A new AI model solves math and coding problems by diffusing directly through vector space, skipping the word-by-word generation that GPT-style (autoregressive) models rely on.
Researchers built the Continuous Embedding Diffusion Reasoner (CEDR), trained with an ELF-based recipe that learns compact representations pulled from multiple layers of a large autoregressive teacher model. A staged curriculum then trains a smaller prompt encoder that replaces that teacher entirely at inference time, since the team found prompts only need to preserve enough information for a correct answer score, not copy the teacher's features exactly. The setup also adapts a technique called DiffusionNFT, adding self-conditioning guidance and gold-solution endpoints to work around sparse reward signals during training. On a 638M-parameter denoising backbone, the supervised models beat other published continuous-diffusion baselines of similar size, and the post-training version hit 63.74% pass@1 (solved on the first try) on GSM8K and 24.6% on MATH500 at 64 denoising steps, plus 32.85% on HumanEval and 30.18% on HumanEval+ at 128 steps.
The real finding isn't the benchmark scores. It's that decoding text accurately doesn't automatically mean good reasoning, which is why the team had to separately optimize the latent representations feeding the diffusion process. Swapping the full teacher Transformer for a compact prompt encoder also matters practically: it removes a heavy dependency from inference and lets different parts of a solution get denoised asynchronously, at different rates, a flexibility autoregressive models don't have.
These numbers still trail flagship autoregressive reasoning models on the same benchmarks by a wide margin, and the promised code repository was not live as of publication, so for now latent diffusion reasoning reads as a promising research direction rather than a deployed replacement.