<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="dalal-etal-2026-test">
    <titleInfo>
        <title>More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Gal</namePart>
        <namePart type="family">Dalal</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Assaf</namePart>
        <namePart type="family">Hallak</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Gal</namePart>
        <namePart type="family">Chechik</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yftah</namePart>
        <namePart type="family">Ziser</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>Wider beam search should improve LLM reasoning, but when should you stop widening? Prior work on beam width selection has focused on inference efficiency (CITATION), without analyzing whether wider search can hurt output quality. We study each search step as scoring n candidates and keeping the top b of them. Best-of-n keeps a single winner (b=1); beam search keeps several. Because the scorer is noisy, a worse candidate can outrank a better one, and that risk grows as more candidates are compared. How far you can usefully widen therefore depends on the scorer: how large the true quality gap is relative to the noise. Keeping several hypotheses is easier than naming one winner, so beam search is more forgiving than best-of-n, but that per-step analysis does not by itself tell you the best width for the full task. We compare perplexity scoring and a process reward model (PRM) across three 7–8B models and all ten MR-BEN subjects (5,737 questions). Perplexity yields no accuracy gain at any tested width; the PRM gains up to 8.9 points and continues improving through beam width 4 (n=16 candidates per step). Across 444K search steps, the spacing of ranked scores agrees with a simple Gaussian ranking model. The same model and algorithm therefore support very different useful widths, depending on how reliable the scorer is.</abstract>
    <identifier type="citekey">dalal-etal-2026-test</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.165/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>3785</start>
            <end>3801</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search
%A Dalal, Gal
%A Hallak, Assaf
%A Chechik, Gal
%A Ziser, Yftah
%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 dalal-etal-2026-test
%X Wider beam search should improve LLM reasoning, but when should you stop widening? Prior work on beam width selection has focused on inference efficiency (CITATION), without analyzing whether wider search can hurt output quality. We study each search step as scoring n candidates and keeping the top b of them. Best-of-n keeps a single winner (b=1); beam search keeps several. Because the scorer is noisy, a worse candidate can outrank a better one, and that risk grows as more candidates are compared. How far you can usefully widen therefore depends on the scorer: how large the true quality gap is relative to the noise. Keeping several hypotheses is easier than naming one winner, so beam search is more forgiving than best-of-n, but that per-step analysis does not by itself tell you the best width for the full task. We compare perplexity scoring and a process reward model (PRM) across three 7–8B models and all ten MR-BEN subjects (5,737 questions). Perplexity yields no accuracy gain at any tested width; the PRM gains up to 8.9 points and continues improving through beam width 4 (n=16 candidates per step). Across 444K search steps, the spacing of ranked scores agrees with a simple Gaussian ranking model. The same model and algorithm therefore support very different useful widths, depending on how reliable the scorer is.
%U https://aclanthology.org/2026.emnlp-main.165/
%P 3785-3801

Markdown (Informal)

[More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search](https://aclanthology.org/2026.emnlp-main.165/) (Dalal et al., EMNLP 2026)

• More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search (Dalal et al., EMNLP 2026)

ACL

• Gal Dalal, Assaf Hallak, Gal Chechik, and Yftah Ziser. 2026. More Test-Time Compute Can Hurt: Overestimation Bias in LLM Beam Search. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 3785–3801, Budapest, Hungary. Association for Computational Linguistics.