On October 24, 2026, a new study was published on the arXiv platform (arXiv:2610.30841v1), introducing TRIAD—the first benchmark specifically designed to evaluate cross-modal (text–image–audio) disambiguation capabilities. Its core objective is to assess whether multimodal large language models can, like humans, resolve semantic ambiguities that remain unresolved in a single modality by leveraging complementary cross-modal cues.
TRIAD comprises 327 hand-crafted tri-modal samples, each consisting of a text passage, an image, and an audio clip, accompanied by a question and several answer options. Its construction principle is that the correct answer can be uniquely determined only when all three modalities are present; removing any one modality results in multiple answer options being compatible with the remaining information. This benchmark covers 86 scene categories and 18 types of cross-modal ambiguity.
An evaluation of 14 multimodal systems reveals that, even when performing strongly at the single‑sample level, the models still lag significantly behind humans in disambiguation at the scene‑group level—the best model group achieved an accuracy more than 20 percentage points lower than human performance. When multiple‑choice options are removed (i.e., open‑ended responses are required), the model group’s accuracy drops further into the single digits, while human performance remains robust. Counterfactual analysis also indicates that current systems generally downplay the use of audio cues when text and images coexist.