Diffusion-GR2: Generative Reasoning Reranker with Diffusion

In the field of recommendation systems, generative re-ranking models have demonstrated exceptional accuracy by incorporating reasoning chains before reordering candidate lists. However, this capability comes at a cost: autoregressive decoding consumes a high number of sequential steps, limiting inference speed and hindering deployment in environments with large data volumes. To overcome this barrier, block diffusion models offer an efficient alternative by decoding multiple positions in parallel through a few denoising steps. Nevertheless, directly converting an autoregressive re-ranker to a diffusion-based one introduces two problems: a structural gap —due to positions being denoised in parallel, generating invalid rankings with duplicates or missing items— and a distributional gap —because fine-tuning with fixed teacher trajectories becomes off-policy during inference. The Diffusion-GR2 proposal closes both gaps by combining a conversion fine-tuning (CFT) that adapts the diffusion model initialized from the autoregressive one, an on-policy distillation (OPD) that supervises the model with its own decoded trajectories, and a reinforcement learning (RL) stage on the re-ranking reward. Experiments on Amazon Beauty show that Diffusion-GR2 recovers accuracy close to the autoregressive re-ranker, while parallel decoding multiplies performance by 2.4 to 3.5 times. This advance is especially relevant for companies seeking to integrate artificial intelligence into their recommendation processes without sacrificing speed. At Q2BSTUDIO, as a software and technology development company, we apply this type of innovation in our AI for business solutions, combining efficient generative models with scalable cloud architectures. Likewise, we implement custom applications that integrate advanced re-ranking techniques, optimized through AWS and Azure cloud services, and enhanced with Power BI dashboards for result analysis. Cybersecurity is also a pillar in our developments, ensuring that sensitive user data is protected even in high-performance deployments. This type of diffusion model, together with autonomous AI agents, opens the door to faster and more accurate recommendation systems, key for real-time decision-making. Ultimately, Diffusion-GR2 represents a firm step towards the democratization of generative re-rankers, making them viable for production environments without compromising quality.