[edit]
Self-Augmenting Retrieval for Diffusion Language Models
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:55110-55124, 2026.
Abstract
Diffusion language models generate text by iteratively denoising an entire response in parallel, holding a noisy draft of the answer at every step. We show that these drafts surface salient entities early in the denoising trajectory, making them a useful lookahead signal for retrieval-augmented generation. We further show that grounding generation in retrieved evidence sharply reduces inter-token dependence, making RAG particularly well suited to parallel decoding. We exploit both through Self-Augmenting Retrieval for Diffusion Language Models (SARDI), a dynamic RAG framework that uses these intermediate drafts to guide retrieval during denoising. SARDI is training-free, retriever-agnostic, and applicable to any reasoning-capable discrete diffusion language model. Across five multi-hop QA benchmarks, SARDI outperforms training-free diffusion and autoregressive retrieval baselines at up to $5\times$ higher throughput.