Hear me out JEV architecture but internal layers loop twice to simulate latent reasoning

tried it didnt work, we need an actual trained head that specialized in latent reasoning , tbh latent reasoning itself is in like very early stages atleast in the oss world so we will have to wait before this becomes an actual thing , what works tho is reasoning normally and then assigning probabilities , like it works really well at cost of latency

Super interesting, so its like system two then?

feels like the model paradigm might come down to dynamic (draft+decode reasoning tokens) vs fixed (latent loops) reasoning

I can't hear you, there is no sound :D

kinda had a similar idea here but with auxiliary outputs, latent reasoning would be cool

JevDiffusion?