benchmarkscritiquebearishExisting KI-VQA benchmarks obscure failure points by reporting only end-task accuracy without isolating sub-problem performanceComputer Vision28 Jul 2026http://arxiv.org/abs/2607.21155v1