I haven't seen models return nothing (void as in this paper) but I have seen OpenAI models, using structured output, emit that structured output followed by kilobytes of whitespace.
My model harness has to have code to detect it - check for runs of whitespace and terminate the API call. Honestly it seems bizarre: not just that it happens but that it's not handled at their layer.
I've seen this about, handwaving, one in a thousand times. Twice during manual test invocations over the past year, twice in early live beta testing that luckily was logged, and presumably it happened much more during later live work (it's caught and logged but we don't retain logs long for privacy reasons and I can't check how often.)
My model harness has to have code to detect it - check for runs of whitespace and terminate the API call. Honestly it seems bizarre: not just that it happens but that it's not handled at their layer.
I've seen this about, handwaving, one in a thousand times. Twice during manual test invocations over the past year, twice in early live beta testing that luckily was logged, and presumably it happened much more during later live work (it's caught and logged but we don't retain logs long for privacy reasons and I can't check how often.)