다단계 질의응답에서 적응형 검색을 위한 동적 게이트 기반 질의 정제 = Dynamic Gate-Based Query Refinement for Adaptive Retrieval in Multi-Hop Question Answering

다국어 초록 (Multilingual Abstract)

Multi-hop Question Answering (MHQA) necessitates that Large Language Models (LLMs) synthesize scattered information through iterative retrieval and reasoning. However, current iterative frameworks face structural bottlenecks: the ``context overflow'' arising from the accumulation of intermediate artifacts, and the ``error propagation'' where initial retrieval inaccuracies cascade into downstream reasoning failures. These challenges often destabilize the reasoning trajectory, leading to suboptimal final answers. To address these limitations, this study presents a lightweight, training-free adaptive pipeline designed to enhance the robustness of iterative MHQA. We first integrate the RECAP mechanism---inspired by cognitive chunking---to summarize reasoning history at each step, thereby stabilizing discourse coherence and mitigating context overflow. Building upon this stabilized baseline, we introduce a ``Similarity-Gated Query Refinement'' module. This component functions as a dynamic monitor, assessing retrieval quality via similarity score distributions and autonomously triggering query rectification only when high uncertainty is detected. This selective intervention strategy ensures corrective re-retrieval without incurring unnecessary computational overhead. Empirical evaluations on three standard benchmarks---HotpotQA, 2WikiMultihopQA, and MuSiQue---demonstrate that our adaptive pipeline consistently outperforms a robust RECAP-based baseline across diverse settings. Crucially, significant performance gains were observed not only in proprietary API-based models but also in offline environments using Llama-3.1-8B, substantiating the model-agnostic nature and generalizability of the proposed framework. While acknowledging limitations such as reliance on heuristic gating, this research underscores the efficacy of selective intervention based on lightweight uncertainty signals. By balancing computational efficiency with rigorous error correction, the proposed framework offers a practical foundation for designing robust, controllable agentic workflows in complex reasoning tasks.

목차 (Table of Contents)

• 1 Introduction 1
• 2 Related Work 4
• 2.1 Multi-hop Question Answering and Benchmark Datasets 4
• 2.2 Retrieval-Augmented Generation and Iterative QA 5
• 2.3 Adaptive Reasoning and Search-based Approaches 6
• 2.4 Adaptive Retrieval and Self-Correction Mechanisms 6
• 3 Proposed Method 8
• 3.1 Overview 8
• 3.2 Baseline: Iterative QA with RECAP 10
• 3.3 Similarity-Gated Query Refinement 12
• 4 Experiments 15
• 4.1 Experimental Setup 15
• 4.1.1 Datasets 15
• 4.1.2 Evaluation Metrics 16
• 4.1.3 Models 17
• 4.1.4 Retriever and Similarity Computation 17
• 4.1.5 Baselines and Compared Methods 18
• 4.2 Main Results 18
• 4.2.1 Baseline Validation 18
• 4.2.2 Main Comparison Results 19
• 4.3 Efficiency Analysis 20
• 4.4 Analysis of Gating Strategy and Parameter Sensitivity 21
• 4.4.1 Effect of Conditional Gating 22
• 4.4.2 Sensitivity Analysis of the Gating Threshold 23
• 4.5 Additional Analysis and Implementation Details 24
• 4.5.1 Prompt Engineering and Operational Strategy 24
• 4.5.2 Qualitative Case Studies 25
• 4.5.3 Discussion 27
• 5 Conclusion 28