GVS5H: Zero-shot self-orchestration boosts Qwen3.8-27B to 92.4% on LiveCodeBench

·2026.10.25 12:56

Key point

A training-free method using five coordinated model instances raises open-weight Qwen3.8-27B accuracy from 66.8% to 92.4% on hard coding tasks.

1 / 3

Introduces ledger-based zero-shot self-orchestration, a training-free method where fresh model instances decompose problems and coordinate via a shared filesystem.
Locally served, open-weight Qwen3.8-27B accuracy rises from 66.8% to 92.4% on the 100 latest hard LiveCodeBench problems.
Orchestrated Qwen3.8 Flash Next reaches 93.0% pass@1, exceeding Claude Fable 5's 90.4% at 9% of the cost.
Gains are not universal; some models show no improvement or performance degradation.
Transcript analysis attributes the accuracy gains to problem decomposition and persistent context.
Code is released under the MIT License, with the paper and figures under CC BY 4.0.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.