I’m testing the new Gemini 3.1 Flash Lite on document extraction tasks, and for some pages it consistently returns early results with completely unfinished extraction.
For example, when extracting line items from a 7-page bank statement, page 1 comes back perfectly, page 2 is unfinished (2 rows out of ~70), pages 3 and 4 are perfect, and pages 5, 6, and 7 are also unfinished.
I’ve tried the same tasks with Gemini 2.5 Flash and Gemini 3 Preview - no problems whatsoever.
This is really frustrating because there are absolutely no logs showing why the model stopped generating output (Finish_reason=STOP for all of them, even the good ones). It just stops.
For context, I’m using it in Vertex AI.