If both ML models perform poorly, we could use a more traditional encoder-decoder transformer

[AML14]

I don't think we'll need to explore this idea anymore. The encoder-decoder models at the token level are working well.