Closed kfsone closed 4 days ago
I just discovered AutoAgent, which seems to do exactly that:
Here is the Paper: https://arxiv.org/pdf/2309.17288.pdf
It could be helpful as an inspiration.
@JieyuZ2 FYI
We are closing this issue due to inactivity; please reopen if the problem persists.
check out our new efforts in agent team formation: https://arxiv.org/abs/2405.19425
When I ran the arxiv-research example to try an engineering-practices research ("research and try methods for building and packaging python modules as wheels") experiment, I got a noteworthy response from the critic up-front:
Let's take a moment to remember what the 'A' stands for, Artificial. The nature of all available training data to date is going to bias the sets of patterns on which these systems could possibly have been trained, especially bias it away from work-sharing collaboration.
If one looks at the prompts used by, say, ChatDev, we can see that users typically do not understand the implications and ramifications of this being text prediction and begin to herd cats.
Analogous to micromanaging in humans where the other party in a conversation floods the exchange with communication parameters and predicates, if we dilute the input with an excess of shaping elements we dilute attention and decrease the likelihood of a good match against the positive patterns we're looking for.
vs
(^- attention here is diluted so badly for GPT-4, bison, and codey that it is near guaranteed to do the opposite of what the prompt may appear to say to a human)
To this end, I want to suggest that the AutoGen team dedicate some aspect of AutoGen to team formation. For this it may be desirable to potentially run some local LLM to act like a HypervisoR agent, capable of rejecting responses or prompts, or at least pushing back on them.