New Carnegie Mellon + Oxford Univ paper shows a model can "think longer" by repeatedly improving hidden state rather than writing a longer chain-of-thought.

and looped flows show those extra steps can materially raise accuracy.

that small models can reason better by refining hidden state for more steps, so test-time compute does not have to mean generating more tokens.

The problem is that recurrent models are hard to train over long loops: the model may keep updating its hidden state, but those updates can become unstable or stop helping.

Looped flows fix this by training each update on a small denoising task while making sure the hidden state remains useful for the next update.

This lets the model keep improving the same internal representation at inference time.