A training-free technique that has one AI model coordinate with fresh copies of itself can push open-weight coding models to near the level of a leading proprietary system, according to a new research paper and code released alongside it.
The method, which the authors call ledger-based zero-shot self-orchestration, works without any additional training. Fresh instances of a single model take turns breaking down a problem and working on it. They share information through a common filesystem that holds a plan, running notes and the current solution.
The team tested the approach across eleven open and closed-weight models on the 100 most recent hard problems in LiveCodeBench, a widely used coding benchmark. The results were uneven. Gains reached as high as 25.6 percentage points on pinned backends, but the authors note that some models were unchanged or performed worse.
The most eye-catching result involves Qwen3.8 Flash Next. With orchestration, the model reached 93.0% pass@1, compared with 90.4% for Fable 5. The paper puts the cost at 9% of what Fable 5 would run.
The second route involves a model that anyone can host. Qwen3.8-27B is open-weight, and when served locally it climbed from 66.8% to 92.4% with the orchestration method. That figure is the one the authors point to when arguing that self-hostable models can slightly exceed frontier accuracy.
To explain the improvement, the authors analyzed the transcripts of the runs. They attribute the gain to two factors: decomposition of the problem and persistent context that carries across the separate instances. In the paper's framing, inference-time organization can approach frontier coding accuracy at a fraction of the cost.
The paper also includes a caveat about one comparison. A run of Opus 5 used an earlier version of the scaffold, in a single pass served through OpenRouter rather than five runs on a pinned backend, so that result is not directly comparable to the others.
The code is available to run with uv and an API key for whichever model is being tested. The repository includes a manager mode, which runs the orchestration, and a single-call baseline for comparison. Results are written to a results file, and each workspace is saved to a separate folder for inspection.
The code is released under the MIT License. The paper, figures and run data fall under CC BY 4.0. The LiveCodeBench fork, the benchmark problem statements and the LaTeX template files carry their own licenses, which are listed in a separate notice file in the repository.
This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.