BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence

Biao Xiang; Soyeon Caren Han; Yihao Ding

cs.CLMar 9, 2026

BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence

Authors: Biao Xiang, Soyeon Caren Han, Yihao Ding

Organizations: The University of Melbourne Melbourne, Australia · The University of Western Australia Perth, Australia

Abstract

Multi-hop question answering (QA) is widely used to evaluate the reasoning capabilities of large language models, yet most benchmarks focus on final answer correctness and overlook intermediate reasoning, especially in long multimodal documents. We introduce BRIDGE, a benchmark for multi-hop reasoning over long scientific papers that require integrating evidence across text, tables, and figures. The dataset supports both chain-like and fan-out structures and provides explicit multi-hop reasoning annotations for step-level evaluation beyond answer accuracy. Experiments with state-of-the-art LLMs and multimodal retrieval-augmented generation (RAG) systems reveal systematic deficiencies in evidence aggregation and grounding that remain hidden under conventional answer-only evaluation. BRIDGE provides a targeted testbed for diagnosing reasoning failures in long multimodal documents.

Figures & tables

Dataset	MH	Type	Ann.	Evl.	MM	Doc. Scope	Scale
MultiTabQA Pal et al.,2023	$\checkmark$	Cp	$\times$	$\times$	$\times$	Relational DB	9k
MEQA Li et al.,2024	$\checkmark$	Cp+Re	$\times$	$\times$	$\times$	Multi-doc text	2,243
SPIQA Pramanick et al.,2025	$\times$	—	$\times$	$\times$	$\checkmark$	Sci. papers	270k
DBQR-QA Nararatwong et al.,2024	$\checkmark$	Cp+Re	$\times$	$\times$	$\checkmark$	Finance reports	400
JEMHopQA Ishii et al.,2024	$\checkmark$	Cp+Re	$\checkmark$	$\times$	$\times$	Wikipedia (JP)	1179
Complex-TR Tan et al.,2024	$\checkmark$	Cp+Re	$\times$	$\times$	$\times$	Temporal facts	60k

Table 1. Comparison with existing multi-hop QA benchmarks. MH: Multi-hop; Type: Question types (Ab: Abstractive; Cp: Comparative; Re: Causal Reasoning); Ann.: Explicit multi-hop annotation; Evl.: Step-by-step multi-hop evaluation; MM: Multimodal evidence.

Figure 1. Representative examples of comparative (Cp), abstractive (Ab), and causal reasoning (Re) questions (top), and the corresponding pages where evidences locate (bottom). Mod.: involved modalities (T: text; Tb: table; F: figure).

Figure 2. Distribution of QA instances by hop depth, number of distinct pages involved, and hop pattern, broken down by question type (Abstractive, Causal, Comparative)

Page Bin	Model	Best Strat.	Audit	Acc	Fid	N
1–2	ChatGPT	CoT	4.692	4.867	4.881	3338
1–2	Qwen	CoT	3.731	3.844	3.706	2417
3–5	ChatGPT	CoT	4.417	4.538	4.632	2739
3–5	Qwen	CoT	3.552	3.617	3.472	1659
6–10	ChatGPT	CoT	4.240	4.319	4.455	2636
6–10	Qwen	CoT	3.518	3.559	3.430	1333

Table 5. Performance across page-depth bins, grouped by the furthest evidence page involved (best strategy per model by audit_score).

Hop Bin	Model	Best Strat.	Audit	Acc	Fid	N
1-hop	ChatGPT	CoT	3.692	3.656	3.966	212
	Gemma	Direct	2.952	3.000	3.317	25
	Gemini	Direct	2.888	2.978	3.245	180
	Qwen	Reflection	2.649	2.729	2.602	59
2-hop	ChatGPT	CoT	4.425	4.548	4.641	3344
	Gemma	CoT	4.200	4.454	4.408	817

Table 6. Performance by hop depth (1-hop, 2-hop, 3+-hop).

Evidence	Model	Audit	Acc	Fid	N
Text	ChatGPT	4.629	4.812	4.816	6933
	Gemma	4.384	4.699	4.643	1859
	Gemini	4.250	4.447	4.564	5643
	Qwen	3.496	3.557	3.437	5145
Table	ChatGPT	4.096	4.180	4.323	11217
	Gemma	3.816	3.955	3.950	1794

Table 7. Performance by primary evidence modality (Text, Table, Figure), aggregated over prompting strategies. N denotes the number of instances whose dominant supporting evidence belongs to each modality.

Explore similar work

DocHop-QA: Towards Multi-Hop Reasoning over Multimodal Document Collections
Aug 20, 2025Jiwon Park, Seohyun Pyeon, Jinwoo Kim +4Multi-Hop QA Scientific QA
Pohang University of Science and Technology, Pohang-si, South Korea · The University of Melbourne, Melbourne, Australia · The University of Western Australia, Perth, Australia
DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents
Sep 2, 2026Zhuoran Yu, Le Thien Phuc Nguyen, Jaden Park +5Multimodal Large Language Models Chart QA
University of Wisconsin-Madison · Massachusetts Institute of Technology · Stanford University +2
DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding
May 9, 2026Xiang Feng, Jiawei Zhou, Zhangfeng Huang +6LLM Grounding LLM Answer Verification
School of Computer Science, National Engineering Research Center for Multimedia Software and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, China · Alibaba Group, Hangzhou, China · Independent Researcher +1

Aug 20, 2025J/K move · Enter open · S save

DocHop-QA: Towards Multi-Hop Reasoning over Multimodal Document Collections

Jiwon Park, Seohyun Pyeon, Jinwoo Kim, Rina Carines Cabral+3

Pohang University of Science and Technology, Pohang-si, South Korea · The University of Melbourne, Melbourne, Australia · The University of Western Australia, Perth, Australia

Multi-Hop QA Scientific QA Multimodal QA Document QA

Open paper AskPDF arXiv

Despite rapid progress in large language models (LLMs), current QA benchmarks still overlook the core challenge of real-world scientific information seeking: synthesizing multimodal evidence scattered across multiple documents and structural formats. Existing QAs remain narrow in scope, relying on unimodal text and short-span reasoning that fail to capture the complexity of real information-seeking. We introduce DocHop-QA, a benchmark of 11,379 instances for evaluating multimodal, multi-document, multi-hop scientific QA. Built from publicly available PubMed articles, DocHop-QA incorporates textual passages, tables, and layout cues, enabling cross-document inference without explicit hyperlinks. To scale realistic QA construction, we develop an LLM-driven generation pipeline grounded in 11 scientific reasoning concepts, producing diverse and coherent question-answer pairs. To highlight the utility and versatility of the dataset, we propose a task-driven evaluation framework spanning four settings, including generative answering, multimodal evidence integration and structured index prediction. Experiments show that current models struggle with DocHop-QA's long-context, multi-evidence demands, establishing it as a rigorous testbed for advancing next-generation scientific QA systems.

Abstract

Explore similar work

DocHop-QA: Towards Multi-Hop Reasoning over Multimodal Document Collections

DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents

DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding