The paper tackles a known trade-off in language model reasoning: generating several chain-of-thought rationales can uncover correct solutions that a single attempt misses, but those rationales are long and repeated autoregressive generation is costly.
To address this, the authors propose Rationale Distribution Distillation, a method that uses flow matching to map the distribution of rationales into a continuous latent space. Rather than sampling and reading full text rationales, the model can reason over the distilled distribution directly.
The work is framed as distributional continuous latent reasoning, aiming to preserve the benefits of diverse sampling while avoiding the expense of generating lengthy rationales. The preprint is available on arXiv.