NavTree
A hierarchical document retrieval method that operates without LLM-generated summaries. It constructs a deterministic balanced segmentation tree as a navigation scaffold and employs a hybrid frontier traversal strategy to achieve efficient long-document retrieval, requiring no LLM calls during the indexing phase.
Timeline (last 90 days)
The arXiv paper "Tree Navigation Without LLM Summaries" argues that the main benefit of summary trees in long-document question answering may come from navigation rather than the content of LLM-generated summaries.
NavTree is a leaf-node-only retriever that builds a deterministic balanced segment tree on top of text chunks; the entire indexing process does not call any language model, making the indexing cost zero.
NavTree's tree itself stores no summaries and is used only as a navigation skeleton; during retrieval it employs a hybrid frontier traversal strategy combining lexical and dense methods, and ultimately outputs only leaf text chunks to the reader.
In equal-cost evaluations, NavTree is the strongest equal-cost hierarchical retriever and ties with the strongest flat baseline.
On long-document multi-hop QA tasks, NavTree is the only hierarchical method that significantly outperforms BM25 at the category-to-category level.
The generative RAPTOR variant, even equipped with high-quality clustering summaries, still loses to NavTree—which has zero index cost—at every multi-chunk budget tier.
NavTree's ranking results remain consistent across stronger readers, open-weight readers, stronger encoders, and full factorial experiments.
All Facts (7)
The arXiv paper "Tree Navigation Without LLM Summaries" argues that the main benefit of summary trees in long-document question answering may come from navigation rather than the content of LLM-generated summaries.
50%UnverifiedNavTree is a leaf-node-only retriever that builds a deterministic balanced segment tree on top of text chunks; the entire indexing process does not call any language model, making the indexing cost zero.
50%UnverifiedNavTree's tree itself stores no summaries and is used only as a navigation skeleton; during retrieval it employs a hybrid frontier traversal strategy combining lexical and dense methods, and ultimately outputs only leaf text chunks to the reader.
50%UnverifiedIn equal-cost evaluations, NavTree is the strongest equal-cost hierarchical retriever and ties with the strongest flat baseline.
50%UnverifiedOn long-document multi-hop QA tasks, NavTree is the only hierarchical method that significantly outperforms BM25 at the category-to-category level.
50%UnverifiedThe generative RAPTOR variant, even equipped with high-quality clustering summaries, still loses to NavTree—which has zero index cost—at every multi-chunk budget tier.
50%UnverifiedNavTree's ranking results remain consistent across stronger readers, open-weight readers, stronger encoders, and full factorial experiments.
50%