Multimodal Fact-Level Attribution for Verifiable Reasoning

Multimodal large language models (MLLMs) are increasingly used for real-world tasks involving multi-step reasoning and long-form generation, where reliability requires grounding model outputs in heterogeneous input sources and verifying individual factual claims. However, existing multimodal grounding benchmarks and evaluation methods focus on simplified, observation-based scenarios or limited modalities and fail to assess attribution in complex multimodal reasoning. We introduce MuRGAt (Multimodal Reasoning with Grounded Attribution), a benchmark for evaluating fact-level multimodal attribution in settings that require reasoning beyond direct observation. Given inputs spanning video, audio, and other modalities, MuRGAt requires models to generate answers with explicit reasoning and precise citations, where each citation specifies both modality and temporal segments. To enable reliable assessment, we introduce an automatic evaluation framework that strongly correlates with human judgments. Benchmarking with human and automated scores reveals that even strong MLLMs frequently hallucinate citations despite correct reasoning. Moreover, we observe a key trade-off: increasing reasoning depth or enforcing structured grounding often degrades accuracy, highlighting a significant gap between internal reasoning and verifiable attribution.

AI systems can now watch videos and listen to audio to answer complex questions, but a serious problem remains: when they explain their answers, do they actually point to the right evidence? An AI might give the correct answer while citing the wrong moment in a video—or inventing evidence that isn't there. Until now, there was no good way to measure this, because existing tests only checked simple cases like "how many flags are in this picture," not questions that require genuine reasoning across video, audio, and graphs. We built MuRGAT, a benchmark that asks AI models to answer hard questions and pinpoint exactly which moment and which source (the audio, the visuals) supports each claim they make. We also created an automatic scoring tool that matches human judgment closely, so this can be measured at scale. Surprisingly, we found that even the strongest AI models frequently "hallucinate" their citations—getting the answer right while grounding it in the wrong evidence. We also discovered that pushing a model to reason harder or cite more carefully can actually make it less accurate. Our work gives researchers the tools to build AI that is both correct and genuinely trustworthy.