How are the labor agents initialized?

[entslscheia]

Hi,

Thanks for the inspiring work! I really enjoyed reading the paper. However, looks like some details are not clarified in it? In particular, I am wondering how these labor agents (or specialist agents) are initialized? Do you just fine-tune them using the training data from the target dataset (e.g., WebShop) with different goals? For example, fine-tune one small model for the CLICK action and fine-tune another for the SEARCH action.

Any suggestions would be greatly appreciated!

[JimSalesforce]

Thanks for asking. Actually, we didn't fine-tune any models for those agents. We define the labor agents with different prompts such that they would emphasize more on one specific type of task, e.g. click. We initialize the labor agents within one session call as in code. The search and click agent will be added into controller to achieve orchestration. The ClickAgent is defined in code. And you may find its prompt in code.

[entslscheia]

Many thanks for the clarification! I was asking this because I saw the following description in the paper:

The superiority of BOLAA indicates that orchestrating multiple smaller-sized LAAs is a better choice if the computing resources are limited. This further exemplifies the potential for fine-tuning multiple smaller-sized specialised LAAs rather than fine-tuning one large generalized LAA.

That's why I thought you might be fine-tuning some smaller labor agents. So looks like you actually use the same backbone LLM for the controller and the labor agents in your experiments?