---
id: 20260712-T0-03
title: "新方法让AI智能体从模仿中学会对齐，无需人工标注"
title_en: "New RL Method Aligns AI Agents via Imitation, Cuts Human Feedback"
url: https://ai.daily.yangsir.net/daily/20260712-T0-03
issue_date: 2026-07-12
publish_date: 2026-07-11T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2607.07859
---

# 新方法让AI智能体从模仿中学会对齐，无需人工标注

arXiv新论文提出Feedback Manipulation Regularization方法，让智能体在模仿学习阶段就能自动实现与人类价值观的对齐，无需额外的RLHF训练。该方法通过操纵反馈信号，使智能体在模仿专家轨迹时自然学会符合人类偏好的行为。研究在多个机器人控制任务上验证，对齐效果接近甚至超过传统的在线RL对齐方法，大幅降低了对齐成本。

## English Version

**New RL Method Aligns AI Agents via Imitation, Cuts Human Feedback**

A new arXiv paper introduces Feedback Manipulation Regularization, enabling AI agents to align with human values during imitation learning without additional RLHF. By manipulating feedback signals, agents learn aligned behaviors from expert trajectories. Tests on robotics tasks show alignment quality comparable to or better than traditional online RL alignment, significantly reducing training costs.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2607.07859)

**详情页**：https://ai.daily.yangsir.net/daily/20260712-T0-03

---

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