LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models

LoopUS post-training framework converts standard pretrained LLMs into looped latent refinement architectures for improved test-time reasoning without retraining or disruptive retrofits.