I'm excited to announce two papers of ours which will be presented this summer at @naaclmeeting.bsky.social https://eting.bsky.social and @iclr-conf.bsky.social !
🧵
🧵
Comments
Thank you, Kanishka and Aaron. I could not have hoped for better collaborators! https://arxiv.org/abs/2410.22590
[👇 https://bsky.app/profile/juand-r.bsky.social/post/3lahqgko2wc2m]
When does CoT help? It turns out that gains are mainly on math and symbolic reasoning.
Check out our paper for a deep dive into MMLU, hundreds of experiments, and a meta-analysis of CoT across 3 conferences covering over 100 papers! https://arxiv.org/abs/2409.12183
CoT is just implementing a basic memory strategy by storing some states in the sequence of tokens.
Another proof that LLMs are a terrible algorithm...