François Fleuret: AR models can do FOL perfectly without any FOL in their definition

francoisfleuret · x · 2026-09-26

Pushing back on the claim that next-token prediction fundamentally caps LLMs, François Fleuret argues a model's formal definition lacking a property doesn't mean the process it implements lacks it too. Example: train an autoregressive model on vast first-order logic examples and it will do FOL perfectly, with no FOL in its formal definition.

Related event: Meta researcher Fleuret debunks "LLMs just predict the next token" in 7-tweet thread(8 posts)→

Original post →

More from AGI Musings

AGI Musings channel →