transformer models (LLaMA and Mistral) encode discourse relations, with a focus on the somewhat opposed relations on causation and antithesis. Framing the task as a next-token prediction task and applying a suite of interpretability techniques to test model internals, our findings show that certain early layers make predictive decisions at mid-sequence tokens, while some mid-level layers finalize their decisions closer to the last token. Most of the remaining layers primarily propagate earlier decisions rather than actively influencing them. Additionally, we observe that some layers exhibit a preference for one answer over alternatives, suggesting asymmetric representation of discourse-based reasoning.
%U https://aclanthology.org/2026.sigdial-1.17/
%P 237-253

Markdown (Informal)

[For What Reason? Interpreting Models’ Encoding of Causation and Antithesis](https://aclanthology.org/2026.sigdial-1.17/) (Bhattacharyya & Wein, SIGDIAL 2026)

• For What Reason? Interpreting Models’ Encoding of Causation and Antithesis (Bhattacharyya & Wein, SIGDIAL 2026)

ACL

• Abhidip Bhattacharyya and Shira Wein. 2026. For What Reason? Interpreting Models’ Encoding of Causation and Antithesis. In Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 237–253, Atlanta, Georgia, USA. Association for Computational Linguistics.