I've skimmed up and down and read some sections a few times - I'm still not 100% sure what the main conclusions are. A summary section would make the article more accessible and valuable to others.
My understanding is that the main point here is:
- some problems require reasoning through many steps of computation,
- transformers have limited depth, so cannot solve problems that require many steps.
My understanding is that the main point here is:
- some problems require reasoning through many steps of computation, - transformers have limited depth, so cannot solve problems that require many steps.
Am I missing anything else?