Researchers introduced InfiLoop, a residual-connection design for looped transformer models intended to prevent reasoning accuracy from degrading as more test-time loop iterations are added. The method uses content-based weighting and a learned temporal decay to determine which prior computations should be preserved and how much weight new updates should receive, while a streaming recurrence keeps memory use constant regardless of loop count. In experiments, a 7-million-parameter InfiLoop model reached 97.9% exact accuracy on the Sudoku-Extreme benchmark and 13.6% pass@2 on ARC-AGI-2. Its accuracy on Sudoku-Extreme continued improving even beyond 20,000 effective test-time computation steps.
