AI 服務暫時不可用，以下為來源正文，待恢復後補全翻譯。
[Submitted on 30 Sep 2026] Title:Tree Navigation Without LLM Summaries: A Matched-Cost Study of Hierarchical Retrieval for Long-Document QA View a PDF of the paper titled Tree Navigation Without LLM Summaries: A Matched-Cost Study of Hierarchical Retrieval for Long-Document QA, by Priyank Jayraj and 2 other authors View PDF HTML (experimental) Abstract:Retrieval-augmented generation grounds language models in external context, but for long documents flat top-$k$ retrieval can cluster on a single region and miss complementary evidence. RAPTOR-style summary trees address this by recursively clustering chunks and using a language model to summarize each cluster at indexing time, then ranking summary nodes alongside raw chunks at query time. We show the main benefit of summary trees in long-document QA can come from navigation rather than the generated summary content. We introduce NavTree, a leaves-only retriever that builds a deterministic balanced segment tree over chunks (zero language-model calls at indexing) and uses the tree purely as a navigation scaffold: a hybrid lexical-and-dense frontier walk, anchored on top retrieved leaves, descends from the root and emits only leaf chunks to the reader. On a matched-cost evaluation against flat retrievers and an extractive re-implementation of RAPTOR, NavTree is the strongest matched-cost hierarchical retriever in our evaluated grid and ties the strongest flat baseline. On long-document multi-hop QA, it is the only hierarchical method that significantly beats BM25 on a class-vs-class basis, corroborated by a reader-free retrieval-recall check. A matched-reader replication of the published abstractive RAPTOR variant, given strong cluster summaries, still loses to NavTree at every multi-chunk budget, at zero indexing cost. The ranking carries across stronger and open-weight readers, a stronger encoder, and a full factorial that isolates leaves-only emission as the structural lever. Subjects: Computation and Language (cs.CL); Information Retrieval (cs.IR) Cite as: arXiv:2610.06902 [cs.CL] (or arXiv:2610.06902v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2610.06902 arXiv-issued DOI via DataCite Submission history From: Priyank Jayraj [view email] [v1] Wed, 30 Sep 2026 08:42:45 UTC (9,192 KB) Full-text links: Access Paper: View a PDF of the paper titled Tree Navigation Without LLM Summaries: A Matched-Cost Study of Hierarchical Retrieval for Long-Document QA, by Priyank Jayraj and 2 other authors View PDF HTML (experimental) TeX Source view license Ancillary-file links: Ancillary files (details): MANIFEST.txt README.md REPRODUCIBILITY.md code/asrmb/init.py code/asrmb/baselines/init.py code/asrmb/baselines/bm25.py code/asrmb/baselines/dense_flat.py code/asrmb/baselines/hybrid_flat.py code/asrmb/baselines/raptor_baseline.py code/asrmb/benchmarks/init.py code/asrmb/benchmarks/base.py code/asrmb/benchmarks/multihop.py code/asrmb/benchmarks/narrativeqa.py code/asrmb/benchmarks/qasper.py code/asrmb/benchmarks/quality.py code/asrmb/core/init.py code/asrmb/core/inference.py code/asrmb/core/packing.py code/asrmb/core/segtree.py code/asrmb/index/init.py code/asrmb/index/faiss_utils.py code/asrmb/retrieval/init.py code/asrmb/retrieval/ce_rerank.py code/asrmb/retrieval/navigated_fill.py code/requirements.txt code/scripts/bench_index_cost.py code/scripts/bench_query_latency.py code/scripts/build_anon_zip.sh code/scripts/build_paper_figures.py code/scripts/build_paper_tables.py code/scripts/compute_confidence_intervals.py code/scripts/diagnostics/qasper_tail_token_stats.py code/scripts/exp_abstractive_raptor.py code/scripts/exp_alpha_kleaf_sweep.sh code/scripts/exp_factorial_w7_quality2000.sh code/scripts/exp_hybrid_flat_multihop_fill.sh code/scripts/exp_multi_seed.sh code/scripts/exp_multihop.sh code/scripts/exp_queue_remaining.sh code/scripts/exp_summary_vs_leaf.py code/scripts/exp_summary_vs_leaf_extractive.py code/scripts/exp_winning_cell_ablation.sh code/scripts/merge_hybrid_flat_multihop.py code/scripts/paper_status.sh code/scripts/phase1_build_contexts.py code/scripts/phase1_longdoc.py code/scripts/phase2_longdoc.py code/scripts/phase2_run_reader.py code/scripts/plot_results.py code/scripts/post_qwen_qasper_finalize.sh code/scripts/render_appendix_table.py code/scripts/render_factorial_w7.py code/scripts/render_table_multihop_recall.py code/scripts/render_table_openweight.py code/scripts/render_table_qasper_recall.py code/scripts/rescore_nqa_summary_leaf.py code/scripts/resume_arr_experiments.sh code/scripts/run_a1_ollama.sh code/scripts/run_a1_openweight.sh code/scripts/run_a3_bge_large.sh code/scripts/run_ablation_qasper256.sh code/scripts/run_benchmarks.py code/scripts/run_canonical_baseline.sh code/scripts/run_multihop_recall.py code/scripts/run_niah.py code/scripts/run_niah_multi.py code/scripts/run_overnight_orchestrator.sh code/scripts/run_paired_tests.py code/scripts/run_paired_tests_openweight.py code/scripts/run_qasper_recall.py code/scripts/sync_asrmb_paper_tables.py code/scripts/sync_quality_asrmb_tables_from_canon.py responsible_nlp_checklist.md results/abstractive_raptor_gpt4o/quality_checkpoint.json results/abstractive_raptor_gpt4o/quality_predictions.json results/abstractive_raptor_gpt4o/quality_results.json results/abstractive_raptor_gpt4o/summary.json results/bge_large_run/qasper/qasper_checkpoint.json results/bge_large_run/qasper/qasper_results.csv results/bge_large_run/qasper/qasper_results.json results/bge_large_run/quality/quality_checkpoint.json results/bge_large_run/quality/quality_results.csv results/bge_large_run/quality/quality_results.json results/canonical_v1/narrativeqa/narrativeqa_checkpoint.json results/canonical_v1/narrativeqa/narrativeqa_results.csv results/canonical_v1/narrativeqa/narrativeqa_results.json results/canonical_v1/qasper/qasper_checkpoint.json results/canonical_v1/qasper/qasper_results.csv results/canonical_v1/qasper/qasper_results.json results/canonical_v1/quality/quality_checkpoint.json results/canonical_v1/quality/quality_results.csv results/canonical_v1/quality/quality_results.json results/factorial_w7_quality2000/A0_B0_C0_D0/quality_checkpoint.json results/factorial_w7_quality2000/A0_B0_C0_D0/quality_results.csv results/factorial_w7_quality2000/A0_B0_C0_D0/quality_results.json results/factorial_w7_quality2000/A0_B0_C0_D1/quality_checkpoint.json results/factorial_w7_quality2000/A0_B0_C0_D1/quality_results.csv results/factorial_w7_quality2000/A0_B0_C0_D1/quality_results.json results/factorial_w7_quality2000/A0_B0_C1_D0/quality_checkpoint.json results/factorial_w7_quality2000/A0_B0_C1_D0/quality_results.csv results/factorial_w7_quality2000/A0_B0_C1_D0/quality_results.json results/factorial_w7_quality2000/A0_B0_C1_D1/quality_checkpoint.json results/factorial_w7_quality2000/A0_B0_C1_D1/quality_results.csv results/factorial_w7_quality2000/A0_B0_C1_D1/quality_results.json results/factorial_w7_quality2000/A0_B1_C0_D0/quality_checkpoint.json results/factorial_w7_quality2000/A0_B1_C0_D0/quality_results.csv results/factorial_w7_quality2000/A0_B1_C0_D0/quality_results.json results/factorial_w7_quality2000/A0_B1_C0_D1/quality_checkpoint.json results/factorial_w7_quality2000/A0_B1_C0_D1/quality_results.csv results/factorial_w7_quality2000/A0_B1_C0_D1/quality_results.json results/factorial_w7_quality2000/A0_B1_C1_D0/quality_checkpoint.json results/factorial_w7_quality2000/A0_B1_C1_D0/quality_results.csv results/factorial_w7_quality2000/A0_B1_C1_D0/quality_results.json results/factorial_w7_quality2000/A0_B1_C1_D1/quality_checkpoint.json results/factorial_w7_quality2000/A0_B1_C1_D1/quality_results.csv results/factorial_w7_quality2000/A0_B1_C1_D1/quality_results.json results/factorial_w7_quality2000/A1_B0_C0_D0/quality_checkpoint.json results/factorial_w7_quality2000/A1_B0_C0_D0/quality_results.csv results/factorial_w7_quality2000/A1_B0_C0_D0/quality_results.json results/factorial_w7_quality2000/A1_B0_C0_D1/quality_checkpoint.json results/factorial_w7_quality2000/A1_B0_C0_D1/quality_results.csv results/factorial_w7_quality2000/A1_B0_C0_D1/quality_results.json results/factorial_w7_quality2000/A1_B0_C1_D0/quality_checkpoint.json results/factorial_w7_quality2000/A1_B0_C1_D0/quality_results.csv results/factorial_w7_quality2000/A1_B0_C1_D0/quality_results.json results/factorial_w7_quality2000/A1_B0_C1_D1/quality_checkpoint.json results/factorial_w7_quality2000/A1_B0_C1_D1/quality_results.csv results/factorial_w7_quality2000/A1_B0_C1_D1/quality_results.json results/factorial_w7_quality2000/A1_B1_C0_D0/quality_checkpoint.json results/factorial_w7_quality2000/A1_B1_C0_D0/quality_results.csv results/factorial_w7_quality2000/A1_B1_C0_D0/quality_results.json results/factorial_w7_quality2000/A1_B1_C0_D1/quality_checkpoint.json results/factorial_w7_quality2000/A1_B1_C0_D1/quality_results.csv results/factorial_w7_quality2000/A1_B1_C0_D1/quality_results.json results/factorial_w7_quality2000/A1_B1_C1_D0/quality_checkpoint.json results/factorial_w7_quality2000/A1_B1_C1_D0/quality_results.csv results/factorial_w7_quality2000/A1_B1_C1_D0/quality_results.json results/factorial_w7_quality2000/A1_B1_C1_D1/quality_checkpoint.json results/factorial_w7_quality2000/A1_B1_C1_D1/quality_results.csv results/factorial_w7_quality2000/A1_B1_C1_D1/quality_results.json results/factorial_w7_quality2000/_summary.json results/final_archive/README.md results/final_archive/longdoc_ci.json results/final_archive/longdoc_results.json results/final_archive/longdoc_results.json.checkpoint results/final_archive/narrativeqa_checkpoint.json results/final_archive/narrativeqa_results.json results/final_archive/qasper_checkpoint.json results/final_archive/qasper_ci.json results/final_archive/qasper_results.json results/final_archive/qasper_unifiedqa_results.json results/final_archive/quality_ci.json results/final_archive/quality_navigated_checkpoint.json results/final_archive/quality_results.json results/final_archive/quality_unifiedqa_results.json results/final_archive/quality_unifiedqa_results.json.checkpoint results/frozen_opt/AUDIT.md results/frozen_opt/ablation_qasper256/_meta/manifest.json results/frozen_opt/ablation_qasper256/row1_classic_steiner/qasper_checkpoint.json results/frozen_opt/ablation_qasper256/row1_classic_steiner/qasper_ci.json results/frozen_opt/ablation_qasper256/row1_classic_steiner/qasper_results.csv results/frozen_opt/ablation_qasper256/row1_classic_steiner/qasper_results.json results/frozen_opt/ablation_qasper256/row2_kleaf48/qasper_checkpoint.json results/frozen_opt/ablation_qasper256/row2_kleaf48/qasper_ci.json results/frozen_opt/ablation_qasper256/row2_kleaf48/qasper_results.csv results/frozen_opt/ablation_qasper256/row2_kleaf48/qasper_results.json results/frozen_opt/ablation_qasper256/row3_hybrid_alpha035/qasper_checkpoint.json results/frozen_opt/ablation_qasper256/row3_hybrid_alpha035/qasper_ci.json results/frozen_opt/ablation_qasper256/row3_hybrid_alpha035/qasper_results.csv results/frozen_opt/ablation_qasper256/row3_hybrid_alpha035/qasper_results.json results/frozen_opt/ablation_qasper256/row4_alpha050/qasper_checkpoint.json results/frozen_opt/ablation_qasper256/row4_alpha050/qasper_ci.json results/frozen_opt/ablation_qasper256/row4_alpha050/qasper_results.csv results/frozen_opt/ablation_qasper256/row4_alpha050/qasper_results.json results/frozen_opt/ablation_qasper256/row5_tight_tail/qasper_checkpoint.json results/frozen_opt/ablation_qasper256/row5_tight_tail/qasper_ci.json results/frozen_opt/ablation_qasper256/row5_tight_tail/qasper_results.csv results/frozen_opt/ablation_qasper256/row5_tight_tail/qasper_results.json results/frozen_opt/ablation_qasper256/summary.md results/frozen_opt/ablation_quality2000/row1_classic_steiner/quality_checkpoint.json r [truncated for AI cost control]