Skip to content

[Missing Paper]: Steer, Don't Solve: Training Small Critic Models for Large Code Agents #3

Description

@shubhamrgandhi

Paper title

Steer, Don't Solve: Training Small Critic Models for Large Code Agents

Short name

Steer, Don't Solve

Authors

Shubham Gandhi, Yiqing Xie, Atharva Naik, Ruichen Zhu, Carolyn Rose

Publication month

2026-06

Venue

arXiv preprint

Category (can select multiple)

inference_scaling, sft

arXiv link

https://arxiv.org/abs/2606.21811

GitHub repository link

https://github.com/shubhamrgandhi/critic-training

HuggingFace link

https://huggingface.co/code-critic-model

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 such as high-level planning may have different optima and remain 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). Code: https://github.com/shubhamrgandhi/critic-training

Submission checklist

  • I have provided at least one valid paper link (arXiv/OpenReview/DOI/Website).
  • I have checked this paper is not already in the repository.

Additional notes

We train separate 4B and 8B critic models (SFT, then DPO) that read the coding agent's trajectory every k steps, identify errors, and return high-level guidance without generating actions, so the primary category is Inference-Time Scaling, next to SWE-PRM and Agentic Rubrics; the SFT category covers how the critic itself is trained. Open weights and training data are on HuggingFace. I am the first author; happy to open a PR against the YAML files instead if you prefer.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions