Recursive AI-Led Development of Frontier Industrial Coding Model

Research Skills · Agent-led post-training · Executable evaluation

Overview · Paper · Framework · Components · Research Skills · Eval Harness · Model · Security

---

Overview

iCoder is a research project on agent-led model development for RTL design and GPU kernel optimization. This repository publishes the project's evaluation harness and Research Skills bundle. The full technical report is available here, and model artifacts are available as iCoder-27B.

The project targets industrial coding: RTL design and GPU kernel optimization, where acceptance is determined by specialized executable toolchains and deployment quality rather than surface similarity alone.

[!NOTE] iCoder is human-guided and agent-led, not a fully autonomous self-improvement system. Its workflows require user-provided objectives, task instructions, verifiers, resource limits, and an isolated execution environment.

Framework

_{Research Skills guide the agent loop; task-native execution feeds evidence back into experimentation.}

The project framework separates instructions supplied before an experiment from evidence produced during execution:

Development path

Data construction first creates a shared executable task pool. Parameter updates then proceed through SFT → OPSD → RLVR:

The stages form a feedback loop: evaluation can send the process back to data construction, verifier work, or an earlier training decision. This repository focuses on the released evaluation and Research Skills components; it does not include the complete training infrastructure.

Repository components

Evaluation harness

The harness turns generated artifacts into task-native evidence:

model endpoint → candidate generation → isolated verification → structured artifacts

It separates generation from resource-intensive verification, records structured outcomes and provenance fields, and keeps benchmark-specific logic behind a shared orchestration interface. See the harness guide.

Getting started

Clone the repository once:

git clone https://github.com/bingreeky/iCoder.git
cd iCoder

To use the agent workflow, install the Skill directories and start with the auto-post-training controller. The Research Skills guide explains installation, bootstrap inputs, the Human Prior boundary, and the role of every Skill.

To run model evaluation, enter eval_harness/ and follow the toolchain, dataset, and endpoint setup.

Repository layout

iCoder/
├── eval_harness/       evaluation and executable-verification infrastructure
├── research_skills/    versioned Research Skills and release manifest
└── README.md            project overview

Security

The evaluation harness compiles and executes model-generated code. Run it only in an isolated, disposable environment with strict filesystem, process, resource, and network controls. Never place personal data, model credentials, or unrelated secrets on an evaluation host.

See the harness security guidance for operational details.

Citation and release resources

@article{yang2026icoder,
  title   = {iCoder-27B: Recursive AI-Led Development of Frontier Industrial Coding Model},
  author  = {Cheng Yang and Jiayang Lyu and Shangyuan Liu and Guibin Zhang and
             Jiong Lin and Xinlei Yu and Junchi Yan and Shuicheng Yan and
             Weinan E and Linfeng Zhang and Linfeng Zhang and Qibing Ren},
  journal = {arXiv preprint arXiv:2609.29626},
  year    = {2026}
}

Citation metadata for the evaluation harness is available in eval_harness/CITATION.cff. Model artifacts are available from the i-Coder organization on Hugging Face.

License

The evaluation harness is released under the MIT License. Vendored components and benchmark datasets retain their original licenses; review the third-party notices before redistribution.