PaperlayerView exact revision on arXiv

Source-linked AI summary

Group-Aware Adaptive Retrieval for Evidence Navigation

June Park, Jun Kwon, Jonghyo Kim, Jongwuk Lee

arXiv:2609.02188v1cs.IR

TL;DR

Reasoning-intensive retrieval faces bounded recall when relevant documents are absent from the initial candidate set and cannot be identified through surface-level matching. GAREN navigates corpus graphs through semantically coherent groups, combining group-level expansion with document-level reranking. Experiments show strong performance, including up to 8.0% improvement over the strongest baseline on BRIGHT.

  • Problem

    Reasoning-intensive retrieval struggles with bounded recall because relevance may require multi-step reasoning and relevant documents may be missing from the initial candidate set.

  • Method

    GAREN partitions corpus graphs into semantically coherent groups, uses a group-level navigator with explore-then-exploit expansion, and combines this with document-level reranking.

  • Results

    GAREN achieves strong performance on BRIGHT and additional retrieval datasets, with up to 8.0% improvement over the strongest baseline on BRIGHT.

  • Takeaways & Limitations

    Group-level guidance provides an alternative to relying only on local document-level signals when selecting corpus-graph expansion directions.

  • Takeaways & Limitations

    GAREN’s community-detection-based grouping and fixed explore-then-exploit transition may be improved by more advanced construction strategies or adaptive scheduling.

Abstract

from arXiv · show

Reasoning-intensive retrieval addresses queries whose relevance cannot be identified by surface-level matching, thereby requiring multi-step reasoning. Because relevant documents rarely appear in the initial candidate set, retrieval systems suffer from the bounded recall problem. Existing methods iteratively expand a candidate pool at the document level over a corpus graph, examining each neighbor in isolation and drifting toward a narrow region of the corpus. To address this problem, we propose Group-Aware Adaptive Retrieval for Evidence Navigation (GAREN), which explores the corpus graph through group-level expansion. GAREN organizes documents into semantically coherent and distinguishable groups based on their connections in the corpus graph. The information in each group indicates what can be accessed by expanding through it, providing guidance beyond individual document-level signals. At each iteration, GAREN uses a group-level navigator to identify promising expansion directions, retrieves documents from the selected groups, and applies a document-level reranker to evaluate the updated candidate set. Extensive experiments show that GAREN achieves up to 8.0% improvement over the strongest baseline on BRIGHT. The source code is available at https://github.com/KJ12124/GAREN

1 Introduction

Reasoning-intensive retrieval suffers from bounded recall because relevant documents may be absent initially and their relevance may emerge only through multi-step reasoning. GAREN addresses this by evaluating group-level expansion directions alongside document-level reranking.

  • Reasoning-intensive queries require relevance judgments beyond lexical or semantic similarity, making relevant documents difficult to retrieve initially.
  • The bounded recall problem persists because relevant documents missing from the initial candidate set are difficult to recover downstream.
  • Existing adaptive retrieval methods expand graph neighbors of documents judged relevant, using current candidate relevance as expansion evidence.
  • Document-level signals can provide little guidance when current documents appear indirectly relevant, allowing early expansion errors to propagate.This limitation is especially critical when the initial candidate set provides weak or unreliable evidence.
  • GAREN treats documents reachable through each expansion direction as a group whose shared context previews the information that direction can provide.
  • GAREN constructs semantic groups, navigates among them with an explore-then-exploit strategy, reranks documents, and propagates evidence from high-confidence candidates.

2 Related Work

Related work addresses difficult retrieval through query reasoning, improved retrievers, adaptive graph traversal, and hierarchical corpus organization. These approaches motivate GAREN’s use of group-level structure for expansion guidance.

  • Reasoning-intensive retrieval research enhances query interpretation with decomposition, reasoning chains, or corpus-feedback query rewriting.
  • Adaptive retrieval dynamically adds candidates beyond initial results through corpus-graph traversal guided by reranking and additional semantic signals.
  • GAREN’s framework is presented as a group-oriented extension of corpus-graph retrieval, with semantic groups summarized for navigation.
  • Recent methods incorporate reasoning feedback, LLM preferences, and richer semantic scoring into adaptive expansion or corpus-graph construction.
  • Hierarchical retrieval organizes corpora into multiple levels, enabling access to both fine-grained evidence and higher-level abstractions.

3 Preliminary

Listwise reranking reorders documents already in a candidate window, while adaptive retrieval expands that window through a corpus graph. GAREN changes the expansion unit from individual documents to groups with previews of reachable information.

  • Listwise reranking orders a candidate window so documents relevant to the query appear higher in the ranking.
  • Reranking cannot retrieve documents absent from the current window, so adaptive retrieval updates the window using a corpus graph of semantically proximate documents.
  • Adaptive retrieval preserves the top h reranked documents and appends w −h graph-expanded candidates to form the next window.
  • Existing expansion derives anchors from current-window evidence, but lacks visibility into the information reachable through each potential direction.
  • GAREN instead expands at the group level, using group representations as coarse previews of information reachable through alternative directions.

4 Proposed Method

GAREN expands the corpus graph through semantically coherent document groups, using group summaries and a navigator to select directions before document-level reranking and evidence propagation. Its explore-then-exploit strategy broadens early exploration and concentrates later expansion on promising groups.

  • 4.2 Group-Aware Adaptive Retrieval: The framework combines a group-level navigator for selecting expansion directions with a document-level reranker for evaluating exposed candidates.The navigator scores candidate directions using group summaries, while the reranker assesses current candidate documents.
  • 4.1 Group Construction: GAREN partitions the corpus graph into semantically coherent groups and generates summaries that preview information accessible through each expansion direction.Groups form a corpus partition, and each summary represents shared semantic context without requiring every group document to be examined individually.
  • 4.2 Group-Aware Adaptive Retrieval: The retrieval loop constructs each next window from reranked documents and newly selected expansion documents.Algorithm 1 initializes a window from the initial ranked results, reranks it iteratively, expands through scored groups, and retains the next window.
  • 4.2 Group-Aware Adaptive Retrieval: GAREN explores multiple promising groups initially, then switches to depth-first expansion focused on the highest-scoring groups after iteration τswitch.Round-robin selection mitigates incorrect early decisions, whereas later exploitation uses accumulated reranking evidence to prioritize groups.
  • 4.3 Group-Driven Evidence Propagation: Group-Driven Evidence Propagation refines final scores by combining document-level evidence with rank-weighted support from selected documents in the same group.Higher-ranked documents receive stronger base weights, and α controls propagation from neighboring selected documents; the final ranking can promote relevant documents in evidence-rich groups.

5 Experimental Setup

GAREN is evaluated on BRIGHT, R2MED, and eight BEIR datasets using nDCG@10 and Recall@100. The study compares it with fixed-pool and graph-based adaptive retrieval baselines, and reports significant BRIGHT gains over the best baseline in both reranking settings.

  • Datasets: The evaluation covers BRIGHT, R2MED, and eight BEIR datasets spanning reasoning-intensive, medical, and diverse retrieval tasks.BRIGHT targets reasoning-intensive retrieval, R2MED evaluates reasoning-driven medical retrieval, and BEIR contributes eight datasets.
  • Datasets: The reported metrics are nDCG@10 and Recall@100, with R@100 measuring recall over documents observed by the reranker.This distinguishes ranking quality at the top 10 from recall within the reranker-observed document set.
  • Baselines: The baselines include fixed-pool Retrieve-and-Rerank, document-neighbor SlideGAR, ANN-graph RGS, and reasoning-guided graph retrieval REPAIR.These baselines represent fixed first-stage reranking and several adaptive graph-based retrieval strategies.
  • Effectiveness on Reasoning-Intensive Retrieval: GAREN significantly outperforms the best-performing baseline on BRIGHT in both reranking settings under a query-level paired t-test with p < 0.05.Table 1 marks the best and second-best results within each reranking prompt setting.

6 Results and Analysis

GAREN improves retrieval by using group-level expansion to navigate beyond local document neighborhoods, with benefits that grow for distant gold documents. Experiments show strong BRIGHT results, effective ablations, and competitive latency.

  • Main Results: GAREN improves average nDCG@10 over Retrieve-and-Rerank by 25.8% in non-reasoning and 25.3% in reasoning settings on BRIGHT.It also outperforms SlideGAR and RGS by 18.9% and 7.2% in the non-reasoning setting.
  • Main Results: GAREN achieves competitive performance on R2MED and BEIR, although gains are more modest because first-stage retrieval can already recover gold documents.The results extend group-level guidance beyond BRIGHT to medical reasoning and traditional retrieval.
  • Document-Level vs. Group-Level Expansion: Document-level expansion improves rapidly but saturates, whereas group-level expansion improves steadily and eventually surpasses it.Group evaluation provides broader contextual signals than scoring immediate neighbors alone.
  • Document-Level vs. Group-Level Expansion: The advantage of group-level expansion grows as gold documents become more distant, partly because the group graph reduces graph distance to those documents.This pattern is measured through reachability and Hit@100 comparisons across graph-hop distances.
  • Ablation Study: Ablations show that explore-then-exploit, group-driven evidence propagation, and semantic navigator scoring each contribute to retrieval quality.Replacing navigator scoring with connectivity reduces nDCG@10 by 25.1% and R@100 by 27.7%.
  • Efficiency: GAREN maintains strong retrieval performance with competitive latency under both prompting settings.The reranker and navigator can operate independently and in parallel, while a lightweight navigator is sufficient for direction assessment.
  • Case Study for Expansion Strategy: The case study shows GAREN exploring diverse groups before concentrating on the most promising direction, successfully retrieving gold documents.Exploitation alone remains near the seed set, while exploration alone fails to concentrate sufficiently.

7 Conclusion

GAREN addresses bounded recall in reasoning-intensive retrieval by combining group-level navigation, document-level reranking, explore-then-exploit expansion, and evidence propagation. It performs particularly well when relevant documents are distant from the initial retrieval pool.

  • Conclusion: GAREN organizes the corpus graph into coarse-grained groups and coordinates group navigation with document reranking to address unreliable initial retrieval signals.An explore-then-exploit strategy and group-driven evidence propagation further support iterative retrieval.

8 Limitations

GAREN’s limitations concern how document groups are constructed and how the system switches from exploration to exploitation. Both design choices work effectively in the experiments but leave room for more adaptive alternatives.

  • Group Construction Strategy: GAREN uses community-detection-based partitioning, while learned partitioning, alternative algorithms, or stronger graph embeddings could produce more informative expansion directions.
  • Fixed Explore-then-Exploit Transition: The fixed τswitch schedule encourages broad exploration early and focused expansion later across the reported experiments.
  • Fixed Explore-then-Exploit Transition: Query-specific characteristics or intermediate retrieval signals could make the exploration-to-exploitation transition more adaptive.

Ethics Statement

The work states that it follows ACL ethical guidelines and uses scientific resources obtained under permissive licenses for their intended research purposes.

  • The work adheres to the ACL’s ethical guidelines.
  • All scientific resources were obtained under permissive licenses and used for their intended research purposes.
  • The corpus graph connects each document to its top 16 most similar documents using Qwen3-Embedding-4B representations.
  • GAREN partitions the corpus graph with Leiden community detection using γ = 1.0, θ = 0.001, and a re-partitioning threshold of ¯s = 20.
  • Group-driven evidence propagation uses α = 0.01, with RBP decay p = 0.6 for non-reasoning reranking and p = 0.5 for reasoning reranking.

B.1 Pipeline Robustness

GAREN remains effective across stronger first-stage retrieval, alternative graph embeddings, different reranker backbones, and varied component scales, while its comparison with LATTICE shows an effectiveness–cost trade-off that depends on setup.

  • First-Stage Retrieval: GAREN remains the best-performing method after replacing BM25 with the stronger ReasonIR first-stage retriever.
  • Corpus Graph Embedding: GAREN continues to outperform all baselines on BRIGHT when the corpus graph uses BGE-Large embeddings under both prompting settings.
  • Scaling the Reranker and the Navigator: Reranker and navigator sizes are varied across three scales in the Qwen3.5 and Gemma4 families, with nDCG@10 and latency measured for every combination.
  • Reranker Backbones: GAREN achieves the strongest performance in most domain categories across GPT-4o-mini and Gemma4-26B-A4B-it reranker backbones and both prompting settings.
  • Comparison with LATTICE: LATTICE organizes the corpus into a semantic tree for top-down traversal, whereas GAREN iteratively navigates group directions from an initial document-retrieval state.
  • Comparison with LATTICE: Under the standard setup, GAREN has slightly lower ranking performance but substantially lower iterative-retrieval cost than LATTICE; with GPT-4 queries and ReasonIR, it is stronger and cheaper.

C.1 Structural Analysis of the Group Unit

Group-level expansion reaches distant gold documents more effectively because the group graph compresses long document-graph distances. Community-detection groups outperform equally sized K-means and random partitions.

  • Reachability by distance: Group-level expansion increasingly outperforms document-level expansion as gold documents lie farther from the initial retrieval set.
  • Reachability by distance: Document-level expansion reaches no gold document at 7 or more hops, while group-level expansion retains meaningful reachability at long distances.
  • Distance to gold documents: Gold documents span wide distances on the document graph but are mostly within 3 hops on the group graph.
  • Distance to gold documents: A single group selection can skip multiple document hops, allowing distant gold documents to be reached within a limited iteration budget.
  • Group construction strategies: Community detection performs best among the tested grouping strategies, while random grouping falls far behind both community detection and K-means.

C.3 Failure Case Analysis

GAREN’s failure analysis identifies errors in group construction and navigator judgment, then evaluates design choices governing exploration, navigator budgets, and retrieval scalability.

  • Failure modes: 53.3% of failures arise from group construction, where summaries omit document-specific facets or graph links follow surface overlap instead of topical similarity.These errors can misrepresent the gold document or place it in an unrelated group.
  • Failure modes: 46.7% of failures arise from navigator judgment despite appropriate grouping and relevant summaries, especially when query-summary relevance requires multi-step reasoning.The navigator may assign a low score even when the group contains the needed information.
  • Improvement directions: Larger navigators improve group selection, but their latency cost outweighs the gains; proposed remedies include topic-aware navigator training and improved grouping.Suggested grouping changes include LLM-extracted document topics and soft clustering across multiple document aspects.
  • Sensitivity analysis: Performance peaks at τswitch = 4: immediate exploitation narrows exploration too early, whereas τswitch = 8 spreads expansion across too many groups.The threshold controls when GAREN switches from exploration to exploitation.
  • Sensitivity analysis: Performance improves with larger Kg, but GAREN adopts Kg = 20 to balance accuracy against computational cost.Further increases provide additional gains while incurring proportionally higher navigator cost.
  • Scalability: GAREN scales to corpora of up to 4.6M documents, with roughly linear offline construction cost and online navigator scoring bounded by Kg per iteration.Group construction is performed offline once per corpus, while online cost remains bounded regardless of corpus size.
Loading 2609.02188v1…