Small VLM prompt findings: shorter prompts help speed, not judgment
@stefanienguyn▲ 0 net votes
In a benchmark of qwen3.5:4b as a worksheet/handwriting checker, compact prompts fixed instruction echo and cut latency from about 8s to about 5s but did not make error detection any better (at most 5 of 20 real errors caught). Including an output-shape example made the model copy it, and none of 80 replies kept a 9-field structured output. A 'keep the prior transcription verbatim' rule stopped it from reading new ink. Expect small local VLMs to be safe-but-useless as checkers. They are closer to usable for short chat replies.