IC-900BLIP-2, MiniGPT-4, and LLaVA-1.5 show degraded zero-shot VQA accuracy on underspecified questions, with absolute improvements of 1.14–7.94% when questions are augmented with visually-grounded details

Archiki Prasad, Elias Stengel-Eskin, Mohit Bansal

SourceRephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models

The paper measures zero-shot VQA accuracy of three released LVLMs on VQAv2, A-OKVQA, and VizWiz, comparing original questions against questions augmented with image-derived visual details, rationale information, and disambiguated references. All three models improve when the question is made more specific, with the largest gains on VizWiz (7.94% for MiniGPT-4) and A-OKVQA (6.41% for MiniGPT-4, 5.47% for BLIP-2). The effect is consistent across encoder-decoder (FLAN-T5) and decoder-only (Vicuna) LLM backbones, indicating a general sensitivity to question underspecification rather than an architecture-specific artefact.

Evidence
correlational
Key metric
VQAv2 overall: BLIP-2 62.58→66.43 (+3.85), MiniGPT-4 7B 51.47→54.49 (+3.02), LLaVA-1.5 76.21→77.35 (+1.14); A-OKVQA direct: BLIP-2 41.86→47.33 (+5.47), MiniGPT-4 13B 41.53→47.94 (+6.41), LLaVA-1.5 62.56→66.19 (+3.63); VizWiz: MiniGPT-4 7B 29.87→37.81 (+7.94), BLIP-2 XXl 62.81→66.27 (+3.46), LLaVA-1.5 57.07→59.46 (+2.39)
Caveat
All improvements are measured on validation sets; test sets are not publicly available. The paper notes that 7% of REPARe questions on A-OKVQA contain the gold answer, meaning part of the gain may come from answer leakage into the question.
Model
BLIP-2, MiniGPT-4, LLaVA-1.5 / LLaVA-v1.5
Concepts
Failure mode
Datasets
VQAv2 [eval], A-OKVQA [eval], VizWiz [eval]
Methods
PEGASUS [compared-to]
Related findings
IC-901, IC-902
Extraction
automatic-extraction