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

Explain Like I'm Normal

Researchers introduced DocHop to test if Multimodal LLMs can handle complex reasoning that requires jumping between document text and visual charts. Most current models fail when the context for a chart is hidden in the narrative, rather than the graphic itself. This benchmark forces AI to prove it actually understands how textual constraints change data interpretation.