Navigating the Mirage: A Dual-Path Agentic Framework for Robust Misleading Chart Question Answering
ChartCynics splits chart reading into visual diagnosis and OCR-grounded data checks to catch deceptive designs.
The paper targets misleading charts that can fool vision-language models through tricks such as inverted axes or distorted representations. Its framework uses cropped visual inspection to flag structural anomalies, then compares that against extracted numerical data. A trained summarizer resolves conflicts between what the chart appears to show and what the data supports. On two benchmarks, the authors report 74.43% and 64.55% accuracy, about 29 points above the Qwen3-VL-8B backbone. HF Daily Papers' note
The paper targets misleading charts that can fool vision-language models through tricks such as inverted axes or distorted representations. Its framework uses cropped visual inspection to flag structural anomalies, then compares that against extracted numerical data. A trained summarizer resolves conflicts between what the chart appears to show and what the data supports. On two benchmarks, the authors report 74.43% and 64.55% accuracy, about 29 points above the Qwen3-VL-8B backbone. HF Daily Papers' note
score 4