I don't think OpenAI is doing looped transformers, though it is possible they're doing something like MoEUT. However, that doesn't really explain how they are actually eliciting significantly denser reasoning. I don't think the RL signal is dense enough to saturate frontier math with minimal reasoning or perform complex actions without reasoning at all. Instead, it's far more likely they're doing synthetic rewriting of CoTs. Rewriting them to be terser, or outright dropping parts of the CoT and asking the next generation pretrain to fill in the middle, essentially internalizing the reasoning patterns deep in the weights. This is likely to me for many reasons, the first is that it gets around the RL signal problem. You turn a RL task into a dense SFT/pretraining task, so the signal is far denser and you get more bits per episode. The second is that it's simple and true breakthroughs are rare, and it's something that needs to happen at scale (you need pretraining scale data for this to work). If you're markov brained, the original CoT is our data generating process, and by dropping parts of it, we force the model to reconstruct the hidden variables (the hidden parts of the CoT) but in the model weights instead. (blog below)