一项研究介绍了一种基于账本的零样本自组织方法，通过协调多个模型实例来提升代码生成性能。在100个最新的困难LiveCodeBench问题上的评测中，本地部署的开源Qwen3.8-27B模型的性能从66.8%提升至92.4%，性能提升达25.6个百分点1。该方法使得多个Qwen3.8-27B模型实例的联合表现超越了Fable 5模型的90.4%准确率，同时成本显著更低1。
Qwen3.8 Flash Next在应用该编排方法后达到93.0%的pass@1准确率，不仅优于Fable 5的90.4%，其成本仅为对方的9%1。研究表明该方法在固定后端和自托管权重两条部署路径上都能接近前沿级编码准确率1。
A research approach utilizing ledger-based zero-shot self-organization through coordinated multiple model instances has demonstrated significant improvements in code generation capabilities 1. Testing on 100 challenging problems from LiveCodeBench, the locally deployed open-source Qwen3.8-27B model showed performance gains rising from 66.8% to 92.4%, surpassing Fable 5's 90.4% accuracy rate while operating at substantially reduced costs 1.
The most impressive results came from Qwen3.8 Flash Next, which achieved 93.0% pass@1 accuracy after orchestration—outperforming Fable 5 at 90.4% while consuming only 9% of the cost 1. The methodology demonstrated performance improvements of up to 25.6 percentage points and proved viable across both fixed backend and self-hosted weight deployment paths, enabling locally operated solutions to approach frontier-level coding accuracy 1.
评论
还没有评论，欢迎留下第一条。