The paper proposes SequenceMatch, an imitation learning framework for sequence generation that addresses the compounding error problem of autoregressive models. It incorporates backtracking and uses SequenceMatch-$\chi^{2}$ divergence as a training objective, leading to improvements over MLE on text generation.

https://arxiv.org/abs//2306.05426

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