Researchers have found a way to watch multimodal AI models decide, token by token, whether to pay attention to an image or the words around it — and the patterns are more structured than the field assumed.
The study, posted to arXiv, tracks attention shifts across two model families and four open-weight vision-language models of varying sizes. The researchers introduce what they call OTaT (One Token at a Time), a framework that monitors how a model's attention moves between image tokens, text tokens, instruction tokens, and previously generated output as a response unfolds. The findings are consistent: attention to the image spikes when the model needs image-derived facts, instruction tokens get revisited at task transitions, and attention to already-generated text grows steadily as generation progresses.
Most interpretability work on large models has asked where computation happens — which layers, which circuits. This paper asks when, which turns out to matter for fixing failures. When the researchers blocked attention pathways to test causality, models fell back on language priors, leaked information across modalities, or flatly denied what was in the image — a taxonomy of failure modes that had no clean explanation before.
Armed with the attention map, the team proposes a test-time intervention: nudge attention toward the relevant modality at the moment it's needed. The result is a measurable boost in multimodal task performance without any retraining. Whether that gain holds on the messier benchmarks outside a controlled lab setting is the next question nobody has answered yet.