Relevant papers

A new relevant paper by several researchers: https://arxiv.org/pdf/2110.08207.pdf
They show promising results on zero-shot task generalization by fine-tuning on task instructions and some examples.
They have a very detailed appendix (~150 pages) discussing constructions of tasks, training strategies, prompts etc. that could be helpful for this work.
This paper can be added as a second reference in README under section: "Any empirical evidence that this might be true?"
Thought it might be useful to create an issue to track relevant papers.