Retrieval-Augmented Generation (RAG) typically appends retrieved documents to the prompt, which increases the cost of self-attention and can reduce the model’s ability to reliably use evidence in long contexts. To address this issue, we propose Dyna...
Retrieval-Augmented Generation (RAG) typically appends retrieved documents to the prompt, which increases the cost of self-attention and can reduce the model’s ability to reliably use evidence in long contexts. To address this issue, we propose Dynamic Reasoning-Aware Parametric RAG, which converts retrieved documents into document-conditioned LoRA parameters and injects them into the model weights rather than the input context. The proposed method uses teacher-generated answers and explanatory reasoning as distillation signals, and adopts a two-stage curriculum that progresses from single-document to multi-document settings to improve training stability. Across four open-domain question answering benchmarks, our approach is competitive with strong dynamic baselines and shows its largest gains on long-tail and compositional question datasets. Ablation studies show that the benefit of reasoning supervision is sensitive to the backbone’s capacity and reasoning quality, and is most pronounced for sufficiently capable language models.