Parcae: First Scaling Laws for Stable Looped Language Models Show Compute-Optimal Alternative to Parameter Scaling

arXiv·medium signal

Prairie et al. derive the first scaling laws for looped language models — architectures that increase effective depth by reusing parameter blocks instead of adding new ones. Traditional models scale by increasing parameterization at the expense of memory; looped architectures scale compute without proportional memory growth. The paper shows stable looping is achievable and provides practitioners a new efficiency frontier for resource-constrained deployments.