Fleuret: Teacher Forcing Trades Temporal Iteration for Depth in Transformers

François Fleuret explains that teacher forcing lets transformers convert temporal iteration over a medium-sized state into depth-wise iteration over a huge state, pleasing GPUs, but at the cost of learning internal state dynamics, leaving only token-by-token reasoning.

2026-10-24 ~ 2026-10-24 · 2 related posts