Looped LMs need not pay for every loop: near fixed points, they can scale FLOPs without scaling memory. Fixed-point shortcuts speed up training (pretraining, post-training) and inference (prefill, dec ...
Controlled Ouro-to-Huginn ablations isolate loop placement, input injection, and recurrent-state design, then test whether the findings transfer to compute-matched MoE models.
Parcae stabilizes loop models with an exact exponential decay for its recurrent state, while its input branch uses an Euler gain. Replacing that gain with exact zero-order hold (ZOH) improves validati ...
