Prompt Echoing Fixes Question-First Paradox in Vision-Language Models
Repeating the question before and after an image boosts VLM accuracy by up to 19 points.
Deep Dive
Researchers discovered the 'question-first paradox' in VLMs: placing the question before the image harms accuracy despite intuition. Their training-free fix, 'prompt echoing' (restating the question on both sides of the image), closes the paradox and surpasses the best single-pass ordering on NaturalBench, POPE, Winoground, and open-ended VQAv2—by up to 19 Winoground group-accuracy points. No training, fine-tuning, or architecture change required.
Key Points
- Question-first prompting in VLMs underperforms image-first ordering by up to 17.5 group-accuracy points due to answer token ignoring distant question tokens.
- Prompt echoing (duplicating the question on both sides of the image) resolves the paradox without training, fine-tuning, or architecture changes.
- Additional gains come from echoing the image itself, restoring whole-image view that causal decoders otherwise lose.
Why It Matters
A zero-cost prompt trick that universally improves VLM accuracy, applicable to any model without retraining.