So at the heart of this architecture is what they call 'Markovian RSA', a combination of two papers RSA[0], which generates a certain amount of reasoning traces for a prompt; and the 'Markovian Thinker'[1] which seems to basically cut the end of those traces to keep context at a reasonable length.
I feel like there's potential to improve that part of just cutting a tunable amount (τ) of tokens off the tail end of those traces, because you may potentially lose valuable insight earlier in the trace? They did train the model (in SFT) to put the relevant information into the tail (τ) of the trace, but I'm not sure this is the best possible way.