diagnostic plotting

This study investigates the robustness of vision-language models (VLMs) against misleading data visualizations. To address this problem, we systematically design eight representative classes of visual deception—including axis truncation, distorted scaling, and misleading color encoding—and conduct the first large-scale, joint quantitative and qualitative evaluation across ten state-of-the-art VLMs. Results demonstrate that most models exhibit significant vulnerability: given identical deceptive charts, they consistently generate semantically incorrect interpretations, revealing structural weaknesses in their visual reasoning capabilities. Our work not only identifies critical blind spots in VLMs’ comprehension of data visualizations but also introduces VisDeceptBench—the first dedicated benchmark for evaluating VLM robustness against misleading charts. Furthermore, we provide empirically grounded insights and methodological guidance to inform the design of deception-resistant mechanisms for future VLM development.

Misleading visualizations severely compromise data interpretation accuracy, yet existing tools lack comprehensive support for detection, attribution, and correction. This paper introduces the first large language model (LLM)-based interactive visualization debugging system, covering all 74 known misleading patterns. It enables fine-grained identification, dynamic adaptation to emerging deceptive strategies, and automatic generation of corrected visualizations. The system integrates Claude and GPT-series models with rule-augmented prompt engineering, a dedicated visualization analysis engine, and a natural language dialogue interface. In benchmark evaluation, it achieves a false positive rate below 4%. An expert user study demonstrates significant improvements in misleading pattern identification accuracy, fact-checking efficiency, and chart credibility. Our core contribution is the first end-to-end, LLM-driven visualization debugging framework—unifying detection, root-cause attribution, and remediation within a single, interactive system.

This study addresses the challenge that current vision-language models struggle to detect misleading data representations in chart captions caused by subtle reasoning errors or flaws in visualization design. To this end, the authors construct the first fine-grained benchmark that combines real-world visualizations with synthetically generated misleading captions, and introduce a novel taxonomy categorizing misleading content into reasoning errors and visualization design errors, enabling precise attribution of deception sources. Systematic evaluation of various open-source and commercial vision-language models reveals that these models are significantly better at identifying visualization design flaws than reasoning-based misinformation, yet they frequently misclassify accurate charts as deceptive. This work bridges the gap between coarse-grained deception detection and fine-grained error-type identification, highlighting limitations in complex semantic understanding.

Deep learning models for time series suffer from poor interpretability, and existing attribution methods yield inconsistent and unreliable results on sequential data. Method: This paper introduces the first systematic, orthogonal six-dimensional evaluation framework for time-series interpretability, covering both classification and segmentation tasks. It proposes six mutually exclusive metrics—categorized by gradient-, propagation-, and perturbation-based explanation methods—and conducts large-scale empirical evaluation across nine state-of-the-art attribution methods and multiple model architectures, using the UCR benchmark and real-world industrial time series. Contribution/Results: Results reveal no globally superior method; explanation quality is highly dependent on task type and model architecture, with L2 regularization and dropout significantly affecting interpretability. The framework confirms low inter-metric correlation, and we publicly release a reproducible toolkit and expert guidance for method selection—establishing a standardized evaluation paradigm for time-series explainability research.

This study addresses the robustness of global post-hoc interpretability methods—such as Partial Dependence Plots (PDP) and Accumulated Local Effects (ALE)—for black-box predictive models on tabular data under data and model perturbations. Motivated by reliability deficits arising from their fragility in model debugging and scientific discovery, we propose the first theoretical bounds quantifying explanation deviation under worst-case and best-case perturbations. Our approach integrates rigorous theoretical analysis with extensive empirical evaluation across multiple real-world and synthetic datasets, systematically assessing explanation sensitivity to controlled perturbations. Results demonstrate that even minor perturbations can substantially distort global feature effect estimates—up to complete reversal of effect direction in adversarial cases. This work establishes the first formal robustness framework for global interpretability, filling a critical gap in the theoretical foundations of trustworthy XAI. It provides the first quantifiable diagnostic benchmark for evaluating the stability and fidelity of global explanations.

This study addresses the limitations of traditional residual plot diagnostics—namely, their reliance on subjective human interpretation, low efficiency, and poor scalability—by introducing computer vision techniques for the first time to automate the assessment of residual plots in linear models. The authors develop the R package autovi and an accompanying Shiny-based interactive web application, autovi.web. Their approach leverages deep visual models to quantify the strength of structural signals in residual plots and produces interpretable diagnostic metrics. This methodology significantly enhances the consistency and efficiency of model fit evaluation, offering statisticians and data analysts a robust, objective, and scalable tool for automated diagnostic assessment in statistical modeling.

This work addresses the challenge of interpreting posterior uncertainty in settings with limited observational data, where it is difficult to assess estimation precision or identify which unobserved variables most effectively reduce uncertainty about a target quantity. The authors propose an interactive visualization system based on “variance deltas,” which, for the first time, structures missing information as a subset tree that explains posterior uncertainty. This approach automatically constructs and organizes subsets of unobserved variables to reveal key sources of uncertainty, integrating Bayesian inference, posterior uncertainty quantification, and a tree-based interactive exploration mechanism that enables users to dynamically split and merge subsets. Experiments on both simulated causal inference tasks and real-world polling prediction demonstrate the system’s effectiveness in uncovering non-obvious yet highly influential factors driving uncertainty.

This study addresses the challenge that current AI agents struggle to spontaneously detect misleading visual designs—such as decorative clutter, axis manipulation, and scale distortion—in charts without explicit prompting. It presents the first systematic evaluation of large language model–driven AI agents’ sensitivity to graphical integrity flaws under unguided conditions, leveraging the BeauVis and PREVis standardized scales to automatically assess their judgments of chart aesthetics and readability. The findings reveal that AI agents consistently assign high scores even when graphical integrity is compromised, indicating a pronounced tendency to overlook deceptive visual practices. This highlights a critical limitation in existing models’ capacity to perceive and evaluate the trustworthiness of data visualizations, underscoring the need for improved mechanisms to support faithful visual interpretation.

This work addresses the lack of intuitive, immediate feedback on fitting errors in existing model-fitting approaches. It proposes an interactive fitting framework that integrates visual and auditory feedback: as users manipulate parametric curves, the system synthesizes audio in real time, with greater model-data discrepancies producing louder and more dissonant sounds. This is the first approach to incorporate auditory cues into model exploration, enabling multisensory assessment of fit quality. Combining interactive visualization, real-time audio synthesis, and Gaussian process regression, the method demonstrates effectiveness and generalizability across four diverse case studies—golf putting, dilution experiments, cosmological parameter estimation, and temperature data fitting—significantly enhancing users’ intuitive perception of model misfit.

Existing research on model visualization predominantly emphasizes data or tasks, lacking a cohesive analytical framework centered on the model itself. This work proposes a model-centric, two-stage visualization approach: first capturing the spatial and temporal behaviors of models through abstract listeners, then integrating these insights into the classical information visualization pipeline. Drawing upon a corpus of 128 papers and 331 annotated visualizations, we systematically construct the first ontology-driven visualization framework oriented toward models. Our analysis reveals a prevailing emphasis on outcome presentation in current literature, while studies exploring model mechanisms—though comparatively scarce—demonstrate disproportionately high impact. This study establishes a new paradigm for the design of future model visualization techniques.