Recurrent Looped Transformer extends loops across tokens

While extending looped recurrence across the full sequence is an elegant theoretical approach to scaling latent depth without inflating per-token compute, RLT remains an untested conceptual design. Unified state transitions across training and inference eliminate prompt-response discrepancies during reinforcement learning, but token-level recurrence poses severe challenges for parallel training throughput. Ultimately, critical claims regarding reasoning improvements, hardware efficiency, and RL scaling remain theoretical until supported by empirical validation.