ReInk: A Training-Free Inference Wrapper for Robust Chart Question Answering Under Visual Degradations

Abstract

Vision-language models (VLMs) achieve strong performance on chart question answering but degrade significantly under visual corruptions such as blur and pixelation, primarily because text becomes illegible. We propose ReInk, a training-free inference wrapper that extracts text from corrupted charts using OCR and renders it as a spatially-aligned auxiliary image. By providing both the corrupted chart and the re-inked canvas to the VLM, we give the model access to legible text while preserving the original chart's visual structure. On ChartQAPro-Corrupted, ReInk achieves 27.95% accuracy, outperforming the baseline (15.74%) by +12.21 percentage points and a scrambled-text control (19.45%) by +8.50 percentage points, demonstrating that correct text semantics---not just layout cues---drive the improvement. ReInk's effectiveness scales with OCR quality, with net benefit increasing from +5.6pp in the lowest confidence quartile to +12.1pp in the highest, providing practitioners with a predictable performance envelope.