SLPO: Scaling Latent Reasoning via a Surrogate Policy
Reinforcement learning for test-time scaling in Chain-of-Thought reasoning is costly because it requires decoding every intermediate step as a token. SLPO introduces Surrogate Latent Policy Optimization, which uses a differentiable surrogate policy interface and a correctness-supervised stopping head to enable outcome-reward RL directly in continuous latent vectors. This allows the model to adapt its computation budget dynamically during inference.