ChartCynics Enhances Chart QA Robustness
HKUST · hf · 2026-07-16
ChartCynics: A More Robust Chart QA Framework
This paper discusses the vulnerability of vision-language models to misleading charts: charts can deceive models through inverted axes, exaggerated scales, or distorted data structures. The authors propose ChartCynics, a "skeptical" reasoning framework that decouples perception from verification.
Methodology
- Diagnostic Vision Path: Captures structural anomalies through strategic ROI cropping.
- OCR-Driven Data Path: Grounds data at the numerical level using OCR to avoid relying solely on visual appearances.
- Agentic Summarizer: Integrates information from both paths and resolves cross-modal conflicts.
Training Approach
- Oracle-Informed SFT: Used for reasoning distillation.
- Deception-Aware GRPO: Used for adversarial alignment to reinforce resistance to visual traps.
Experimental Results
- Achieved accuracies of 74.43% and 64.55% on two benchmarks respectively.
- Represents an absolute improvement of roughly 29% over the Qwen3-VL-8B backbone.
- Even outperforms certain top-tier closed-source models.
Conclusion
The authors conclude that specialized agentic workflows can equip smaller open-source models with stronger robustness in chart understanding, offering a new perspective for trustworthy chart interpretation.
More from Multimodal
- Gemini Omni Flash turns a boat cabin into a cave in Flow by Google — chrisfirst · 2026-07-22
- A simple workflow to turn a photo into an image prompt using Gemini, Grok, or GPT Image — harshitagu72595 · 2026-07-22
- A Reddit user proposes a consistency LoRA to keep anime and game scenes visually stable — ThirdWorldBoy21 · 2026-07-22
- Hand-painted figurines run through Seedance look eerily alive — cocktailpeanut · 2026-07-22
- An AI agent-made bayou country music video is making the rounds on Reddit — LazyKaleidoscope4696 · 2026-07-22
- Testing Qwen 3 Image: Map Borders Shift Based on Prompts, Includes Chinese Labels — NirantK · 2026-07-22