Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces GRAIL, a novel retrieval paradigm for multimodal multi-hop question answering that addresses semantic anchoring by performing context-subtractive query steering at the embedding level. GRAIL adaptively routes queries based on task type (evidence set completion or sequential pool construction), using context subtraction for compositional reasoning and additive embedding updates for localized information aggregation. Experiments on MultimodalQA show a 40.3% macro-averaged performance gain, demonstrating GRAIL's ability to retrieve relevant evidence in a noise-resilient manner and expand the search horizon.
Stop getting stuck in semantic echo chambers: GRAIL's "context subtraction" technique rewrites embeddings on the fly to find missing puzzle pieces in multimodal reasoning.
In multimodal multi-hop question answering, we focus on the initial retrieval stage via two distinct tasks: (1) evidence set completion, retrieving missing evidence given context, and (2) sequential pool construction, iteratively building the top-$K$ pool from the scratch. Under these settings, we point out that conventional iterative retrieval frameworks often suffer from Semantic Anchoring, where previously fetched evidence traps the retriever and yields entity-centric redundancy. To break this trap, we propose GRAIL (Gap-aware Retrieval via Adaptive Implicit Localization), a paradigm that performs implicit query rewriting directly at the embedding level. By context-subtractive query steering, GRAIL excels at compositional cross-modal reasoning, while additive embedding updates show strength on localized information aggregation. By dynamically routing queries based on task type, our Hybrid Framework achieves a 40.3\% macro-averaged performance gain on MultimodalQA. Extensive evaluations demonstrate that sequential GRAIL retrieves in a superior, noise-resilient manner, significantly expanding the search horizon through iterative gap-aware optimization.