# AutoDesign：弱模型靠脚手架逼近前沿模型

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-24 00:05
- AIHOT 分数：36
- AIHOT 链接：https://aihot.virxact.com/items/cmt60ham60vo3ro73ji46zi6b
- 原文链接：https://x.com/rohanpaul_ai/status/2091557326364852677

## AI 摘要

AutoDesign 系统通过智能体在真实任务上运行、分析失误并自动重写提示词、工具、验证检查或重试规则，实现脚手架自动优化。在论文转会议海报任务中，七个智能体得分提升 5 至 19.6 点，最便宜模型获益最大，并发布 PosterBench 基准。研究表明，弱模型配良好脚手架可弥合与前沿模型的大部分差距，升级模型前应先优化现有代码。

## 正文

A weak model with well-built scaffolding around it can close most of the gap to a frontier model, so fix your code before you upgrade your model.

Scaffolding pays off in inverse proportion to model strength, so the money you save by switching to a cheap model can be recovered by engineering around it.

They built a system called AutoDesign that does this automatically. It runs an agent on real tasks, looks at what went wrong, then rewrites one piece of the surrounding setup: a prompt, a tool, a validation check, a retry rule.

A change survives only if it improves scores on the training tasks and doesn't hurt a held-out set, so the system can't just overfit its way upward.

They tested it on turning papers into conference posters, and released a benchmark, PosterBench, to score them.

The result is that scaffolding is worth more than most people assume, and worth the most to weak models. Seven agents each gained 5 to 19.6 points, with the biggest jumps going to the cheapest models.

So before you upgrade to a pricier model, spend a week improving the checks and retry logic around the one you have.

– arxiv. org/abs/2608.13560

Title: "AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design"