Latent-GRPO: Reinforcement Learning in Continuous Thought Space

A new method, Latent-GRPO, replaces discrete text reasoning tokens with continuous recurrent thought vectors in embedding space. This removes the token tax: 80–90% of generation time went to unread prose, and 19–34% of rollouts were truncated by token limits and scored zero reward.

• Discrete Chain-of-Thought spends 80–90% of generation time on text no one reads
• 19–34% of rollouts were cut off by token limits and received zero reward
• Continuous 5,120-dim thought vectors are fed back into the model instead of words
• The method solves the action-density problem for policy gradients via two-pass replay