Papers.Data.Code
Paper

CodeMidas: Scaling Agentic Coding RL Environments from Code Itself

Bowen Ye, Lei Li, Shicheng Li, Zihao Yue, Linghao Zhang, Hanglong Lv, Yuanxin Liu, Wenhan Ma

Uses source code alone to create executable coding RL tasks, with behavior-derived specs, reference-executed tests, and rollout-based task filtering.

arXiv ↗Sep 18, 2026

  • llm
  • code generation
  • reinforcement learning
  • software engineering
Pipeline diagram showing data filtering funnel from source code to RL environments
Pipeline diagram showing data filtering funnel from source code to RL environments

What it is

Builds coding RL environments directly from source code by extracting implemented functionality, synthesizing tests, and filtering tasks for reliable execution rewards. Instead of relying on issues, commits, tests, or docs, it turns code itself into RL tasks: agents infer behavioral specs from existing functionality, build execution-grounded tests, and aggressively filter for leakage and verifier defects.

Key results

  • DeepSWE pass rate rose 10.0%→21.7% (+11.7 pp)
  • ProgramBench Almost Solved rose 4.5→21.5 (+17.0 pp)
  • Terminal-Bench v2.1 pass rate rose 63.7%→72.2% (+8.5 pp)

Shows that implemented functionality in real codebases can be converted into reliable RL training tasks that transfer across diverse coding-agent benchmarks.

One useful ML email a week.

More from Sep 21 – Sep 27

The whole week →

Picked Sep 25

  • Dataset
    microsoft/ProgramDistill

    Unlike GUI-agent benchmarks, it evaluates coding agents by restoring masked web-app features with replayable browser traces and executable verifiers across 4,063 tasks.

    Hugging FaceSep 24, 2026
    →
  • Paper
    PACT: From Credit Assignment to Critic Alignment

    PACT gets 72.87% avg on 4 math benchmarks, +8.80 over GRPO and +13.16 over PPO; 67.4% on SWE-bench Verified, +2.0 over GRPO.

    arXivSep 22, 2026
    →
  • Paper
    Recursive self-improvement of AI research agents

    In 8 days it found 7 accepted rewrites, raised grade 0.703→0.778, matched/exceeded a human-built agent on 4 held-out benchmarks, and cut reward hacking 55%→32%.

    arXivSep 22, 2026
    →
  • Repo
    TencentARC/GameHorizon

    Covers 5,000 gameplay hours from 21 AAA titles, 5,000 offline MCQs, 20 online tasks, and benchmarks 47 models in one suite.

    GitHubSep 20, 2026
    →

Picked Sep 24

  • Dataset
    Harland/OmniVChat

    Makes native audio-visual dialogue measurable without text queries, ASR, or captioning, with fine-grained rubric scores shared across eval and RL.

    Hugging FaceSep 18, 2026
    →
  • Repo
    jaredpalmer/kev

    Unlike recent typed-decision repos, it packs many question types into one calibrated, decode-free pass with 2.0× speedup and 4e-6 output parity.

    GitHubSep 17, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.