English Español Français Deutsch
← Back to the list
Multilingual evaluation Foundational figures Directly upstream of ai100's methodology

Colin Raffel

Whether a single language model can do all NLP tasks at once when they're all framed as text-in-text-out — and whether that unification holds across languages.

T5 (Text-to-Text Transfer Transformer, 2020, lead author) made a methodological argument that the field had been resisting: every NLP task can be framed as the same kind of problem — give the model a string, ask for a string — and a single model trained on that framing handles them all. The follow-up T0 paper (2021) extended the argument to zero-shot multilingual generalization, showing that the same approach scales across languages with only modest performance loss. The combined line is the methodological precursor to almost every instruction-tuned multilingual LLM shipped since, including the ones ai100 currently evaluates.

Worth following when
you want to understand the architectural and training-data choices that made multilingual zero-shot generalization possible — and what those choices imply for evaluation methodology.
Topics
unified text-to-text framing of NLP tasks (T5); multilingual zero-shot generalization (T0); the lineage from research models to deployed multilingual LLMs.
Key works
T5: Text-to-Text Transfer Transformer (2020, lead author); T0 multitask-prompted multilingual generalization (2021, co-author); ongoing publications on model memorization and dataset attribution.