-
Search when generating something with LLM
-
Make Every Move Count: LLM-based High-Quality RTL Code Generation Using MCTS
- Like (blu3mo)
-
Derivative of chain-of-thoughts
-
It’s all about how to define rewards~~ (blu3mo)
- Thinking about tasks that seem easy to define
-
I might want to try this with a fictional paper title + abstract
- Think about that chain
- Wait, this is definitely going to be interesting (blu3mo)(blu3mo)(blu3mo)
- Citations could be the reward or something
-
Ideas to do later
- It seems possible to create several types of researchers
- Like technical type vs vision type, bottom-up vs top-down, etc.
- It seems like updating the type=prompt itself genetically could be possible (blu3mo)(blu3mo)
- “Conferences” could also be simulated
- It seems possible to create several types of researchers
-
It might be interesting in the context of SciSci as well
- You could say, with this kind of system, this kind of paper was created
-
Different from MCTS
- Take 3
- Instead of traversing the tree, just take the top 3 normally
- Just taking them would make the leaf nodes grow too much, so exclude leaf nodes without 3
- Want to lighten the weighting as you go down to descendants.
- When there are two parents and there is a stretched lineage, it’s delicate for the score of the parent who hasn’t contributed much to increase
- Like with α=0.7, multiply α^depth by the value to be added to count and sum
-
This could be interesting if integrated with human feedback
- I remember discussing an idea with (ritar) about using gaze information
- If it catches attention