清华开源VeriLoop Coder-E1，证据治理螺旋驱动可验证递归自改进

精选理由

清华开源了个能自己改代码的模型VeriLoop Coder-E1，SWE-bench拿85.2分，比很多大模型都强，想搞代码修复的可以试试。

AI 摘要

清华团队开源VeriLoop Coder-E1，基于Qwen3.6-27B模型。模型在SWE-bench Verified获得85.20分，SWE-bench Pro为62.38分，Terminal-Bench 2.0达76.40分，DeepSWE为33.63分。该模型采用窄领域PEFT和Self-Harness技术，实现仓库级代码修复的可验证递归自改进。

原文 · pandaily

Tsinghua Team Open-Sources VeriLoop Coder-E1: Evidence-Governed Spiral Drives Verifiable Recursive Self-Improvement for Repository-Level Code Repair

VeriLoop Coder-E1 based on Qwen3.6-27B achieves SWE-bench Verified 85.20, SWE-bench Pro 62.38, Terminal-Bench 2.0 76.40, and DeepSWE 33.63 with narrow-domain PEFT and Self-Harness.

阅读原文