GVS5H introduces ledger-based zero-shot self-orchestration, a training-free method where multiple fresh instances of a single model decompose problems and coordinate through a shared filesystem holding a plan, notes, and current solution. Across eleven open and closed-weight models on the 100 latest hard Live Code Bench problems, the method yields gains of up to 25.6 percentage points on pinned backends. Orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1 against Fable 5's 90.4% at 9% of the cost, while locally served open-weight Qwen3.8-27B rises from 66.8% to 92.4%. Gains are not universal: some models are unchanged or worse. Transcript analysis attributes the gain to decomposition and persistent context.
The paper demonstrates that inference-time organization can approach frontier coding accuracy at a fraction of the cost, or slightly exceed it on self-hostable weights. The code requires uv and an API key for the model being tested. The codebase uses the multiagent engine by default; --engine single is the one-call baseline.
Results are written to runs/results.json, workspaces to runs/ws/. Code is under the MIT License; the paper, figures, and run data are under CC BY 4.0. The Live Code Bench fork and benchmark problem statements retain their own licenses. See NOTICE.md for license details.
Source: Hacker News · Summarized by HeadlinesBriefing