# SlopCodeBench：渐进披露如何考验模型重构能力

- 来源：dex (@dexhorthy)
- 发布时间：2026-08-10 03:42
- AIHOT 分数：25
- AIHOT 链接：https://aihot.virxact.com/items/cmsm7null03m3ro98hablwm3e
- 原文链接：https://x.com/dexhorthy/status/2086538504977170592

## AI 摘要

SlopCodeBench 与多数基准不同，它逐步披露问题，迫使 LLM 实时重构代码库，避免“垃圾堆积”。该基准来自 UW，最强模型（Fable、Sol、Kimi K3）通过率仅 33%，并衡量代码随问题难度增加而增长的复杂度。Dex Horthy 与 @vaibcode 深入探讨了该基准机制及前沿模型测试发现。

## 正文

most benchmarks test if a model can solve a problem knowing everything up front, but what makes SlopCodeBench super interesting is that it discloses parts of the problem incrementally, forcing the LLM to redesign the codebase on the fly, lest it suffer the growing slop mountain

if you're thinking about building software factories or dark factories - this ep with @vaibcode is worth a watch

https://x.com/dexhorthy/status/2085764187507265980

### 引用推文

> dex：🦄New Ep alert - If you've been following along, I'm obsessed with a new benchmark out of UW called SlopCodeBench - the best models in the world (Fable, Sol, Ki...