arXiv 2606.19315 introduces Diffusion-Proof, the first diffusion language model (dLLM) applied to formal theorem proving. The standout: the dLLM beats autoregressive (AR) models on theorem-proving benchmarks, marking the first time a dLLM has surpassed AR on a "structured reasoning" task.
The "dLLM for theorem proving" angle: theorem proving is a structured reasoning task — the model must generate a step-by-step proof in a formal language (Lean, Coq, Isabelle). The structured nature of the task is well-suited to dLLM's "iterative denoising" process, which can refine a proof step by step.
The technical details: Diffusion-Proof uses a "proof-aware" noise schedule that respects the structure of the proof. The dLLM first generates a rough proof, then iteratively refines it, with the refinement guided by a "proof verifier" that checks each step. The result is a proof that is both formally correct and structurally clean.
The benchmark: on the miniF2F benchmark (formal math olympiad problems), Diffusion-Proof solves 73.2% of problems, beating the previous AR SOTA (67.8%) and matching the best closed-source model (LeanDojo, 71.5%). The biggest improvement is on "long proof" problems, where the iterative refinement shines.
The bigger takeaway: "dLLM for structured reasoning" is a significant new direction. The dLLM paradigm has been mostly used for open-ended generation (text, code, image), and theorem proving is a clean, structured test case. The Diffusion-Proof result suggests that dLLMs are particularly well-suited to structured reasoning, and the next round of dLLM research will likely focus on this direction.