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

TL;DR

BRIDGE is a new benchmark designed to evaluate multi-hop reasoning in long multimodal scientific documents, emphasizing intermediate reasoning steps and evidence grounding beyond just final answers.

Contribution

It introduces a comprehensive dataset with explicit reasoning annotations for multi-hop questions across text, tables, and figures in scientific papers.

Findings

State-of-the-art models show deficiencies in evidence aggregation.

Traditional evaluation overlooks reasoning process failures.

BRIDGE enables targeted diagnosis of reasoning capabilities.

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.