A recent arXiv paper introduces a gap-aware retrieval method for multimodal multi-hop question answering. The work focuses on the initial retrieval stage, where the goal is to gather the right evidence before reasoning across multiple modalities and steps.

The authors define two distinct tasks: evidence set completion, which retrieves missing evidence given a partial context, and sequential pool construction, which builds a pool of evidence incrementally. The title's "subtraction" hints at a strategy that identifies what is absent from the current evidence set, rather than only adding more.

This gap-aware approach could help systems avoid missing critical pieces of information in complex QA pipelines. The paper is available on arXiv under identifier 2605.28641.