Unified multitask models
One model, one format, many tasks.
| Concept | Introduced by | Years | Papers |
|---|---|---|---|
| Every task as question answering Cast ten NLP tasks as QA over a context so one model handles them all. | decaNLP | 2018 | 0 |
| Multi-task vision-language learning Train one VL model on many tasks at once and share the representation. | Multi-task hierarchical VL | 2018–2021 | 2 |
| Text-to-text framing Cast every task as text in, text out: one model, one loss. | T5 | 2019–2022 | 19 |
| All VL tasks as text generation Answer every vision-language task by generating its label as text. | VL-T5 | 2021 | 1 |