<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="gandhi-etal-2026-steer">
    <titleInfo>
        <title>Steer, Don’t Solve: Training Small Critic Models for Large Code Agents</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Shubham</namePart>
        <namePart type="family">Gandhi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Yiqing</namePart>
        <namePart type="family">Xie</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Atharva</namePart>
        <namePart type="family">Naik</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Ruichen</namePart>
        <namePart type="family">Zhu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Carolyn</namePart>
        <namePart type="family">Rose</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>Coding tasks are typically complicated and require multiple capabilities, ranging from high-level planning to low-level implementation. While coding agents are optimized for the joint capabilities, individual capabilities may have different optima, and high-level planning remains a major bottleneck. To address this challenge, we train a separate critic model that is specialized in high-level planning to steer the coding agent in inference. We construct SFT and DPO data to train the critic model to identify errors made by the coding agent and provide correct and clear high-level guidance without generating concrete actions. Experiments show that our fine-tuned 4B and 8B critic models significantly improve the performance of 6 larger coding agents (e.g., improving the resolved rates of GLM-4.7-Flash-30B-A3B and GPT-OSS-120B by 16.0% and 14.4% on SWE-Bench Verified). The critic model also reduces the total inference costs for some coding agents by solving tasks in fewer steps (e.g., reducing the per-example inference cost for GPT-OSS-20B from 0.07 to 0.03). We provide the code in https://github.com/shubhamrgandhi/critic-training.</abstract>
    <identifier type="citekey">gandhi-etal-2026-steer</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2478/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>56259</start>
            <end>56276</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Steer, Don’t Solve: Training Small Critic Models for Large Code Agents
%A Gandhi, Shubham
%A Xie, Yiqing
%A Naik, Atharva
%A Zhu, Ruichen
%A Rose, Carolyn
%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 gandhi-etal-2026-steer
%X Coding tasks are typically complicated and require multiple capabilities, ranging from high-level planning to low-level implementation. While coding agents are optimized for the joint capabilities, individual capabilities may have different optima, and high-level planning remains a major bottleneck. To address this challenge, we train a separate critic model that is specialized in high-level planning to steer the coding agent in inference. We construct SFT and DPO data to train the critic model to identify errors made by the coding agent and provide correct and clear high-level guidance without generating concrete actions. Experiments show that our fine-tuned 4B and 8B critic models significantly improve the performance of 6 larger coding agents (e.g., improving the resolved rates of GLM-4.7-Flash-30B-A3B and GPT-OSS-120B by 16.0% and 14.4% on SWE-Bench Verified). The critic model also reduces the total inference costs for some coding agents by solving tasks in fewer steps (e.g., reducing the per-example inference cost for GPT-OSS-20B from 0.07 to 0.03). We provide the code in https://github.com/shubhamrgandhi/critic-training.
%U https://aclanthology.org/2026.emnlp-main.2478/
%P 56259-56276

Markdown (Informal)

[Steer, Don’t Solve: Training Small Critic Models for Large Code Agents](https://aclanthology.org/2026.emnlp-main.2478/) (Gandhi et al., EMNLP 2026)

• Steer, Don’t Solve: Training Small Critic Models for Large Code Agents (Gandhi et al., EMNLP 2026)

ACL

• Shubham Gandhi, Yiqing Xie, Atharva Naik, Ruichen Zhu, and Carolyn Rose. 2026. Steer, Don’t Solve: Training Small Critic Models for Large Code Agents. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 56259–56276, Budapest, Hungary. Association for Computational Linguistics.