InfiLoop Maintains Reasoning Accuracy in Long-Horizon Transformer Recurrence

InfiLoop introduces loop-native residual connections to prevent performance degradation during tens of thousands of test-time iterations. By combining content-based weighting with learned temporal decay, the mechanism maintains a constant-size running summary of recurrent states to prevent noisy updates from overwriting correct deductions.