<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gurung-etal-2026-lightweight">
    <titleInfo>
        <title>Lightweight Latent Reasoning for Narrative Tasks</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Alexander</namePart>
        <namePart type="family">Gurung</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Esmeralda</namePart>
        <namePart type="given">S</namePart>
        <namePart type="family">Whitammer</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Mirella</namePart>
        <namePart type="family">Lapata</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <genre authority="bibutilsgt">journal article</genre>
    <relatedItem type="host">
        <titleInfo>
            <title>Transactions of the Association for Computational Linguistics</title>
        </titleInfo>
        <originInfo>
            <issuance>continuing</issuance>
            <publisher>MIT Press</publisher>
            <place>
                <placeTerm type="text">Cambridge, MA</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">periodical</genre>
        <genre authority="bibutilsgt">academic journal</genre>
    </relatedItem>
    <abstract>Large language models (LLMs) tackle complex tasks by generating long chains of thought or “reasoning traces” that act as latent variables in the generation of an output given a query. A model’s ability to generate such traces can be optimized with reinforcement learning (RL) to improve their utility in predicting an answer. This optimization comes at a high computational cost, especially for narrative-related tasks that involve retrieving and processing many tokens. To this end, we propose LiteReason, a latent reasoning method that can be interleaved with standard token sampling and easily combined with RL techniques. LiteReason employs a lightweight Reasoning Projector module, trained to produce continuous latent tokens that help the model ‘skip’ reasoning steps. During RL, the policy model decides when to activate the projector, switching between latent and discrete reasoning as needed. Experimental results on plot hole detection and book chapter generation show that our method outperforms latent reasoning baselines and comes close to matching non-latent RL training, while reducing final reasoning length by 77–92%. Overall, LiteReason guides RL training to a more efficient part of the performance-computation tradeoff curve.1</abstract>
    <identifier type="citekey">gurung-etal-2026-lightweight</identifier>
    <identifier type="doi">10.1162/tacl.a.796</identifier>
    <location>
        <url>https://aclanthology.org/2026.tacl-1.98/</url>
    </location>
    <part>
        <date>2026</date>
        <detail type="volume"><number>14</number></detail>
        <extent unit="page">
            <start>2163</start>
            <end>2186</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Journal Article
%T Lightweight Latent Reasoning for Narrative Tasks
%A Gurung, Alexander
%A Whitammer, Esmeralda S.
%A Lapata, Mirella
%J Transactions of the Association for Computational Linguistics
%D 2026
%V 14
%I MIT Press
%C Cambridge, MA
%F gurung-etal-2026-lightweight
%X Large language models (LLMs) tackle complex tasks by generating long chains of thought or “reasoning traces” that act as latent variables in the generation of an output given a query. A model’s ability to generate such traces can be optimized with reinforcement learning (RL) to improve their utility in predicting an answer. This optimization comes at a high computational cost, especially for narrative-related tasks that involve retrieving and processing many tokens. To this end, we propose LiteReason, a latent reasoning method that can be interleaved with standard token sampling and easily combined with RL techniques. LiteReason employs a lightweight Reasoning Projector module, trained to produce continuous latent tokens that help the model ‘skip’ reasoning steps. During RL, the policy model decides when to activate the projector, switching between latent and discrete reasoning as needed. Experimental results on plot hole detection and book chapter generation show that our method outperforms latent reasoning baselines and comes close to matching non-latent RL training, while reducing final reasoning length by 77–92%. Overall, LiteReason guides RL training to a more efficient part of the performance-computation tradeoff curve.1
%R 10.1162/tacl.a.796
%U https://aclanthology.org/2026.tacl-1.98/
%U https://doi.org/10.1162/tacl.a.796
%P 2163-2186

Markdown (Informal)

[Lightweight Latent Reasoning for Narrative Tasks](https://aclanthology.org/2026.tacl-1.98/) (Gurung et al., TACL 2026)

• Lightweight Latent Reasoning for Narrative Tasks (Gurung et al., TACL 2026)

ACL

• Alexander Gurung, Esmeralda S. Whitammer, and Mirella Lapata. 2026. Lightweight Latent Reasoning for Narrative Tasks. Transactions of the Association for Computational Linguistics, 14:2163–2186.