Navigating the Mirage: A Dual-Path Agentic Framework for Robust Misleading Chart Question Answering

The paper introduces ChartCynics, an agentic dual-path framework for answering questions about misleading charts. It separates structural diagnosis from numerical verification: a Diagnostic Vision Path uses targeted ROI crops to detect anomalies such as inverted axes, while an OCR-Driven Data Path grounds answers in extracted numbers. An Agentic Summarizer reconciles conflicts through Oracle-Informed SFT and Deception-Aware GRPO. The authors report 74.43% and 64.55% accuracy on two benchmarks, an absolute improvement of about 29% over Qwen3-VL-8B, and higher performance than the proprietary models evaluated.