Soft Latent Thinking Enables Continuous Reasoning in Embedding Space

Soft Latent Thinking replaces the standard vocabulary head with a lightweight projector to allow autoregressive reasoning in continuous embedding space. Tested on DeepSeek-Qwen-1.5B and LLaMA-3.2-3B, the method improves pass@k metrics and reduces per-step compute compared to discrete token-based chain-of-thought.