InfiLoop reportedly keeps improving with more loops beyond 20,000 effective steps on Sudoku-Extreme

The researchers describe “latent overthinking,” in which extra iterations can overwrite correct intermediate deductions in looped Transformers. They say InfiLoop learns which past computations to retain and how much to accept from each new update. Their paper reports that a 7M-parameter InfiLoop model reached 97.9% exact accuracy on Sudoku-Extreme and 13.6% pass@2 on ARC-AGI-2. On Sudoku-Extreme, they report continued improvement with test-time looping beyond 20,000 effective steps.