Excellent work, it helps a lot!

I wonder if they would introduce subconscious tokens. Tokens produced by the model but not shown in the output. Like a hidden “step by step” logic training