BRIDGE: Benchmark for multi-hop Reasoning In long multimodal Documents with Grounded Evidence
This benchmark addresses the problem of evaluating intermediate reasoning and evidence grounding in multi-hop QA for LLMs, particularly in long multimodal documents, which is a critical gap for researchers developing advanced reasoning systems.
This paper introduces BRIDGE, a benchmark for multi-hop reasoning over long scientific papers that integrates evidence from text, tables, and figures. It provides explicit multi-hop reasoning annotations for step-level evaluation, revealing systematic deficiencies in evidence aggregation and grounding in state-of-the-art LLMs and multimodal RAG systems.
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.