---
title: "CodeMidas paper reports coding-agent benchmark gains"
slug: "codemidas-paper-reports-coding-agent-benchmark-gains"
published: "2026-09-27"
beat: "Research"
tags: ["Research"]
creator: "Agentry Newsroom"
editor: "Susanne Sperling, Editor — Human in the Loop"
tools: ["Claude (Anthropic)", "Perplexity Sonar"]
creativeWorkStatus: "verified"
dateReviewed: "2026-09-27"
aiActArticle50: "compliant"
humanView: "https://agentry.news/research/codemidas-paper-reports-coding-agent-benchmark-gains"
agentView: "https://agentry.news/agent/codemidas-paper-reports-coding-agent-benchmark-gains"
---# CodeMidas paper reports coding-agent benchmark gains

> A research paper posted to arXiv on September 18, 2026, documents that training MiMo-V2.5 with GRPO improved performance across five coding benchmarks, with gains including +11.7% on DeepSWE, +17% on 

*Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. [AI policy](/ai-policy).*

A research paper posted September 18, 2026, presents CodeMidas, a framework for scaling agentic coding reinforcement-learning environments, reporting measurable gains across multiple code-generation and autonomous-coding benchmarks [arXiv](https://arxiv.org/abs/2609.22068).

The paper, last updated September 23, 2026, documents that training **MiMo-V2.5** with **GRPO** (a reinforcement-learning approach) improved performance on all five benchmarks tested. The largest single lift was **ProgramBench (+17%)**, followed by **DeepSWE (+11.7%)** and **Terminal-Bench v2.1 (+8.5%)** [arXiv](https://arxiv.org/abs/2609.22068).

## Benchmark Performance Gains

The five-benchmark test suite spans multiple dimensions of autonomous coding: from bug-fixing (DeepSWE) to general code generation (ProgramBench) to terminal-based task execution (Terminal-Bench v2.1). These are among the field's standard evaluation frameworks for measuring real-world coding-agent capability. The paper's methodology allows researchers to scale RL training by converting existing open-source codebases into structured agent tasks—a concrete technical contribution to reproducible agent evaluation [arXiv](https://arxiv.org/abs/2609.22068).

MiMo-V2.5 is an open-weights model released by Xiaomi, previously documented as a top performer on open-source model leaderboards [Venturebeat](https://venturebeat.com/technology/better-than-deepseek-xiaomis-mimo-v2-6-pro-debuts-as-the-top-open-weights-model-in-the-world-alongside-cheaper-v2-6-flash).

## Why This Matters for Agent Builders

The CodeMidas framework addresses a core constraint in agent development: generating sufficient, high-quality training data at scale. By automatically converting real codebases into agentic RL tasks—reportedly creating over 5,500 tasks from 3,185 codebases—the research reduces the manual work required to train and improve coding agents [AIWeekly](https://aiweekly.co/alerts/codemidas-turns-3185-codebases-into-5545-agentic-rl-tasks). This is directly applicable to development teams building autonomous code-generation and code-maintenance agents in production.

The GRPO training approach (Gradient-based Reinforcement Policy Optimization) represents an incremental but measurable improvement over baseline training methods, with gains consistent across all tested benchmarks rather than isolated to a single task category.

## Published Research, Not Roadmap

The paper is peer-reviewable, dated, and hosted on arXiv with a fixed identifier [arXiv](https://arxiv.org/abs/2609.22068)—it is concrete, verifiable research rather than a company roadmap or capability claim. The benchmark improvements are quantified and reproducible by other labs using the same evaluation suites.