IC-902BLIP-2 and MiniGPT-4 confidence-based question selection underperforms the original question for paraphrased candidates but succeeds for semantically enriched REPARe questions
The paper compares three selection strategies: using the original question, selecting among PEGASUS-generated paraphrases via model confidence, and selecting among REPARe-generated semantically enriched questions via model confidence. For BLIP-2, paraphrase selection yields 40.23% on A-OKVQA direct (vs. 41.86% baseline, −1.63) and 73.57% on A-OKVQA MC (vs. 73.89%, −0.32), while REPARe selection reaches 44.87% and 77.20% respectively. MiniGPT-4 shows the same pattern: paraphrase selection gives 26.28% (A-OKVQA direct, vs. 27.51% baseline) and 40.52% (MC, vs. 41.66%), while REPARe selection reaches 33.23% and 63.20%. This indicates the models' self-assessed confidence is unreliable for choosing among cosmetic rewrites but effective for semantically enriched candidates.
The paraphrase baseline uses a single PEGASUS model; the paper does not test whether a stronger paraphraser or more paraphrase candidates would close the gap. The oracle setting (using gold answers for selection) does show paraphrases can help, confirming the issue is in the selection mechanism, not the paraphrases themselves.