💡 This paper proposes LoopLM, a framework integrating implicit iterative reasoning into pre-training via latent space computation and entropy-regularized depth allocation, claiming small LMs can match l