Retrieve-then-rerank pipelines restrict large language models (LLMs) to a fixed top-k candidate set, making their reasoning ability fundamentally limited by the recall ceiling of the first-stage dense retriever. On reasoning-intensive retrieval, we make two observations: first, relevant documents missed by dense top-k retrieval are frequently reachable within a few hops on the native approximate nearest neighbor (ANN) graph; second, LLM relevance judgments can identify high-yield anchors whose neighborhoods contain these missed documents. Based on these findings, we propose NaviRank, an LLM-guided navigation framework that performs greedy expansion over the HNSW graph: it uses highly parallel pointwise LLM judgments to steer anchor selection and local expansion, followed by listwise calibration for final ranking. Experiments on BRIGHT show that NaviRank achieves 27.0 NDCG@10, while establishing the strongest accuracy–efficiency Pareto frontier in both token usage and latency among all evaluated methods.

Visual document retrieval (VDR) bridges language-based systems with visually grounded knowledge. Recent VDR models adopt multi-vector late interaction, representing each document with multiple visual token embeddings and scoring relevance with MaxSim. However, existing late interaction engines struggle to support multi-vector VDR at scale due to the higher dimensionality and larger number of visual tokens per document.

We propose VISTA, a VDR-specific late-interaction engine that exploits the correlation among visual tokens within each document to optimize both compression and search. To the best of our knowledge, VISTA is the first VDR-specific late-interaction engine for scalable multi-vector visual document retrieval. Experiments on ViDoRe V2 and ViDoRe V3 show that VISTA achieves comparable or even better retrieval quality than compressed baselines, while consistently reducing latency by 1.2–6.7× and index size by 1.4–6.1×.