You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I love the idea, haven't look at the code but. What do you think of refining the break case for the recursion as undivideable task ? This to prevent delegation of a single task.
A task that can no longer be divided into subtasks cannot be delegated.
Pretty sure you have thought about it, what problem did you encountered when using that approach ?
The text was updated successfully, but these errors were encountered:
Yeah, the happy path is when the LLM decides not to decompose a simple-enough task. The biggest challenge is deciding when a task is "simple enough" - most LLMs don't have a good sense of what they do/don't "know" so they tend to overcommit to a too complex task or undercommit (see §5 of our paper). It might be solvable using better prompting (though this requires more human labor), or on demand if a node fails a subtask and you have an oracle (e.g., https://arxiv.org/abs/2311.05772).
I love the idea, haven't look at the code but. What do you think of refining the break case for the recursion as undivideable task ? This to prevent delegation of a single task.
A task that can no longer be divided into subtasks cannot be delegated.
Pretty sure you have thought about it, what problem did you encountered when using that approach ?
The text was updated successfully, but these errors were encountered: