Replies: 1 comment 2 replies
-
|
Thank you very much for your suggestion. Indeed, errors caused by syntax mistakes outside of the program code can severely hinder the evolution process. We also discovered these issues with OpenEvolve, and therefore we've made numerous improvements to error handling in LoongFlow. For example, as you mentioned, we've added retries (you can control the number of retries yourself); optimized generation in React; and improved the evaluation mechanism, etc. You can try running your task using the LoongFlow framework; you'll see significant improvements in both efficiency and performance. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
I have used openevolve quite a bit for experiments with code that's a bit larger and more complex than the typical examples. One challenge I found was that sometimes (maybe 5% of iterations, very roughly speaking) the generated code fails for uninteresting reasons (syntax error, misusing some API). In these cases, stock openevolve puts the program in the archive with some default bad metrics. This seemed bad to me, so I implemented a mechanism to retry the LLM query instead. In my experience one retry is almost always enough to get a program that works, and I don't think I've ever seen more than 2 retries. Is there any interest in such a feature in LoongFlow?
Beta Was this translation helpful? Give feedback.
All reactions