“no deanonymized preprint may be posted in the month prior to submission.” Is that a mistake?

how well do *you* understand how AI reasoning works? test yourself here:
do-you-understand-ai.com

This is a limited study, only considering one language and only evaluating summarization performance, but so far these findings seem to hold even in follow-up experiments in other setups. There’ll be more on this coming up next year!

Even with the quite limited data I used, having more trainable parameters consistently lead to better scores.

The best setup was LoRA in the feed-forward module, followed by bottleneck adapters. LoRA in the attention module was less stable & performed worse, especially considering the number of parameters added. Prefix tuning and (IA)^3 didn’t really work in comparison.

Great points! The example prompt on the first page is truly painful💀

I’d like to be on there :)

bsky.app/profile/did:...