Search papers, labs, and topics across Lattice.
ChartCynics, a dual-path agentic framework, tackles the challenge of misleading charts by decoupling visual perception and data verification. A Diagnostic Vision Path identifies structural anomalies using ROI cropping, while an OCR-Driven Data Path ensures numerical grounding. An Agentic Summarizer, trained with Oracle-Informed SFT and Deception-Aware GRPO, resolves cross-modal conflicts, leading to a ~29% accuracy improvement over the Qwen3-VL-8B backbone.
Smaller open-source models can outperform proprietary VLMs on misleading charts by strategically decoupling perception and verification within a specialized agentic workflow.
Despite the success of Vision-Language Models (VLMs), misleading charts remain a significant challenge due to their deceptive visual structures and distorted data representations. We present ChartCynics, an agentic dual-path framework designed to unmask visual deception via a"skeptical"reasoning paradigm. Unlike holistic models, ChartCynics decouples perception from verification: a Diagnostic Vision Path captures structural anomalies (e.g., inverted axes) through strategic ROI cropping, while an OCR-Driven Data Path ensures numerical grounding. To resolve cross-modal conflicts, we introduce an Agentic Summarizer optimized via a two-stage protocol: Oracle-Informed SFT for reasoning distillation and Deception-Aware GRPO for adversarial alignment. This pipeline effectively penalizes visual traps and enforces logical consistency. Evaluations on two benchmarks show that ChartCynics achieves 74.43% and 64.55% accuracy, providing an absolute performance boost of ~29% over the Qwen3-VL-8B backbone, outperforming state-of-the-art proprietary models. Our results demonstrate that specialized agentic workflows can grant smaller open-source models superior robustness, establishing a new foundation for trustworthy chart interpretation.
