MisVisFix: Fixing Misleading Visualizations
- MisVisFix is an interactive dashboard that detects, explains, and corrects misleading visualizations using multimodal LLMs.
- It analyzes bitmap images to identify and categorize issues as major, minor, or potential, enhancing chart reliability.
- The system generates corrected charts via Python code and offers an interactive chat interface for iterative refinement.
MisVisFix is an interactive dashboard for detecting, explaining, and correcting misleading visualizations using LLMs. It analyzes existing bitmap images rather than chart specifications, uses both Claude and GPT models, classifies issues as major, minor, or potential concerns, and combines issue detection with visual localization, code-based chart regeneration, and an interactive chat interface. In the reported evaluation, MisVisFix achieved Precision 0.94, Recall 0.98, and F1 Score 0.96 on a dataset spanning all 74 issue categories from Lo et al.’s taxonomy (Das et al., 6 Aug 2025).
1. Scope, definition, and problem setting
MisVisFix addresses misleading visualizations as charts or graphics that distort how viewers understand data. The distortion may be intentional or unintentional, and the paper places these issues on a broad spectrum that includes obvious visual tricks such as truncated axes and subtler problems such as inappropriate encodings, selective data presentation, missing context, or misrepresentation of findings. A central distinction in the system design is between structural problems, which are visible in the image itself, and contextual problems, which require reasoning about what the chart claims, what data is shown or omitted, and whether the visual framing supports unsupported conclusions (Das et al., 6 Aug 2025).
The motivation is explicitly framed against three limitations in prior work: a detection-only focus, limited issue coverage, and the lack of interactive, user-facing tools. MisVisFix is therefore designed to support the full workflow of inspecting a chart image, detecting issues, categorizing them by severity, explaining why they matter, localizing them on the chart, proposing improvements, generating corrected charts, and iteratively refining results through conversation. The paper also contrasts this objective with specification-based tooling, emphasizing that MisVisFix works on existing bitmap images, which is crucial when only a posted image is available.
The taxonomy used by MisVisFix contains all 74 potential misleading visualization types from Lo et al. The system groups detected issues into three severity levels. Major Issues are problems that substantially distort data perception or pose serious risk of misinterpretation; examples given include truncated axes, misleading color encodings, and fabricated data. Minor Issues are problems that reduce clarity or readability without fundamentally changing interpretation; examples include missing titles and inconsistent label formatting. Potential Issues are elements that may be problematic depending on context or audience, such as specialized encoding choices or technical terminology.
2. System architecture and operational workflow
MisVisFix is organized as a modular pipeline with four major components: the Input Processing Module, Visualization Analysis Engine, Visualization Correction and Generation Module, and Interactive User Interface (Das et al., 6 Aug 2025).
| Component | Function |
|---|---|
| Input Processing Module | Normalize image and extract chart elements |
| Visualization Analysis Engine | Detect issues, assign severity, localize regions |
| Visualization Correction and Generation Module | Generate corrected charts through Python code |
| Interactive User Interface | Compare outputs, inspect issues, refine through chat |
The operational workflow begins when a chart image is uploaded. The system first normalizes the image and extracts visible chart elements such as title, axes, legends, and annotations. It also detects the chart type. The analysis engine then passes the chart, extracted content, and the issue catalog to multimodal LLMs, which identify issues and categorize them by severity. A later stage asks the models to provide detailed explanations and improvement suggestions, and another stage asks for percentage-based coordinates so the dashboard can highlight the problematic region.
The paper describes a typical session as a sequence of nine steps: Upload a chart image; Preprocessing and extraction; Issue detection; Severity classification; Explanation; Localization; Correction; Comparison; and Refinement. The refinement stage includes chat-based requests, optional dataset upload when extraction is inaccurate, and user feedback about missed issues. This design makes MisVisFix both a diagnostic tool and an iterative redesign environment.
A notable architectural choice is the dual-model architecture. Claude-3.7 is the primary model for data extraction from charts because it showed the strongest VLAT performance, while Claude-3.7 and GPT-4.5 together are used for issue detection and correction generation. The reported model comparison attributes stronger contextual issue detection to Claude and stronger structural issue detection and code generation to GPT-4.5.
3. Detection logic, severity assignment, and localization
MisVisFix uses LLM-based issue detection with structured prompting with Chain-of-Thought guidance. The prompt instructs the model to: identify chart type and chart elements, examine axis properties, analyze color usage, review data representation, and list detected issues by severity category (Das et al., 6 Aug 2025). The system does not initially ask for free-form explanations. Instead, it first returns only the names of identified issues, then prompts again for detailed explanations, and then organizes the results by severity. This staged design is intended to simplify downstream parsing and reduce output variability.
The severity labels are not produced by a formal scoring rule. The paper states that GPT-4.5 and Claude 3.7 are prompted to assign issues to the three severity levels, but it does not provide a deterministic algorithm for that classification. The resulting categorization accuracy is reported separately from raw detection performance, which reflects the fact that borderline cases remain harder than straightforward structural defects.
Localization is implemented through a prompt-based coordinate mechanism rather than a separately trained vision detector. After issue identification, the system asks the multimodal model to output percentage-based top_gap and left_gap coordinates for each issue. These values are then used to highlight regions responsively across different display sizes. The examples given in the paper include axis-focused localization for “Truncated Axis”, mark-focused localization for “Color Blind Unfriendly”, and top-area highlighting for “Missing Title”. The paper does not introduce a runtime confidence score, voting rule, or formal decision function for combining model outputs.
The paper also reports differentiated behavior across issue classes. Structural issues are easier to detect than contextual issues. This is consistent with the system’s internal organization: some tasks depend mostly on visible chart structure, while others depend on inferring omitted context, selective framing, or misrepresentation of findings.
4. Explanation, correction generation, and interactive use
Once issue names are available, MisVisFix generates detailed explanations that state what the issue is, why it is misleading or problematic, how it affects interpretation, and what should be improved (Das et al., 6 Aug 2025). These explanations are exposed through the dashboard rather than hidden in model logs. Clicking an issue expands a panel with its description and improvement guidance, turning the system into both an inspection tool and a visualization-literacy aid.
Correction is code-based rather than image-synthesis-based. The correction module takes the original chart image, extracted data, and the list of identified issues, then prompts the model to generate Python code using matplotlib, seaborn, or similar plotting libraries. The reported correction process has three stated steps: Extract data from the original visualization; Generate visualization code with fixes for identified issues; and Apply styling to match the original. Generated code is executed in a controlled environment; runtime errors are sent back to the LLM for repair; resulting plots are converted to base64 images for rendering; and large images are stored using Firebase when token limits make direct passing impractical.
The paper explicitly states that early experiments with DALL-E were unsuccessful because such models are optimized for creative image generation rather than precise data visualization. This design choice is significant because it keeps correction tied to reconstructed plotting code rather than unconstrained image editing. The correction module also prioritizes chart types considered less prone to misinterpretation—bar charts, line charts, and scatterplots—and may replace a fundamentally problematic chart type rather than merely restyle it. A pie chart with too many similar-colored segments may therefore be converted into a sorted bar chart, and a line chart with dual y-axes may be replaced by single-axis line charts.
The interface is organized into seven panels: Panel A shows the original misleading visualization with issue localization overlays; Panel B and Panel C show corrected versions generated by Claude and GPT; Panel D is the dataset upload panel; Panel E and Panel F list issues detected by GPT and Claude; and Panel G is the interactive chat interface. The main interaction pattern is issue-to-region linking: hovering over an issue highlights the corresponding chart region, and clicking reveals the explanation and improvement advice. The chat interface supports follow-up requests such as changing design details, exploring alternative visual forms, or adding missed issues. The paper presents this as an interactive learning mechanism, although it also notes that no formal online-learning algorithm or longitudinal validation of improvement over time is provided.
5. Evaluation and reported performance
The evaluation dataset contains 450 visualizations: 360 misleading and 90 valid. Misleading charts were sourced from Lo et al.’s collection and designed to cover all 74 issue categories, with at least three examples per category and additional examples for the 20 most common issue types. Valid visualizations came from academic journals, government reports, and major news outlets, and were manually verified by three visualization experts (Das et al., 6 Aug 2025).
| Evaluation aspect | Reported result |
|---|---|
| Overall detection | Precision 0.94; Recall 0.98; F1 Score 0.96 |
| Issue categorization | Major 94.2%; Minor 85.6%; Potential 82.8%; Overall 87.5% |
| Localization precision | 91.3% |
| Correction success | 82.5% of identified issues were successfully addressed |
Performance by chart type shows stronger behavior on common forms than on more complex ones. The reported F1 scores are 0.97 for bar charts, 0.96 for line charts, 0.93 for pie/donut charts, 0.92 for scatterplots, and 0.91 for other chart types. Performance by issue category includes 0.98 for Truncated Axis, 0.97 for 3D Effects, 0.96 for Dual Axis, 0.95 for Missing Title, 0.94 for Missing Axis Labels, 0.94 for Inappropriate Color Use, 0.93 for Inconsistent Scale, 0.93 for Selective Data Presentation, 0.92 for Data Manipulation, and 0.91 for Misrepresentation of Findings. Localization precision is also broken down by region type: 96.4% for axes, 94.8% for textual components, 85.2% for color schemes, and 88.9% for data point encodings.
Model selection before final deployment used VLAT (Visualization Literacy Assessment Test) with a custom three-step prompting technique: extract data from graphs, sort extracted data, and answer questions based on the extracted data. The reported VLAT scores are 50.17 for Claude-3.7-sonnet, 34.33 for GPT-4.5-preview, 33.50 for GEMINI-2.0-pro, and 28.82 for the human baseline under the paper’s comparison scheme. The final expert evaluation involved 5 visualization experts with a mean 13.4 years of experience; the quantitative ratings were 8.5/10 for Detection Accuracy, 8.0/10 for Usefulness of Suggested Improvements, and 8.0/10 for Likelihood to Use in Professional Work.
The paper also reports operational limitations. Performance decreases by approximately 12% on low-resolution images, and end-to-end analysis takes about 2–3 minutes because the pipeline makes multiple sequential LLM calls across two models. The authors further note difficulties with multiple interacting issues, domain-specific scientific visualizations, false positives when unusual domain conventions are acceptable, and style-preservation trade-offs in corrected charts.
6. Relation to adjacent systems and stated limitations
MisVisFix occupies a distinct position in the broader literature on visualization quality, misvisualization diagnosis, and chart integrity. VizLinter is a linter and fixer for Vega-Lite JSON specifications that translates a chart into ASP facts, detects violations in a 41-rule base, and selects repairs through optimization; its scope is invalid or illegal specification structure rather than bitmap-image misinformation (Chen et al., 2021). VisQualdex offers a manual codex with 7 categories and 60 criteria/questions, answered as yes/no or NA, for evaluating static visualizations through a structured web audit (Sawicki et al., 2022). VizDefender targets a different threat model again: proactive detection and interpretation of tampered visualization images through a semi-fragile watermark module and an intent analysis module (Song et al., 21 Dec 2025). MisVisFix differs from all three by centering on existing bitmap charts, using multimodal LLMs to reason over a 74-type misinformation taxonomy, and coupling critique to automatic chart regeneration (Das et al., 6 Aug 2025).
The paper is explicit that MisVisFix still has important limits. It struggles more with specialized domains such as scientific visualizations, performs worse on low-resolution images, and is less reliable when multiple issues interact in complicated ways. It can also produce false positives when a design choice is acceptable under domain conventions, and corrected charts may lose some of the original aesthetic identity. A plausible implication is that the system is strongest as a high-coverage interactive reviewer for broadly recognizable chart forms, rather than as a domain-complete authority for all visualization contexts.
Another limitation is formalization. The paper does not provide formal algorithms, pseudocode, optimization objectives, or LaTeX equations for the core MisVisFix pipeline. It also does not give mathematical formulas for F1 score, precision/recall, severity scoring, coordinate computation, or model selection. The adaptation mechanism is presented as an interface workflow rather than as a validated online-learning method. Finally, because the system depends on LLM reasoning, risks noted in related work on multimodal models—such as hallucination or inconsistency—remain relevant, even though the paper’s structured prompting substantially improves over direct LLM-only baselines.
In that sense, MisVisFix is best understood as an LLM-powered visual misinformation assistant for charts: a practical system that moves misleading-visualization analysis from isolated detection experiments toward a complete interactive workflow of diagnosis, explanation, localization, correction, and iterative refinement.