A streaming assistant must decide, moment by moment, whether to speak and what to say. Standard training uses recorded examples of ideal turns. The model's own live attempts differ from those examples, so it never learns from its own imperfect timing.
Live-OPSD makes the model learn from itself. After it produces a response from the live video alone, the same model gets a training-only hint describing the relevant evidence. The hinted response becomes the target for the version without the hint, and training focuses on moments where they disagree.
On the larger model tested, Live-OPSD scored 59.64 on a streaming benchmark, 5.26 points above the strongest existing method, and remained competitive on another. Teaching a model to correct its own provisional responses improves both when it speaks and what it says.