Reinforcement Learning as a Co-Design of Product and Research


2025-04-09

主讲人:Karina Nguyen

Product

Canvas:从普通的 next-token prediction 到 complex tasks 的转变。让用户和模型交流的方式更加多样(Training the model to become a collaborator)

  1. 能 preview 代码。不仅是代码可以直接运行,前端代码可以直接看效果
  2. 可以方便地对某个词语,某个句子选中进行 QA

Two ways of building research-driven products:

  1. Familiar form factor for unfamiliar capability:直接套壳
    1. 100K Context + File uploads: 这个其实现在比较稀松平常了?就是说把 context 加长来支持长 PDF claude acts as business analyst - YouTube
    2. \(\mathbb{P}(\text{I Know})\) / Calibration:其实这部分她语速太快稍稍有点没太听懂。好像说的是,比如一个 unreliable 的长文章,我通过模型去把有用的(或者说是保真的)信息快速提取出来。越保真就越高亮,来支持快速阅读
    3. Streaming model's thoughts:这个也很平常了,就是用输出些中间过程来防止用户等的太急
  2. Start with product belief & vision -> make the model do that:根据需求 finetune 模型 - I just think product designers and model trainers need to collaborate more with each other
    1. Redesign 新闻 page(没咋听懂):Most effective way to get people to read more of storylines coverage is by adding a layer of context to our product and coverage-across major surfaces - that connects people to the stories and information they need.
    2. more humane command line
    3. Claude 每个 chat session 的标题是根据用户语言习惯去标的,也就是说,他在生成 chat session 的时候用到了 user info。
    4. Claude in slack:群助手,好像功能更多,比如每周五 summarize channel,然后还有很多其他的 agent 功能。

这个是 \(\mathbb{P}(\text{I Know})\) / Calibration 的例子:

P(IK) / Calibration:在 AI 新闻文章中用不同深浅的高亮展示信息置信度

这个是 more humane command line 的例子:

Mesh 命令行界面:将命令执行过程、运行状态和完成的任务呈现为独立卡片

Model Behaviors

Refusing

Gives opinions but with caveats

尽量不要直接说 “I don't actually have a point of view” 在有争议的问题上,而是要适当的发表自己的观点,并提出这个观点有 bias

More nuanced refusals:

拒绝回答的示例:助手用 I do not feel comfortable 表达自己的感受

模型拒绝的时候,应该多用 “我感觉这个问题很难回答”,而不是 “你的问题很难回答”。说自己的 feeling 而非直接评价问题。

Building evals that we would trust

用 RL 来让模型学会正确的拒绝

另外几种奇怪的需要拒绝的类别:

  • Function calling refusals:需要告诉模型他没有 physical body to perform tasks in real-world
  • Long doc refusals:Data in the form of "I don't have vision capabilities to do XXX"

Construct RL env and Rewards

RL 的时候,不一定是只给用户输入,而是用户输入+用户信息。因为不同用户会有不同的 preference。

难以衡量的主观任务:写作、视觉设计与审美、情商与主动性,以及语音模式中的社交智能

有了用户信息,这些东西就可以变得 objective 了。就是去迎合用户。

Reward Hacks

没细讲,但提了两篇文章,还没来得及认真看:

  1. Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation (Baker et al., 2025)
  2. Reward Hacking in Reinforcement Learning (Weng, 2024)

References

Baker, B., Huizinga, J., Gao, L., Dou, Z., Guan, M. Y., Madry, A., Zaremba, W., Pachocki, J., & Farhi, D. (2025). Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation. doi.org
Röttger, P., Kirk, H., Vidgen, B., Attanasio, G., Bianchi, F., & Hovy, D. (2024). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 5377–5400. doi.org
Weng, L. (2024). Reward Hacking in Reinforcement Learning. Lil’Log. lilianweng.github.io
Zhao, W., Ren, X., Hessel, J., Cardie, C., Choi, Y., & Deng, Y. (2024). WildChat: 1M ChatGPT Interaction Logs in the Wild. The Twelfth International Conference on Learning Representations. openreview.net

Cite this post

@misc{pu2025rlinproducts,
  author = {Pu, Fanyi},
  title  = {Reinforcement Learning as a Co-Design of Product and Research},
  year   = {2025},
  month  = {4},
  url    = {https://pufanyi.com/blog/rl-in-products}
}