Diffusion LLMs can match autoregressive (AR) reasoning performance by using AR-generated plans as globally visible scaffolds.
March 17, 2026
Original Paper
Think First, Diffuse Fast: Improving Diffusion Language Model Reasoning via Autoregressive Plan Conditioning
arXiv · 2603.13243
AI-generated illustration
The Takeaway
This identifies a fundamental coordination problem in diffusion-based text generation and solves it through hybrid plan conditioning. It enables diffusion models to achieve a 10-12% accuracy boost in reasoning tasks like GSM8K and HumanEval, bridging a major gap between the two architectures.
From the abstract
Diffusion large language models (dLLMs) generate text via iterative denoising but consistently underperform on multi-step reasoning. We hypothesize this gap stems from a coordination problem: AR models build coherence token-by-token, while diffusion models must coordinate all positions simultaneously. We propose plan conditioning, a training-free method that prepends a short (~100-token) natural-language plan from an AR model to the diffusion model's prompt. The plan serves as a frozen scaffold