Source-linked AI summary
Beyond One-Shot Expansion: Contrastive Evidence Exploration for Multi-Hop Retrieval
JungMin Yun, YoungBin Kim
TL;DR
Multi-hop QA retrieval must find linked evidence as reasoning unfolds, yet existing single-intent and one-shot approaches adapt poorly to newly retrieved evidence. The paper proposes a training-free framework combining evidence-conditioned exploration, contrastive refinement, and coverage-aware selection, and reports consistent gains across three multi-hop QA datasets. Its main costs are additional offline facet construction and repeated inference-time LLM calls.
Problem
Multi-hop retrieval must uncover complementary passages linked through intermediate entities and relations, while existing methods often rely on a single intent and static expansion.
Method
The training-free framework iteratively retrieves evidence-conditioned probes, refines candidates with passage-specific contrastive facets, and ranks complementary passages using coverage.
Results
The framework consistently improves retrieval quality and downstream QA performance across MuSiQue, HotpotQA, and 2WikiMultihopQA.
Takeaways & Limitations
Adaptive exploration, contrastive refinement, and coverage-aware selection together provide a retrieval strategy for progressively recovering and ranking multi-hop evidence.
Takeaways & Limitations
Contrastive facet construction and sequential exploration increase offline processing and inference latency, especially for large or frequently updated corpora.
Abstract
from arXiv · showhide
Retrieval-augmented generation (RAG) critically depends on retrieving the evidence necessary for effective reasoning. However, this remains particularly challenging in multi-hop question answering (QA), where supporting passages are often linked through intermediate entities and relations that must be progressively uncovered. Existing retrieval approaches typically rely on a single retrieval intent or one-shot query expansion, limiting their ability to adapt to newly retrieved evidence and potentially introducing noisy or redundant retrieval signals. To address these limitations, we propose a training-free multi-hop retrieval framework that integrates evidence-conditioned exploration, passage-specific contrastive refinement, and coverage-aware final ranking. During offline indexing, the framework constructs passage-specific contrastive facets that characterize each passage relative to its semantically similar neighbors, providing fine-grained signals to distinguish closely related candidates. At inference time, the framework iteratively retrieves evidence, generates probes targeting unresolved information needs, refines candidate relevance using the contrastive facets, and selects a complementary set of passages that collectively cover diverse evidence-seeking intents. Experiments on MuSiQue, HotpotQA, and 2WikiMultihopQA demonstrate consistent improvements in retrieval quality and downstream QA performance over baselines.
1 Introduction
Multi-hop retrieval must uncover linked evidence through evolving entities and relations, but one-shot or single-intent methods can miss complementary passages. The proposed training-free framework combines adaptive exploration, passage-specific contrastive refinement, and coverage-aware selection to improve retrieval and downstream QA.
- Multi-hop QA requires connecting passages through intermediate entities, relations, or reasoning steps that the original question may not state explicitly.
- Single-intent retrieval and static one-shot expansion can overlook evidence needed later and cannot adequately adapt to newly uncovered information.
- Iterative retrieval can still use noisy or weakly grounded signals, while isolated ranking may return redundant passages instead of covering distinct evidence-seeking intents.
- The framework progressively explores evidence, distinguishes semantically similar passages through contrastive facets, and selects passages that collectively cover diverse retrieval intents.
- Experiments on MuSiQue, HotpotQA, and 2WikiMultihopQA show consistent improvements in retrieval quality and downstream QA performance.
2 Methodology
The framework combines offline passage-specific contrastive indexing with online evidence-conditioned multi-hop retrieval. It refines candidates using contrastive signals and selects passages that balance relevance with complementary evidence coverage.
- Problem Formulation: The framework defines multi-hop retrieval as returning passages that collectively provide evidence for different stages of a reasoning chain.
- Pipeline Overview: The pipeline has two stages: offline contrastive facet indexing and online multi-hop retrieval.
- Contrastive Facet Construction & Indexing: Offline indexing creates natural-language facets that distinguish each passage from semantically similar neighbors and stores them for inference.
- Sequential Evidence Exploration: Online exploration iteratively generates probes from retrieved evidence to target information that remains unresolved.
- Contrastive Evidence Refinement: Candidate relevance combines strongest-probe similarity with passage-specific contrastive refinement while preserving the dense retrieval signal.
- Coverage-Aware Final Evidence Selection: Coverage-aware greedy ranking favors candidates that add support for probes not already well covered, producing complementary evidence.
3 Experiments
Experiments across three multi-hop QA datasets show that the framework improves retrieval quality and downstream QA performance, with gains surviving ablation and limited retrieval budgets.
- Experimental Setup: The evaluation covers MuSiQue, HotpotQA, and 2WikiMultihopQA against sparse, dense, reranking, query-expansion, and iterative-retrieval baselines.Baselines include BM25, e5-large-v2, bge-reranker, HyDE, Query2Doc, LameR, IRCoT, and Self-Ask.
- Main Results: Across all datasets and metrics, the method achieves the best retrieval performance, improving both evidence coverage and ranking quality.Improvements span R@5 and FSR@10 as well as MAP@10 and nDCG@10.
- Main Results: 57.00 FSR@10 on MuSiQue exceeds the best baseline score of 39.20, while MAP@10 rises from 54.63 to 63.31 and nDCG@10 from 64.66 to 72.67.On 2WikiMultihopQA, FSR@10 also increases from 75.60 to 91.50.
- Main Results: 96.20 FSR@10 on HotpotQA exceeds 93.30, despite strong baseline performance and smaller gains than on MuSiQue and 2WikiMultihopQA.The method remains highest across all reported HotpotQA metrics.
- Ablation Study: Removing sequential evidence exploration lowers R@10 from 80.41 to 68.75, nDCG@10 from 72.67 to 65.98, and FSR@10 from 57.00 to 36.50.The largest decrease is in FSR@10, indicating the importance of subsequent evidence-conditioned retrieval rounds for complete support recovery.
- Ablation Study: Coverage-aware ranking and contrastive refinement provide smaller complementary gains, with their removal reducing FSR@10 from 57.00 to 54.60 and 55.60, respectively.Coverage-aware ranking also lowers nDCG@10 from 72.67 to 71.12, while removing contrastive refinement lowers it to 72.38.
- Downstream QA Performance: Using only the top-5 retrieved passages, the method achieves the highest EM and F1 across all three datasets and improves downstream QA performance.On MuSiQue, it reaches 24.0 EM and 36.61 F1 versus best-baseline scores of 20.5 EM and 31.67 F1.
- Downstream QA Performance: Gains are consistent across all three datasets, largest on MuSiQue, while answer quality also depends on the reader’s ability to integrate and reason over retrieved passages.The reported QA evaluation uses the top-5 retrieved passages.
4 Conclusion
The paper concludes that its training-free framework improves multi-hop retrieval and downstream QA by combining sequential evidence exploration, contrastive refinement, and coverage-aware selection. These gains come with computational costs from offline facet construction and repeated inference-time LLM calls.
- Conclusion: The framework integrates evidence-conditioned sequential exploration, passage-specific contrastive refinement, and coverage-aware final evidence selection.It progressively adapts retrieval to newly retrieved evidence and selects complementary passages that jointly support reasoning.
- Conclusion: Experiments on three multi-hop QA datasets demonstrate consistent improvements in retrieval quality and downstream QA performance.The evaluated datasets are MuSiQue, HotpotQA, and 2WikiMultihopQA.
- Conclusion: Sequential exploration contributes most to recovering complete supporting evidence, while contrastive refinement and coverage-aware ranking provide additional improvements.This result comes from the reported ablation findings.
- Limitations: The method incurs computational costs from contrastive facet construction and sequential exploration.Facet construction uses LLM generation for target–neighbor pairs, while sequential exploration requires repeated calls for evidence summarization and probe generation.
- Limitations: These costs may increase offline processing and inference latency for large or frequently updated corpora.The limitation is especially relevant when corpus scale or update frequency increases.
GenAI Usage Disclosure
The manuscript used generative AI for grammatical correction and language editing, and used specified instruction-tuned models within the retrieval and downstream QA experiments.
- GenAI Usage Disclosure: Generative AI was used solely for grammatical correction and language editing during manuscript preparation.The disclosure distinguishes manuscript editing from experimental model use.
- GenAI Usage Disclosure: Qwen2.5-7B-Instruct generated contrastive facet queries, evidence-seeking probes, and evidence summaries in the proposed retrieval framework.The model was used within the retrieval framework for these generation steps.
- GenAI Usage Disclosure: Llama-3.1-8B-Instruct served as the reader model for downstream QA evaluation.Its disclosed role was downstream answer evaluation.