PaperlayerView exact revision on arXiv

Source-linked AI summary

VizAnchor: Decoding Manipulation Intent from Tampering Visualizations via Dual-Anchor Reasoning

Xiaotian Zhang, Huayuan Ye, Haiyang Zhang, Chenhui Li, Changbo Wang, Sicheng Song

arXiv:2608.24535v1cs.CVcs.HC

TL;DR

Visualization tampering can change data-driven interpretations, yet existing methods do not explain how edits alter a chart’s message or why they mislead. VizAnchor combines provenance-verified semantic and spatial anchors with multi-agent VLM reasoning, and experiments show reliable localization and evidence-grounded explanations of manipulation tactics and misleading effects.

  • Problem

    Existing methods locate tampered regions or recover hidden information but do not explain how visualization edits alter messages or why they mislead viewers.

  • Method

    VizAnchor constructs provenance-verified semantic and spatial anchors, then uses multi-agent VLM reasoning to ground edits, reconstruct narratives, and infer misleading intent.

  • Results

    Experiments demonstrate reliable metadata recovery, accurate tamper localization, and evidence-grounded explanations of manipulation tactics and misleading effects.

  • Takeaways & Limitations

    VizAnchor highlights the potential of trustworthy evidence construction and multimodal reasoning for reliable visualization communication.

  • Takeaways & Limitations

    VizAnchor requires pre-embedded metadata and is ineffective for legacy charts or wholly regenerated charts that destroy embedded signals and local forensic traces.

Abstract

from arXiv · show

Data visualizations are widely used for communicating information, but they are also vulnerable to intentional manipulations that induce misleading interpretations. Existing methods focus on locating tampered regions or recovering hidden information, without explaining how the visualization has been manipulated or why the resulting changes may mislead viewers. We propose \textbf{VizAnchor}, a framework for visualization manipulation understanding through dual-anchor evidence construction and VLM-based reasoning. In the first stage, VizAnchor constructs a semantic anchor to recover authentic chart information and a spatial anchor to localize tampered regions. In the second stage, three specialized agents decode the manipulation. The misleader grounding agent analyzes a four-panel visual prompt to predict the misleader information. The chart narrative reconstruction agent takes the original and tampered charts as inputs and reconstructs their respective visual narratives. Finally, the intent inferring agent integrates the visual evidence and misleader information to infer the misleading intent. We further construct a dataset for tampering localization and a dataset for misleading intent inferring. Evaluation shows that VizAnchor accurately localizes manipulations and produces faithful explanations of their manipulation, misleaders, and misleading intents.

Introduction

VizAnchor addresses the semantic difficulty of understanding visualization manipulation by combining authentic-reference recovery, spatial localization, and multi-agent reasoning. Its dual-anchor, two-stage framework reconstructs chart narratives and infers how manipulations create misleading interpretations.

  • Problem: Visualization manipulations can alter viewers’ conclusions through subtle edits to data marks, axes, legends, labels, colormaps, or annotations while retaining visual plausibility.Their effects are semantic rather than merely visual.
  • Limitations: Existing methods emphasize provenance verification or data recovery, while tamper-localization approaches lack an authentic visual reference for explaining deviations from the original message.This limits analysis of why the resulting interpretation is misleading.
  • Framework: VizAnchor uses a two-stage pipeline that embeds recoverable provenance metadata and a crop-synchronization cue, then decodes position information to restore manipulated charts to a canonical canvas.The process produces an aligned tampered chart and recovers metadata from it.
  • Reasoning: Three agents decode tamper type, affected components, and manipulation process, reconstruct narratives for authentic and tampered charts, and infer the underlying misleading intent.They operate through a four-panel prompt containing the Semantic Anchor, tampered chart, Tamper Map, and Localized Comparison.
  • Framework: Its dual-anchor evidence combines a provenance-verified Semantic Anchor with a pixel-level Spatial Anchor for authentic information recovery and precise tampering localization.This mechanism is introduced to address missing authentic references and insufficient spatial evidence.

Related Work

Related work studies misleading visualizations, visualization protection, and generic image tampering detection, but these lines largely address initial chart design, provenance, localization, or forensic traces separately. VizAnchor frames post-creation chart tampering as requiring joint identification of changes, message alteration, and misleading intent.

  • Misleading Visualization: Prior misleading-visualization work examines distorted scales, encodings, proportions, and rhetorical designs that bias chart interpretation.Recent benchmarks also assess whether MLLMs detect or explain misleading cues in standalone charts.
  • Misleading Visualization: Tampered visualizations are defined as initially valid charts intentionally modified at the image level after creation to alter their original message.This setting requires recognizing misleadingness, identifying changes relative to the authentic chart, and inferring modification intent.
  • Visualization Protection: Visualization protection methods embed metadata for provenance verification and source-data recovery or signals for detecting and localizing post-creation tampering.These methods address protection-oriented objectives rather than fully explaining how edits change chart messages or what misleading intent they serve.
  • Image Tampering Detection: Passive image-tampering detectors localize edits using forensic traces such as noise residuals, boundary artifacts, and cross-channel inconsistencies.Although effective for generic image manipulation, these methods are not tailored to visualization semantics.

Method

VizAnchor uses dual-anchor evidence construction to recover authentic chart semantics and localize cropping and edits, then applies a three-agent VLM pipeline to decode manipulation intent. Its invertible watermarking and crop-aware alignment support semantic-anchor retrieval, while localization combines geometric and content-change evidence.

  • Framework Overview: VizAnchor comprises dual-anchor evidence construction followed by VLM-based manipulation-intent decoding.The first stage constructs semantic and spatial anchors; the second uses specialized agents for grounding, narrative reconstruction, and intent inference.
  • Semantic Anchor Construction: The invertible watermarking module embeds metadata and positional information into charts for crop-robust recovery.It uses affine-coupling transformations, tiled metadata redundancy, invertible token shuffling, and separate metadata and position flow blocks.
  • Semantic Anchor Construction: Recovered metadata retrieves the registered original chart as a Semantic Anchor and pairs the untampered watermarked chart with the aligned tampered chart for edit localization.The Crop-Aware Module estimates crop location and scale from recovered position maps, restores the chart to a canonical canvas, and retains the missing region as a crop mask.
  • Spatial Anchor Construction: The Spatial Anchor unions crop and local-edit masks to capture both geometric cropping and fine-grained content modifications.A U-Net-based Localization Module compares the untampered watermarked and aligned tampered charts using RGB content and binary difference cues.
  • Intent Decoding: Three specialized VLM agents use a four-panel prompt to ground visual edits, reconstruct authentic and tampered chart narratives, and infer misleading intent.The prompt contains the authentic chart, tampered chart, and predicted tamper mask.

Experiments

Experiments evaluate VizAnchor’s dual-anchor evidence construction, VLM-based manipulation understanding, and ablations across existing and newly constructed chart-tampering datasets. Results show strong fidelity, metadata recovery, localization, manipulation recognition, process description, and intent inference, while ablations validate the anchors and multi-agent design.

  • Experimental Design: Experiments follow VizAnchor’s two-stage design, evaluating evidence construction, VLM-based manipulation understanding, and ablations.The evaluation covers fidelity, metadata recovery, localization, tamper understanding, and misleading-intent inference.
  • Datasets: VAD contains 1,500 automatically generated chart pairs with pixel-level masks and 120 manually created chart pairs for localization and manipulation-understanding evaluation.VAD-Eval additionally includes crop-based manipulations compared with VDD-Eval.
  • Dual-Anchor Evidence Construction: 43.28 PSNR is achieved by VizAnchor on VGD, while SSIM and LPIPS remain competitive with the best baselines.The results indicate that jointly embedding metadata and position information causes limited visual distortion.
  • Dual-Anchor Evidence Construction: 99.43% BitAcc remains after 60% pixel modification, and 99.36% after 50% chart-area cropping, with VizAnchor consistently leading metadata recovery.The evaluation considers local tampering ratios from 15% to 60% and crop ratios from 10% to 50%.
  • Dual-Anchor Evidence Construction: 0.7418 overall IoU and 0.8375 F1 are achieved for local-edit localization, while crop localization reaches 0.9893 F1 on VAD-LocEval.VizAnchor outperforms the applicable baselines across the reported crop and local-edit evaluations.
  • VLM-Based Manipulation Understanding: VizAnchor reaches 0.91 Accuracy and 0.90 Macro-F1 for tamper-type classification, improves component Exact Match Accuracy from 0.34 to 0.63, and obtains process and intent Cos-FA scores of 0.70 and 0.75.Process and intent AI-FA scores are 0.91 and 0.86, respectively, exceeding VizDefender’s corresponding results.
  • Ablation Studies: Ablations show the Semantic Anchor raises tamper-type Macro-F1 from 0.473 to 0.778, while the full multi-agent pipeline achieves Intent Cos. FA 0.753 and AI FA 0.856.Removing MGA or CNRA lowers performance, and using IIA Only produces the lowest scores.

User Study

A blind user study with 30 participants found that VizAnchor received the highest average ratings among the compared methods for explanation quality across two evaluated dimensions.

  • Study design: Thirty participants rated 10 anonymized manipulation examples on a five-point Likert scale in a blind user study.Participants averaged 24.07 years of age (σ = 4.28); 16 were male, 13 female, and 1 preferred not to disclose gender.
  • Compared methods: The study compared VizAnchor against VizDefender and Dual-Anchor with a vanilla VLM.
  • Results: VizAnchor achieved the highest average ratings across both dimensions, scoring 1.437 for manipulation process and 1.370 for the second dimension.

Discussion

VizAnchor combines proactive watermarking with multi-agent reasoning but is limited by its dependence on pre-embedded metadata and its focus on partial manipulations.

  • Limitations: Pre-embedded metadata restricts VizAnchor to charts protected at creation, preventing direct application to legacy or third-party visualizations without such signals.The framework cannot directly process visualizations already circulating online without embedded metadata.
  • Limitations: VizAnchor’s anchors are designed to localize partial manipulations, including cropping and localized pixel edits.The stated anchor design targets localized rather than unspecified manipulation patterns.

Conclusion … A.2 Robustness-Test Sample Construction

VizAnchor combines semantic and spatial anchors with multi-agent multimodal reasoning to explain visualization manipulations, their misleading effects, and inferred intents. Its robustness samples are deterministically constructed from held-out charts using controlled pixel-modification and crop procedures.

  • Conclusion: VizAnchor constructs semantic anchors from provenance-verified originals and spatial anchors from localized tamper evidence for reasoning over authentic and manipulated visual messages.Its multi-agent framework grounds manipulation attributes and reconstructs visualization narratives.
  • Conclusion: VizAnchor infers misleading intents and achieves reliable metadata recovery, accurate tamper localization, and evidence-grounded explanations of manipulation tactics and effects.The framework combines trustworthy evidence construction with multimodal reasoning for reliable visualization communication.
  • A.1 IWM Training and Test Split: The VisGuard dataset contains 17,957 chart images, split into 14,964 training images and 2,993 held-out test images using a deterministic 5:1 split.The split uses random seed 20260710.
  • A.2 Robustness-Test Sample Construction: Robustness samples are generated on the fly from all 2,993 held-out images after watermarking at model input resolution and mapping residuals to original resolution.Attack generation uses seed 20260713 for source order, payload, donor selection, and attack geometry.
  • A.2 Robustness-Test Sample Construction: Modified-pixel tests use irregular binary masks whose modified-pixel ratio is evaluated at rmod ∈{0.15, 0.30, 0.45, 0.60}.Masks derive from seeded 32 × 32 random fields that are upsampled, smoothed, and thresholded by largest values.
  • A.2 Robustness-Test Sample Construction: Each modified-pixel attack replaces only masked pixels with content from a deterministically selected different test chart, producing 2,993 samples per ratio.This yields 11,972 modified-pixel test instances per evaluated method.
  • A.2 Robustness-Test Sample Construction: Crop robustness reports the fraction of original image area removed, while retaining a seeded rectangular region directly from the original-resolution watermarked chart.The retained rectangle is stored as the ground-truth crop box; restoration occurs later during evaluation.

A.3 Localization Data Construction and Evaluation Splits … B.1 Invertible Watermarking Module

The paper constructs controlled localization and reasoning benchmarks with verified annotations, then specifies an invertible watermarking module for crop-aware training. Localization uses generated tampering pairs for training and separate manually constructed or independent datasets for evaluation.

  • A.3 Localization Data Construction and Evaluation Splits: VAD-LocTrain contains 1,500 automatically generated local-tampering pairs spanning text, bar, and scatter-chart edit families.Each edit is chart-aware and deterministically sampled using seed 20260617 plus the sample index.
  • A.3 Localization Data Construction and Evaluation Splits: Each sample stores clean and tampered charts, an edit-derived binary mask, an overlay, and JSONL provenance, while localization receives a seven-channel input.The seven channels comprise the clean chart, tampered chart, and binary RGB difference; all 1,500 requested samples completed successfully.
  • A.3 Localization Data Construction and Evaluation Splits: The generated pairs are split into 1,200 training and 300 validation pairs, with final localization evaluated separately on VAD-LocEval and VDD-LocEval.Validation is used only for checkpoint selection and hyperparameter monitoring, not as reported final test performance.
  • A.3 Localization Data Construction and Evaluation Splits: VAD-Eval contains 120 manually created chart pairs, partitioned for localization into 20 crop samples and 100 non-crop local-edit samples.All 120 pairs are used as VAD-ReasonEval, and VAD-LocEval does not update localization-network parameters.
  • A.3 Localization Data Construction and Evaluation Splits: VDD-LocEval contains 1,000 local-edit localization samples, while VDD-ReasonEval contains 100 manipulation-understanding samples.These datasets provide independent evaluation coverage for localization and manipulation understanding.
  • A.4 Reasoning-Annotation Protocol: Two visualization experts produced and verified reference annotations covering tamper type, tampered components, manipulation process, and misleading intent.The second pass checked label consistency and factual correspondence with each original–tampered chart pair.
  • A.4 Reasoning-Annotation Protocol: The annotation protocol defines nine chart-specific tamper types and seven tampered-component categories for evaluation.Tamper types include MDV, ARD, MCV, DAA, ML, HL, ARL, DVD, and MC; components include Data Label, Region, Axis, Annotation, Legend, Logo, and Colormap.
  • B.1 Invertible Watermarking Module: The invertible watermarking module uses K = 81 bits as a 9 × 9 map, tiled into a 27 × 27 redundant module map at 512 × 512 chart resolution.Its configuration includes patch size p = 16, token dimension d = 768, latent dimension dz = 64, and L = 4 invertible transformation blocks; crop side length is sampled from [102, 512].

B.2 IWM Training Configuration … C.5 Single-Call Visual-Evidence Ablation

The paper specifies reproducible training configurations for watermarking and localization, then defines a staged VLM reasoning pipeline and controlled visual-prompt protocols for decoding chart manipulation and misleading intent.

  • B.2 IWM Training Configuration: The IWM and Crop-Aware Module are jointly optimized with losses for chart fidelity, metadata prediction, perceptual similarity, and position-map accuracy.The objective combines Lsteg, Lssim, Lmeta, Llpips, and Lpf.
  • B.2 IWM Training Configuration: The IWM trains from scratch for 40 epochs using AdamW, batch size 4, initial learning rate 1 × 10−4, weight decay 1 × 10−5, and StepLR decay by 0.9 each epoch.Training uses seed 20260701, saves every epoch, and reports the final epoch-40 checkpoint without validation-based early stopping.
  • B.3 Localization U-Net: The localization U-Net receives a seven-channel concatenation of authentic RGB, tampered RGB, and a binary RGB-difference map, resized to 512 × 512.The network uses four encoder levels with widths 32, 64, 128, and 256, followed by a bilinear-upsampled decoder with skip connections.
  • B.3 Localization U-Net: The localization model trains for 65 epochs with AdamW at 1×10−4 learning rate and 1×10−4 weight decay, using an 80/20 split of 1,500 samples.BCE and Dice have equal external weights, BCE uses positive-class weight 5, and checkpoint selection uses validation IoU at threshold 0.5.
  • B.4 VLM Inference Protocol: The complete reasoning pipeline uses three sequential gemini-3.5-flash calls: MGA identifies manipulation, CNRA reconstructs narratives, and IIA infers misleading intent.Agents receive no reference tamper type, component, process, or intent during inference; ground truth is used only for evaluation.
  • C.5 Single-Call Visual-Evidence Ablation: The A1–A5 visual-evidence ablation uses a single model call to predict all four outputs under each specified panel setting and is not directly comparable with the complete three-agent pipeline.The prompt restricts predictions to visible evidence, treats overlays as localization aids, and defines allowed tamper labels including 10mix and 1modifying_data_point_values through 9modifying_colormap.
  • C Prompts and Output Schemas: The visual prompt is a two-by-two PNG collage of four 512 × 512 panels: Semantic Anchor, Tampered Chart, Tamper Map, and Localized Comparison.The panels remain at 512 × 512 rather than being downsampled into one 512 × 512 image, and outputs are constrained to strict JSON schemas.
  • C Prompts and Output Schemas: MGA grounds tamper type, component, and concrete process; CNRA separately summarizes authentic and tampered chart narratives; IIA compares these evidences to produce one natural-language intent.The agents are instructed to rely on visible evidence, avoid inventing unreadable values or unsupported motives, and keep manipulation description distinct from intent inference.

D Additional Tampering-Localization Results … E.1 Task-Level Results

VizAnchor achieves strong tamper-region localization on both evaluated benchmarks and qualitatively identifies diverse manipulation types and misleading intents. Additional reasoning results define evaluation metrics and report task-level and aggregate evaluations for tamper recognition, manipulation-process understanding, and intent inference.

  • D Additional Tampering-Localization Results: VizAnchor obtains the highest IoU, recall, and F1 on both VAD-LocEval and VDD-LocEval.Its IoU/F1 scores are 0.7775/0.8609 on VAD-LocEval and 0.7382/0.8352 on VDD-LocEval.
  • D Additional Tampering-Localization Results: On VAD-LocEval, VisGuard has higher precision, but VizAnchor’s higher recall produces the stronger overall F1.Precision is 0.8481 versus 0.8303, while recall is 0.6339 versus 0.9312 for VisGuard and VizAnchor, respectively.
  • D.1 Qualitative Localization Comparison: Qualitative 512×512 examples compare original, watermarked, tampered, ground-truth, and VizAnchor localization results across representative tamper types.The examples include modified values or coordinates, altered data points, deceptive annotations, and visual disproportion.
  • D.1 Qualitative Localization Comparison: The qualitative predictions describe intents involving distorted perceptions, hidden values, exaggerated trends, and misleading category or speed comparisons.Examples include exaggerating negative COVID-19 perceptions, hiding Australia’s 84% approval rating, halving apparent poverty shares, and emphasizing rising CO2 emissions.
  • D.1 Qualitative Localization Comparison: Other qualitative intents involve obscuring quantitative scale or provenance, downplaying categories, creating inconsistent encodings, and altering presented data.Predictions include hiding the y-axis scale, removing branding and source attribution, downplaying nuclear power, mismatching legend colors, and omitting Cairns Airport.
  • E Additional VLM Reasoning Results: Additional VLM reasoning evaluations use cosine-based and strong-model free-form answer metrics, with AI FA normalized to [0, 1].Cos. FA and AI FA denote cosine-based and strong-model free-form answer evaluation, respectively.
  • E.1 Task-Level Results: Task-level results evaluate tamper type, component recognition, manipulation process, and misleading intent, while aggregate tables report sample-weighted results across VAD-ReasonEval and VDD-ReasonEval.Exact denotes exact-match accuracy, M-F1 denotes Macro-F1, and AI FA is normalized to [0, 1] in the aggregate manipulation-process and intent table.

E.2 Qualitative Comparisons … F.1 Category-Level Failure Modes

Qualitative comparisons show VizAnchor more faithfully identifies tampering and misleading intent than VizDefender across illustrated cases, while category-level analysis exposes residual errors in globally encoded manipulations. The section also defines evaluation groupings and ablation settings for visual evidence and multi-agent reasoning.

  • E.2 Qualitative Comparisons: 1.00/1.00 normalized AI scores were achieved by VizAnchor for modifying data-point values and adding or removing data points, exceeding VizDefender in both cases.VizDefender scored 0.00/0.80 and 0.00/0.00, respectively, for Process/Intent.
  • E.2 Qualitative Comparisons: VizAnchor correctly explains modified axis labels, oversized labels, hidden values, swapped colors, and combined bar or row edits in representative comparisons.VizDefender instead attributes several examples to OCR noise, label–bar mismatch, altered values, or unrelated regions.
  • E.2 Qualitative Comparisons: VizAnchor’s explanations connect local edits to global interpretations, including obscured poverty values and altered country rankings or perceived COVID-19 case distributions.The described intents involve hiding exact quantities, swapping United States and United Kingdom colors, shortening the United States bar, and removing China.
  • E.3 Visual-Evidence Ablation: The visual-evidence analysis stratifies complete VizDefender and VizAnchor pipeline results by annotated tamper type and audited original-chart type, never treating cropping as a class.For tamper-type rows, Type Acc. is class recall and Type F1 is one-vs-rest F1; chart-type rows use subgroup accuracy and Macro-F1, while Component F1 covers seven labels.
  • E.3 Visual-Evidence Ablation: Tables 17 and 18 report pooled pipeline results by tamper type and chart type, with AI scores normalized to [0, 1].The reported conventions distinguish per-class recall, one-vs-rest class F1, subgroup Macro-F1, and seven-component Macro-F1.
  • E.4 Multi-Agent Ablation: The multi-agent ablation compares intent inference from the complete visual prompt alone, variants removing CNRA or MGA, and the complete MGA–CNRA–IIA pipeline.Table 19 evaluates normalized strong-model Intent AI on VAD and VDD, while Table 20 reports sample-weighted Intent AI FA across both ReasonEval datasets.
  • F.1 Category-Level Failure Modes: 0.471 and 0.600 tamper-type recall were attained by VizDefender on Mix and modifying data-point values, while component exact-match accuracy fell to 0.088 on Mix and 0.100 on modifying colormap and data–visual disproportion.These results concentrate errors in manipulations requiring local edits to be connected with global chart encoding.
  • F.1 Category-Level Failure Modes: 0.650 recall made modifying colormap VizAnchor’s hardest tamper-type category, while component exact-match accuracy was 0.292 for adding or removing data points, 0.350 for modifying colormap, and 0.529 for Mix.The strict multi-label metric counts a prediction as incorrect when any affected component is wrong.

F.2 Intent-Relation Errors and Metric Disagreement … G.2 Statistical Analysis

VizAnchor’s intent understanding varies by manipulation and chart type, with metric disagreements exposing relation-level errors. A controlled user study compares VizAnchor with baselines using participant-level ratings and paired nonparametric tests.

  • F.2 Intent-Relation Errors and Metric Disagreement: VizAnchor’s category-averaged intent cosine similarity ranges from 0.691 to 0.816, while normalized AI scores range from 0.700 to 0.946.Data–visual disproportion achieves the highest cosine similarity (0.816), whereas hiding labels achieves the highest AI score (0.946).
  • F.2 Intent-Relation Errors and Metric Disagreement: A compound manipulation changing a map title and recoloring two countries receives a normalized AI score of 0.20 because VizAnchor mispredicts the cross-country relation direction.The model correctly identifies the urban-to-rural semantic shift and both affected countries.
  • F.2 Intent-Relation Errors and Metric Disagreement: Cosine similarity captures topical alignment, whereas the AI judge more strongly penalizes errors in relation direction, entity attribution, and contradiction.VizDefender illustrates this disagreement with modifying-colormap scores of 0.584 cosine and 0.180 AI, and modifying-coordinate-values scores of 0.602 cosine and 0.320 AI.
  • F.3 Chart-Dependent Behavior: VizAnchor’s intent cosine similarity is 0.769 on bar charts and 0.770 on line charts but falls to 0.687 on pie and donut charts.Component Macro-F1 ranges from 0.480 on pie and donut charts to 0.964 on area charts.
  • F.3 Chart-Dependent Behavior: Chart-level differences are confounded by subgroup size, since pie and donut charts contain eight samples versus 99 bar-chart and 32 line-chart samples.The reported differences therefore may reflect sampling variation as well as chart-dependent difficulty.
  • F.4 Representative Intent Error Visualizations: Figure 18 replaces VizAnchor’s predicted Tamper Map with the ground-truth overlay to diagnose the compound error, although the mask was unavailable during inference.The example retains intent cosine 0.834 and normalized AI score 0.20 while showing the altered topic and affected countries.
  • G.1 Study Design: The user study samples ten manipulation examples, evenly split between VAD-ReasonEval and VDD-ReasonEval, and evaluates manipulation-process description and misleading-intent inference.Each example includes the original and tampered charts, and participants compare VizDefender, a vanilla VLM, and VizAnchor with hidden method identities.
  • G.2 Statistical Analysis: Ratings use a five-point agreement scale, with participant-level task–method means as the statistical unit and Student-t confidence intervals, Friedman tests, and paired Wilcoxon signed-rank comparisons.Zero differences are excluded, average ranks handle ties, and the large-sample normal approximation is used without continuity correction.

G.3 Results · H Disclosure of AI-Assisted Editing

The results show that VizAnchor was rated higher than both comparison methods for process and intent, with statistically significant method differences. The authors also disclose AI-assisted language and LaTeX editing while retaining responsibility for all technical content and results.

  • G.3 Results: The study included 30 participants: 16 men, 13 women, and one person who did not disclose gender.Valid age data for 29 participants had a mean of 24.07 years (SD = 4.28), median of 22, and range of 18–35.
  • G.3 Results: Participants represented backgrounds in data visualization or visual analytics, computer vision, human–computer interaction, education, design or art, chemistry, and other areas.The corresponding counts were 13, 4, 3, 2, 1, 1, and 6, respectively.
  • G.3 Results: No Likert-scale responses were missing.Participant-level descriptive statistics used scores ranging from −2 to 2.
  • G.3 Results: VizAnchor received higher ratings than both comparison methods for both process and intent tasks.The Friedman tests found differences among methods for process (χ2(2) = 42.660, p = 5.45 × 10−10, Kendall’s W = 0.711) and intent (χ2(2) = 39.138, p = 3.17×10−9, Kendall’s W = 0.652).
  • G.3 Results: Pairwise comparisons used two-sided paired Wilcoxon signed-rank tests with Holm correction across three comparisons within each task.These pairwise results are reported in Table 22.
  • H Disclosure of AI-Assisted Editing: A generative AI assistant supported language editing and LaTeX consistency checking.The authors verified the technical content, experimental settings, analyses, and reported results.
  • H Disclosure of AI-Assisted Editing: The authors take full responsibility for the manuscript.This responsibility follows their verification of all technical content, experimental settings, analyses, and reported results.
Loading 2608.24535v1…