---
id: 20260922-T0-11
title: "代码生成RL的测试太宽松？GT-Anchored Verifier用信息增益筛选测试"
title_en: "GT-Anchored Verifier Co-Training Improves Code Generation RL Reliability"
url: https://ai.daily.yangsir.net/daily/20260922-T0-11
issue_date: 2026-09-22
publish_date: 2026-09-21T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2609.21208
---

# 代码生成RL的测试太宽松？GT-Anchored Verifier用信息增益筛选测试

让同一个语言模型同时扮演代码作者和测试作者的自博弈方法，有望让代码生成强化学习摆脱固定测试集，但存在两个问题：一是宽容度崩溃，通过率奖励被最大化后测试变得过于宽松；二是测试多样性不足。研究者提出GT-Anchored Verifier Co-Training，用信息增益奖励对多样性剪枝后的测试进行筛选，并以真实标注（GT）锚定验证器。论文编号arXiv:2609.21208，未公布在HumanEval等基准上的具体提升幅度。

## English Version

**GT-Anchored Verifier Co-Training Improves Code Generation RL Reliability**

Self-play methods where a single language model acts as both coder and test author promise to move code-generation RL beyond fixed test suites, but suffer from two coupled pathologies: permissiveness collapse, where pass-rate rewards are maximized and tests become too lenient, and insufficient test diversity. Researchers propose GT-Anchored Verifier Co-Training, which uses information-gain rewards over diversity-pruned tests and anchors the verifier to ground truth. Posted as arXiv:2609.21208, no concrete improvement figures on benchmarks like HumanEval are yet available.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2609.21208)

**详情页**：https://ai.daily.yangsir.net/daily/20260922-T0-11

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*