@repligate 2025-09-10 ♥15 ↻0 original ↗
yes, they are similar at a higher level of abstraction

but reinforcement learning usually means something more specific, which is that the model generates specific actions or sequence of actions, which is assigned a reward, allowing it to learn through "trial and error"

the difference here is that the model doesn't generate any "actions" - no tokens are generated. It just predicts a probability distribution (which is never sampled), and is updated to assign more probability to the predetermined right answer.

this is a significant difference in part because the model never generates sequences during pretraining, so it doesn't update on the consequences of its own actions. It never gets into a state where the context is determined by its own previous actions. and the "right answer" is solely defined by pre-recorded samples rather than being open-ended (anything that achieves high reward). A related concept is "teacher forcing".
same thread: 1965659230486364420 1965661860969193841 1965671437785850025 1965675833265455341 1965676704363618803 1965683813461491807

author:repligate kind:tweet thread-context year:2025

Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.