Think about what happens when a language model writes one word at a time. All the rich computation inside the network gets squeezed back into a single token before the next step starts, like summarizing a whole train of thought into one word and then beginning again. The authors argue this makes it hard to hold onto abstract intermediate states while reasoning. Their fix, T2MLR, caches a middle-layer representation from the previous token and folds it into an earlier layer of the next one, so partial thinking survives across steps with little added cost.

What stands out is how targeted it is. Recurring just a small middle block, about 20 percent of the network, often beats looping over every layer, and you can retrofit the pathway onto an existing 1.7B model with a brief finetune rather than retraining from scratch.

That is my reading of the abstract, so check the paper for the actual numbers and where the gains hold up.