<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bandyopadhyay-etal-2026-language">
    <titleInfo>
        <title>When Should Language Models Search? A Topology-Aware Analysis of Reasoning Strategies</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Dibyanayan</namePart>
        <namePart type="family">Bandyopadhyay</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">MD</namePart>
        <namePart type="given">Tanzeel</namePart>
        <namePart type="given">Adam</namePart>
        <namePart type="family">Khan</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Asif</namePart>
        <namePart type="family">Ekbal</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-11</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Yue</namePart>
            <namePart type="family">Zhang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ruifeng</namePart>
            <namePart type="family">Xu</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ayu</namePart>
            <namePart type="family">Purwarianti</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Tong</namePart>
            <namePart type="family">Rong</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jun</namePart>
            <namePart type="family">Suzuki</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">Hengqin, China</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-444-6</identifier>
    </relatedItem>
    <abstract>Search-based reasoning and chain-of-thought (CoT) prompting offer two contrasting inference-time paradigms for language models: CoT commits to a single linear trajectory, while search explores alternative partial reasoning paths before selecting an answer. In this work, we study this trade-off using A* as a controlled instantiation of heuristic-guided search. Across StrategyQA, LogicQA, HotpotQA, GSM8K, and MATH, and across three model scales, we find that search is not a universal upgrade over CoT; instead, the two strategies solve substantially different subsets of instances. To explain this complementarity, we introduce a Reasoning Topology Framework based on structural features computable before generation. A derived branching coefficient is strongly associated with when search helps across mixed-benchmark topology clusters, suggesting that the CoT–search boundary is partly explained by problem structure. We further train a topology-aware router that selects between CoT and A* without observing generated traces. Despite this restriction, it outperforms post-hoc routers based on semantic entropy and LLM-as-critic judgments. We attribute this to a Fluency–Correctness Asymmetry: fluent traces are often incorrect, making trace-level adjudication unreliable. Overall, our results suggest that adaptive reasoning should route by problem structure, deploying search only when the input topology warrants its higher cost. Code: https://github.com/newcodevelop/astar-cot-reasoning.</abstract>
    <identifier type="citekey">bandyopadhyay-etal-2026-language</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-aacl.237/</url>
    </location>
    <part>
        <date>2026-11</date>
        <extent unit="page">
            <start>4663</start>
            <end>4681</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T When Should Language Models Search? A Topology-Aware Analysis of Reasoning Strategies
%A Bandyopadhyay, Dibyanayan
%A Khan, MD Tanzeel Adam
%A Ekbal, Asif
%Y Zhang, Yue
%Y Xu, Ruifeng
%Y Purwarianti, Ayu
%Y Rong, Tong
%Y Suzuki, Jun
%S Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-444-6
%F bandyopadhyay-etal-2026-language
%X Search-based reasoning and chain-of-thought (CoT) prompting offer two contrasting inference-time paradigms for language models: CoT commits to a single linear trajectory, while search explores alternative partial reasoning paths before selecting an answer. In this work, we study this trade-off using A* as a controlled instantiation of heuristic-guided search. Across StrategyQA, LogicQA, HotpotQA, GSM8K, and MATH, and across three model scales, we find that search is not a universal upgrade over CoT; instead, the two strategies solve substantially different subsets of instances. To explain this complementarity, we introduce a Reasoning Topology Framework based on structural features computable before generation. A derived branching coefficient is strongly associated with when search helps across mixed-benchmark topology clusters, suggesting that the CoT–search boundary is partly explained by problem structure. We further train a topology-aware router that selects between CoT and A* without observing generated traces. Despite this restriction, it outperforms post-hoc routers based on semantic entropy and LLM-as-critic judgments. We attribute this to a Fluency–Correctness Asymmetry: fluent traces are often incorrect, making trace-level adjudication unreliable. Overall, our results suggest that adaptive reasoning should route by problem structure, deploying search only when the input topology warrants its higher cost. Code: https://github.com/newcodevelop/astar-cot-reasoning.
%U https://aclanthology.org/2026.findings-aacl.237/
%P 4663-4681

Markdown (Informal)

[When Should Language Models Search? A Topology-Aware Analysis of Reasoning Strategies](https://aclanthology.org/2026.findings-aacl.237/) (Bandyopadhyay et al., Findings 2026)

• When Should Language Models Search? A Topology-Aware Analysis of Reasoning Strategies (Bandyopadhyay et al., Findings 2026)

ACL

• Dibyanayan Bandyopadhyay, MD Tanzeel Adam Khan, and Asif Ekbal. 2026. When Should Language Models Search? A Topology-Aware Analysis of Reasoning Strategies. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026, pages 4663–4681, Hengqin, China. Association for Computational Linguistics.