STEC Evidence Compression for Deep Search in Open-domain Multi-Hop QA

Jul 16, 2026

Open-domain multi-hop question answering with LLM-based search agents combines retrieval and reasoning but generates multiple search trajectories whose evidence is heterogeneous, redundant, incomplete, or conflicting, making reliable final answer selection difficult. Directly comparing raw trajectories exposes the verifier to noisy and unaligned content, while comparing answer strings ignores evidence quality. STEC proposes an evidence compression framework that groups trajectories by normalized answer identity, builds answer-specific evidence representations, and performs evidence-guided answer verification to select the final candidate.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

Introduction

Open-domain multi-hop QA and retrieval limits

Deep search as interleaved search and reasoning

Final answer selection challenge

arXiv :2607 . 10795v 1 [ cs .AI] 12 Jul 2026 STEC: Evidence Compression for Deep Search in Open-domain Multi-Hop QA Xinkang Li 1 , Rong Jiang 1 (􀀌), Xin Song 1 , Ye Wang2 , Yue Han 1 , and Changjian Li 1 1 National University of Defense Technology, China [xinkangli@nudt.edu.cn](xinkangli@nudt.edu.cn) , [jiangrong@nudt.edu.cn](jiangrong@nudt.edu.cn) , [songxin@nudt.edu.cn](songxin@nudt.edu.cn) , [hanyue@nudt.edu.cn](hanyue@nudt.edu.cn) , [lichangjian23@nudt.edu.cn](lichangjian23@nudt.edu.cn) 2 Harbin Institute of Technology, Shenzhen, China [wangye2020@hit.edu.cn](wangye2020@hit.edu.cn) Abstract. In open-domain multi-hop question answering (QA), LLMbased search agents oﬀer a promising approach to knowledge-intensive QA by combining retrieval with reasoning. Existing methods mainly improve open-domain multi-hop QA through reasoning paradigms, retrieval interaction, and search strategy optimization. However, using multiple search trajectories introduces a challenging ﬁnal answer selection problem. Diﬀerent trajectories may support diﬀerent candidates, and the retrieved information can be heterogeneous, redundant, incomplete, or conﬂicting. Directly comparing raw trajectories exposes the veriﬁer to noisy and unaligned content, while comparing answer strings ignores the evidence supporting each candidate, making reliable ﬁnal selection diﬃcult. To address this challenge, we propose STEC, an evidence compression framework for ﬁnal answer selection in multi-hop QA. STEC selects the ﬁnal answer from the existing candidate set through two mechanisms: (1) Answer-Level Evidence Compression, which groups trajectories by normalized answer identity and converts each answer group into a candidatespeciﬁc evidence representation; and (2) Evidence-Guided Answer Veriﬁcation, which compares these representations and selects the ﬁnal answer from the candidate set. The design shifts ﬁnal selection from raw trajectory comparison to candidate-level evidence comparison. We evaluate STEC on four open-domain multi-hop QA benchmarks against representative baselines. Experimental results show that STEC performs best overall among the compared methods, and ablation results provide evidence that answer-level evidence compression contributes to ﬁnal answer selection. Keywords: Deep Search · Multi-Hop Question Answering · Search Trajectory · Evidence Compression · Answer Veriﬁcation 1 Introduction Open-domain multi-hop question answering is a challenging knowledge-intensive task. A system needs to retrieve relevant information from large-scale open corpora and integrate it across entities, relations, and events to support coherent 2 X. Li et al. 7[KYZOUT | 2GXMK􀀓YIGRK 5VKT )UXV[Y | | | | | --- | --- | --- | --- | | | | | | | | | Ș | | 'TY]KX Fig. 1. Illustration of open-domain multi-hop question answering. A system retrieves relevant information from a large-scale open corpus and reasons across multiple hops to derive the ﬁnal answer. multi-step reasoning. Unlike single-hop question answering, a multi-hop question usually cannot be answered with a single retrieved passage. The model must combine information from multiple sources through reliable multi-step reasoning, and retrieval omissions, information noise, or intermediate reasoning errors can all lead to an incorrect ﬁnal answer. Fig. 1 illustrates this task setting. Traditional retrieval-augmented generation provides a basic paradigm for knowledge-intensive question answering. The model ﬁrst retrieves external context and then generates an answer based on the retrieved results, which mitigates knowledge gaps and factual errors that arise when LLMs rely only on parametric knowledge [1] . However, open-domain multi-hop question answering often involves entities, relations, and constraints that emerge gradually during reasoning. As reasoning proceeds, the model may need to resolve new intermediate entities, compare diﬀerent facts, or verify conditions introduc

FAQs

What problem does STEC address in open-domain multi-hop QA?

STEC targets the final answer selection difficulty caused by multiple deep-search trajectories producing conflicting or incomplete evidence for different candidates.

How does STEC compress evidence for final answer selection?

STEC groups trajectories by normalized answer identity and converts each answer group into a candidate-specific evidence representation.

How does STEC choose the final answer among candidates?

STEC uses evidence-guided answer verification by comparing the compressed evidence representations and selecting the final answer from the candidate set.

Jul 16, 2026

Upload by River Wang