Bayze v0.14 adds five multi-model collaboration modes
Bayze developer reports that v0.14 expands the workflow engine from a draft-review-revision sequence into five multi-model collaboration modes: waterfall review, parallel selection, MoA aggregation, revision with escalation, and debate. The developer says six rounds of real testing produced seven fixes, exposing issues including lost context, conflicting token and character limits, same-side anchoring, and fabricated quantitative claims. The report also warns that using multiple models does not inherently improve results and presents measured costs for each mode.
The detailed testing account offers practical lessons for orchestrating multiple models, including why task-specific collaboration patterns, deterministic validation, explicit stances, and skepticism toward unverifiable model claims matter.
上一篇文章里，Bayze 还只是一个「草稿→评审→修订」的瀑布工作流引擎。三个月迭代到 v0.14，它已经有了五种多 AI 协作模式——而这五种模式里最有趣的 debate（辩论），是被六轮真实测试硬生生磨出来的。这篇文章记录全部过程，包括每一次翻车。
一、为什么「多模型」不等于「效果更好」
先泼冷水。我们把所有主流协作模式各跑了真实任务，结论浓缩成一张表：
| 模式 | 机制 | 实测成本 | 什么时候真的有用 |
|---|---|---|---|
| waterfall 复核 | A 起草→B 结构化评审→按需修订 | ~$0.001/次 | 高质量档位的默认选择 |
| parallel + 评委 | N 模型并行答→评委按硬约束选优 | $0.0013/次 | 带硬约束的任务（字数/格式） |
| MoA 聚合 | N 模型答→强模型合成一份 | $0.006/次 | 开放生成（无长度上限） |
| 修订循环+升级 | 评审打回→循环修订→用尽换强模型 | $0.002/次升级 | 初稿不稳的复杂任务 |
| debate 辩论 | 双方交锋 K 轮→裁决者终审 | $0.02-0.04/次 | 分析/决策/风险评估 |
#Bayze#Bayze v0.14#DeepSeek Chat#Qwen Plus#Claude Sonnet 4.5