The paper introduces a benchmark for evaluating the transferability of language models to new tasks. It proposes a learner that leverages selective initialization from past task checkpoints to maximize positive transfer and minimize negative transfer.

https://arxiv.org/abs//2307.05741

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

PODCASTS:

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

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