This paper addresses the limitations of augmented generative search models that achieve robust performance in knowledge-intensive tasks where query-document relevance can be identified through direct lexical or semantic matching. Existing retrieval systems struggle to capture many real-world queries involving abstract reasoning, analogical thinking, or multi-step reasoning. To address this challenge, we present DIVER , a retrieval pipeline designed for inference-intensive information retrieval . DIVER consists of four components: document processing to improve input quality, LLM-based query expansion via iterative document interaction, an inference-enhanced searcher fine-tuned on synthetic multidomain data using hard negatives, and a point-wise reranker that combines LLM-assigned usefulness scores with retrieval scores. On the BRIGHT benchmark, DIVER consistently outperforms competing inference-aware models, achieving state-of-the-art nDCG@10 scores of 41.6 and 28.9 for the original query. These results demonstrate the effectiveness of inference-aware retrieval strategies on complex, real-world tasks. The code and retrieval model will be made public shortly.
