Omni-modal language models can answer questions about multimodal documents, recordings, or videos, but they cannot show which part of the input the answer came from. When asked to cite its evidence, the model produces something, but that output may not be grounded in the actual source material.
OmniTrace is introduced as a unified framework for generation-time source attribution. It aims to provide a systematic way to attribute an answer to specific parts of the input across different modalities, addressing a key gap in transparency for these models.
The significance of this work lies in improving verifiability for omni-modal LLMs, which are increasingly used in tasks that combine text, audio, and visual information. The abstract does not detail the internal mechanisms, but the stated goal is clear: make source attribution a built-in part of generation rather than an afterthought.