<aside>

• https://storage.googleapis.com/deepmind-media/Era-of-Experience /The Era of Experience Paper.pdf
• https://readpaper.com/pdf-annotate/note?pdfId=2798338000613368576&noteId=2801092661522993152
• 中文翻译【https://mp.weixin.qq.com/s/P3gBkdQXu12Zhm71HRvs0w】
• [2025 waic展示： xxx] </aside>

AI agents learn from their interactions with the world

TL;DR

DeepMind's 2024 paper introducing "Experience Replay" for AI agents. Unlike traditional human-centric RLHF, this approach grounds rewards in real-world signals (health metrics, exam results, environmental data) rather than human preferences. The key insight: agents that learn from environmental consequences can surpass human knowledge and discover novel strategies.

<aside>

• 如何打造真实Enviroment，帮助AI进行自主的Action？
  • 环境应该和真实的物理世界链接，给Agent配备足够多的工具可用
• 如何设计丰富的Reward，帮助AI进行自主的Reflection？
  • Reward 应该设置多种多样的，既要有客观的也要可以主观的； </aside>

ReAct ： Observations and Actions

Ultimately, experiential data will eclipse the scale and quality of human generated data.

Rewards

<aside>

What if experiential agents could learn from external events and signals, and not just human preferences?

</aside>