Search systems are increasingly used for reasoning-intensive queries, where what makes a document relevant requires understanding or reasoning over the query-document relation rather than relying on surface vocabulary or topical similarity. The standard recipe - a cheap embedding-based retriever followed by an LLM verifier - works only when the embedding model places the right documents in its top-k, an assumption that recent reasoning-intensive IR benchmarks show often fails to hold even for SOTA embedding models. Recent query-side fixes such as query rewriting and agentic loops keep the LLM upstream of the cheap retriever and remain brittle to the embedder's failures and to the LLM's ability to rewrite the query from its parametric knowledge. In this paper, we explore a different paradigm - LLM-guided hierarchical search - in which an LLM interacts with the corpus directly via a hierarchically navigable search index, with no embedding model in the loop at search time. We propose LATTICE, an instantiation with two technical contributions: (i) a top-down LLM-guided construction of the search index using LLM judgements over multi-level document summaries, and (ii) a calibrated, path-aggregated LLM-guided traversal that mitigates noisy, slate-dependent LLM scores via cross-branch reference nodes. On the reasoning-intensive BRIGHT benchmark, base LATTICE with a single off-the-shelf LLM achieves 46.7 nDCG@10 - matching the best fine-tuned ensemble baseline overall - and a lightweight ensemble LATTICE++ that fuses LATTICE with cheap retrieval reaches 49.1 nDCG@10. A controlled same-LLM comparison against sliding-window reranking shows reranking offers a better tradeoff at low token budgets, but LATTICE converges to a higher asymptote after a moderate budget. LATTICE also works with open-weight LLMs and remains competitive on traditional IR benchmarks (NQ, SciFact, SciDocs).

翻译：搜索系统越来越多地用于推理密集型查询，此时文档的相关性判定需要理解或推理查询与文档的关系，而非依赖表层词汇或主题相似性。标准方案——基于廉价嵌入的检索器后接LLM验证器——仅在嵌入模型能将正确文档纳入其top-k时有效，而近期推理密集型信息检索基准测试表明，即便最先进的嵌入模型也常无法满足这一假设。近期针对查询端的改进（如查询重写和智能体循环）将LLM置于廉价检索器上游，但仍易受嵌入器失效及LLM依赖参数化知识重写查询能力的制约。本文探索了全新范式——LLM引导的分层搜索——其中LLM通过可分层导航的搜索索引直接与语料库交互，搜索过程中无需嵌入模型参与。我们提出LATTICE框架，其包含两项技术贡献：(i) 基于LLM对多级文档摘要判断的自顶向下索引构建方法；(ii) 通过跨分支参考节点校准路径聚合的LLM引导遍历策略，以缓解存在噪声的、与候选列表相关的LLM评分。在推理密集型BRIGHT基准测试中，使用单一现成LLM的基础LATTICE达到46.7 nDCG@10——匹配最佳微调集成基线整体性能——而融合LATTICE与廉价检索的轻量级集成LATTICE++达到49.1 nDCG@10。与滑动窗口重排序的受控同LLM对比表明，重排序在低标记预算下具有更优的权衡，但LATTICE在适度预算后收敛至更高渐近线。LATTICE还可与开放权重LLM协同工作，并在传统信息检索基准测试（NQ、SciFact、SciDocs）中保持竞争力。