Reasoning with Continuous Latent Diffusion
The paper proposes Latent Flow Reasoning Models, a continuous-diffusion approach that refines full reasoning solutions in latent space.
The author says accurate decoding is not enough for strong reasoning, so LFRMs learn compact representations from multiple layers of an autoregressive teacher. The method uses asynchronous denoising, a compact prompt encoder, learned self-conditioning guidance, and gold-solution endpoints to improve sparse reward training. Reported results beat recent continuous-diffusion baselines at comparable backbone sizes on math reasoning and HumanEval code generation. With a 638M-parameter denoising backbone, LFRM-L reaches 63.74% pass@1 on GSM8K, 24.6% on MATH500, 32.85% on HumanEval, and 30.18% on HumanEval+. ArXiv · AI/CL/LG's note
The author says accurate decoding is not enough for strong reasoning, so LFRMs learn compact representations from multiple layers of an autoregressive teacher. The method uses asynchronous denoising, a compact prompt encoder, learned self-conditioning guidance, and gold-solution endpoints to improve sparse reward training. Reported results beat recent continuous-diffusion baselines at comparable backbone sizes on math reasoning and HumanEval code generation. With a 638M-parameter denoising backbone, LFRM-L reaches 63.74% pass@1 on GSM8K, 24.6% on MATH500, 32.85% on HumanEval, and 30.18% on HumanEval+. ArXiv · AI/CL/LG's note
score 6