wtf is loop-ttt?

Loop-TTT is the code and data behind a research blog post about a type of AI language model called a looped, or recurrent depth, model. Instead of running through a fixed set of layers once, a looped model like the one studied here, called Ouro, can apply the same core block of layers multiple times before producing an answer, so how many times it loops becomes an adjustable setting rather than something fixed by the model's design. The research finds that answer accuracy on math problems improves as you increase the number of loops, but only up to the depth the model was originally trained at. Beyond that point, accuracy starts to fall again, as if the model overthinks the problem. Right at that peak point, the researchers try a small adjustment: updating about 100 tiny scaling values inside the model, done fresh for each batch of examples and then reset afterward, using a single step of a common optimization method. This small tweak improves the model's accuracy on generated math answers. The study's more surprising finding is that exactly where in the model this adjustment is applied matters more than whether it happens at all, and that much of the improvement behaves less like the model actually learning something new from each example and more like a reusable adjustment that works across many prompts. The repository includes the core code library for loading the Ouro models, preparing datasets and prompts, and the test time training update itself, along with the exact scripts used to run each experiment and produce each figure in the blog post. Curated result files are included so the reported numbers can be checked without rerunning every experiment, though some raw experiment folders behind certain plots are not part of this release. Running the experiments requires downloading Ouro model checkpoints from Hugging Face and using a GPU. The code is released under the MIT license, while the underlying model checkpoints and benchmark datasets remain under their own separate licenses.