MultAttnAttrib Training-Free Attribution for Long Document Multimodal QA
June 30, 2026
MultAttnAttrib is a training-free method that uses attention heads and calibrated thresholds during the prefill pass to locate source evidence in multimodal documents. It includes MultAttrEval, a new benchmark for fine-grained, ground-truth multimodal attribution.
HOW THIS AFFECTS YOU
●
builderYou can implement evidence attribution in multimodal assistants without the need for additional model training.
●
researcherYou can use MultAttrEval to benchmark the grounding capabilities of multimodal question-answering models.