<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ta-etal-2026-reinforced">
    <titleInfo>
        <title>Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Anh</namePart>
        <namePart type="family">Ta</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Junjie</namePart>
        <namePart type="family">Zhu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Shahin</namePart>
        <namePart type="family">Shayandeh</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Proceedings of the Fifth Workshop on Generation, Evaluation and Metrics (GEM)</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Simon</namePart>
            <namePart type="family">Mille</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Sebastian</namePart>
            <namePart type="family">Gehrmann</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Patrícia</namePart>
            <namePart type="family">Schmidtová</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Ondřej</namePart>
            <namePart type="family">Dušek</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Marzieh</namePart>
            <namePart type="family">Fadaee</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Kyle</namePart>
            <namePart type="family">Lo</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Enrico</namePart>
            <namePart type="family">Santus</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Gabriel</namePart>
            <namePart type="family">Stanovsky</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">San Diego, California, USA</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-423-1</identifier>
    </relatedItem>
    <abstract>Tool-calling agents are evaluated on tool selection, parameter accuracy, and scope recognition, yet LLM trajectory assessments remain inherently *post-hoc*. Disconnected from the active execution loop, such assessments identify errors that are usually addressed through prompt-tuning or retraining, and fundamentally cannot course-correct the agent in real time. To close this gap, we move evaluation into the execution loop at *inference time*: a specialized reviewer agent evaluates provisional tool calls *prior to* execution, shifting the paradigm from post-hoc recovery to proactive evaluation and error mitigation.In practice, this architecture establishes a clear separation of concerns between the primary execution agent and a secondary review agent. As with any multi-agent system, the reviewer can introduce new errors while correcting others, yet no prior work to our knowledge has systematically measured this tradeoff. To quantify this tradeoff, we introduce *Helpfulness-Harmfulness metrics*: helpfulness measures the percentage of base agent errors that feedback corrects; harmfulness measures the percentage of correct responses that feedback degrades. These metrics directly inform reviewer design by revealing whether a given model or prompt provides net positive value.We evaluate our approach on BFCL (single-turn) and τ²-Bench (multi-turn stateful scenarios), achieving +5.5% on irrelevance detection and +7.1% on multi-turn tasks. Our metrics reveal that reviewer model choice is critical: the reasoning model o3-mini achieves a 3:1 benefit-to-risk ratio versus 2.1:1 for GPT-4o. Automated prompt optimization via GEPA provides an additional +1.5–2.8%. Together, these results demonstrate a core advantage of separating execution and review: the reviewer can be systematically improved through model selection and prompt optimization, without retraining the base agent.</abstract>
    <identifier type="citekey">ta-etal-2026-reinforced</identifier>
    <location>
        <url>https://aclanthology.org/2026.gem-main.13/</url>
    </location>
    <part>
        <date>2026-07</date>
        <extent unit="page">
            <start>136</start>
            <end>147</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents
%A Ta, Anh
%A Zhu, Junjie
%A Shayandeh, Shahin
%Y Mille, Simon
%Y Gehrmann, Sebastian
%Y Schmidtová, Patrícia
%Y Dušek, Ondřej
%Y Fadaee, Marzieh
%Y Lo, Kyle
%Y Santus, Enrico
%Y Stanovsky, Gabriel
%S Proceedings of the Fifth Workshop on Generation, Evaluation and Metrics (GEM)
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, USA
%@ 979-8-89176-423-1
%F ta-etal-2026-reinforced
%X Tool-calling agents are evaluated on tool selection, parameter accuracy, and scope recognition, yet LLM trajectory assessments remain inherently *post-hoc*. Disconnected from the active execution loop, such assessments identify errors that are usually addressed through prompt-tuning or retraining, and fundamentally cannot course-correct the agent in real time. To close this gap, we move evaluation into the execution loop at *inference time*: a specialized reviewer agent evaluates provisional tool calls *prior to* execution, shifting the paradigm from post-hoc recovery to proactive evaluation and error mitigation.In practice, this architecture establishes a clear separation of concerns between the primary execution agent and a secondary review agent. As with any multi-agent system, the reviewer can introduce new errors while correcting others, yet no prior work to our knowledge has systematically measured this tradeoff. To quantify this tradeoff, we introduce *Helpfulness-Harmfulness metrics*: helpfulness measures the percentage of base agent errors that feedback corrects; harmfulness measures the percentage of correct responses that feedback degrades. These metrics directly inform reviewer design by revealing whether a given model or prompt provides net positive value.We evaluate our approach on BFCL (single-turn) and τ²-Bench (multi-turn stateful scenarios), achieving +5.5% on irrelevance detection and +7.1% on multi-turn tasks. Our metrics reveal that reviewer model choice is critical: the reasoning model o3-mini achieves a 3:1 benefit-to-risk ratio versus 2.1:1 for GPT-4o. Automated prompt optimization via GEPA provides an additional +1.5–2.8%. Together, these results demonstrate a core advantage of separating execution and review: the reviewer can be systematically improved through model selection and prompt optimization, without retraining the base agent.
%U https://aclanthology.org/2026.gem-main.13/
%P 136-147

Markdown (Informal)

[Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents](https://aclanthology.org/2026.gem-main.13/) (Ta et al., GEM 2026)

• Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents (Ta et al., GEM 2026)

ACL

• Anh Ta, Junjie Zhu, and Shahin Shayandeh. 2026. Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents. In Proceedings of the Fifth Workshop on Generation, Evaluation and Metrics (GEM), pages 136–147, San Diego, California, USA. Association for Computational Linguistics.