1 paper · 1 filter
Roman Kniazev, Nathanaël Fijalkow
Do transformers, when trained on sequential reasoning traces, build internal models of the underlying task? And if so, does the structure of those internal representations mirror t…