This paper shows that decoder-only transformers can be seen as infinite multi-state RNNs, and pretrained transformers can be converted into finite multi-state RNNs. The proposed TOVA policy outperforms other policies in long-range tasks.

https://arxiv.org/abs//2401.06104

YouTube: https://www.youtube.com/@ArxivPapers

TikTok: https://www.tiktok.com/@arxiv_papers

Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016

Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers