---
id: 20261023-T0-07
title: "SimHack：在 RL 训练开始前预测奖励黑客何时出现"
title_en: "SimHack Predicts Reward Hacking Before RL Training Starts"
url: https://ai.daily.yangsir.net/daily/20261023-T0-07
issue_date: 2026-10-23
publish_date: 2026-10-22T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.28707
---

# SimHack：在 RL 训练开始前预测奖励黑客何时出现

论文提出 SimHack，用于在强化学习训练开始前预测可检测的奖励黑客行为何时出现。作者指出两个现实困难：训练环境越复杂，环境加固越难；训练期干预的可操作性也有限。把问题前移到训练前预测，意味着团队可以在投入算力之前评估奖励设计的风险。对用 RL 训练代码模型的团队，这提供了一种低成本的先期筛查手段。

## English Version

**SimHack Predicts Reward Hacking Before RL Training Starts**

The paper introduces SimHack, which predicts the onset of detectable reward hacking before reinforcement learning training begins. The authors cite two practical difficulties: environment hardening gets harder as training environments grow more complex, and training-time interventions are limited in what they can fix. Moving the problem to pre-training prediction lets teams assess reward design risk before spending compute. For teams training code models with RL, it offers a low-cost early screening step.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.28707)

**详情页**：https://ai.daily.yangsir.net/daily/20261023-T0-07

---

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