A joint plan lifts Qwen3.8-27B from 66.8% to 92.4% on code tests
Copies of one small open-weight model that use a shared plan can almost equal a large paid model, but not all models get an increase.
Claimed, not confirmed
Researchers report a method that uses no training. New copies of one model divide a hard code problem into parts. They work through a shared folder that holds a plan, notes and the current answer. On the 100 new hard problems of LiveCodeBench, Qwen3.8 Flash Next solved 93.0%. Fable 5 solved 90.4%. The cost was 9% of the cost of Fable 5, the researchers report.
How it works
The method has the name ledger-based zero-shot self-orchestration. A manager copy divides the problem. The other copies read and write the shared folder. The researchers report that the increase comes from the procedure that divides work and from the notes.
The numbers
The test included 11 models. The increase has a maximum of 25.6 percentage points on pinned backends, which are set servers. Qwen3.8-27B, which a team can operate on its hardware, went from 66.8% to 92.4%.
What is not known
Some models did not change, and some became not as good. Opus 5 operated on an early version, in one step through OpenRouter, not 5 on a pinned backend. The test is not the same for this model. No other group has verified the paper.
Insight: Method may matter more than model size
The paper reports that top code results usually come from large paid models. A related paper reports that changes to the harness repair procedure errors, but only changes to weights repair content errors. This reads as a move of the edge from who makes the model to who makes the control software, which helps open-weight models. Test this on your tasks before you sign a large contract.
Sources
Posted
