VISxGenAI: 2nd GenAI, Agents, and the Future of VISbookmark_border

We adapt Stevens’s power law to measure the implicit ability of AI models to read visualizations, which can reveal the built-in perceptual mechanisms of algorithmic models. In our pilot study, AIs see no legend. In the color conditions, no colormap name is provided either. GPT-5.5 first views a reference visual representation and estimates its magnitude, then estimates the magnitude of each subsequent image of the same representation relative to that reference. Our evaluation of twelve visual variables makes how algorithmic models read visual encodings measurable, comparable with human perception, and more transparent to humans. osf.io/wsrzm has source code, images, magnitude data samples, and analyses.

The rapid adoption of multi-agent Large Language Model (LLM) architectures introduces new challenges for understanding communication, processing activity, and failures across interacting agents. Recent developer tools and debugging approaches provide workflow construction, conversation inspection, trace navigation, and intervention mechanisms, but the dynamic distribution and propagation of activity across the multi-agent architecture remains difficult to inspect as a coherent visual process. We present AgentPulse, a brain-inspired Visual Analytics (VA) approach that treats inter-agent communication and activity as the primary objects of visual analysis. Functionally related agents are organized into cognitive regions, while animated communication, live cluster activity, an fMRI-inspired agent-level temporal heatmap, semantic anomaly indicators, and a coordinated sequence view expose complementary structural and temporal aspects of an execution. A formative qualitative evaluation with four experts examines the comprehensibility and perceived utility of the proposed visual encodings. Our work contributes a complementary perspective to existing agent debugging interfaces by emphasizing the visual analysis of how activity propagates through and accumulates within multi-agent workflows.

Editing agents on 2D infinite canvases must reason about application state that pixels may not reveal. Targets may be offscreen or covered, while visual containment and touching arrows may disagree with recorded frame membership or connector bindings. We ask how a structured observation layer can support accurate object selection during agentic editing on 2D infinite canvases. We present CanvasObs, which augments visual context with a structured payload and guidance for reading four components: Identity & State, Visibility & Structure, Relations, and Offscreen Object bounds in action coordinates. A Remote Team Handoff case study illustrates retrieving an editable offscreen copy, choosing a covered chart by color, separating cards by frame membership, and preserving an evidence connection. We turn these cases into ten controlled configurations and compare CanvasObs with the baseline agent in 1,200 browser trials across three models. CanvasObs improves Correct target for every model. Combined across models, it raises Correct target from 16.5% to 48.3%, an increase of 31.8 percentage points (95% confidence interval [29.0, 34.7]), and Full task success from 4.3% to 25.0%. These results show that compact structured observation can improve object choice on 2D infinite canvases without redesigning the editing pipeline.

There are often multiple valid visualizations, spanning different chart types, visual encodings, and aesthetic choices, that can all support a user’s goal of analyzing a dataset. Determining which of these candidate visualizations is most appropriate, however, remains challenging. Prior work has leveraged expert-curated taxonomies to guide visualization design and subsequent quality evaluation, a practical evaluation model must also account for human-centric factors. Latest visualization tools integrate Large Language Models (LLMs) to generate and refine visualizations, using LLM-as-a-judge to provide design feedback. Without a deeper understanding of whether, and in what ways, human and LLM perceptions of visualizations align, there is a real risk of deploying low-quality agents that do not reflect human judgment. We study this question through a pilot experiment - we formulate the problem as a preference task over curated chart pairs and compare the judgments of state-of-the-art vision models with those of a panel of human annotators. We establish a cautionary tale: such judges risk filtering out informative, compelling displays. We find that frontier models collapse human subjectivity, a rich signal that reveals contrasting design choices, into near-deterministic verdicts.

The combination of generative AI and autonomous agents is shifting visual analytics from systems that primarily support user actions toward systems that can independently plan and execute substantial parts of an analysis. This can reduce the effort required to move from analytical intent to results, but it also redistributes consequential choices from users to agents, including what data to inspect, which questions to pursue, what evidence to prioritize, and which alternatives to leave unexplored. The challenge is therefore not merely preserving human agency, but ensuring that this changing distribution of agency remains understandable and renegotiable. We introduce fluid agency as the capacity to understand and continually renegotiate the distribution of analytical agency among people and agents, and characterize it through five dimensions, Revisable Delegation, Reflective Friction, Analytical Bound, Distributed Agency, and Agency Realization. We then apply these dimensions to three existing agentic visualization systems to illustrate their value by revealing how agency is currently distributed and identifying where and how support for fluid agency could be strengthened. For future agentic visualization, fluid agency shifts the design question from how much work agents can take on to how people and agents can continually negotiate their roles as an analysis unfolds.

AI agents can now analyze, synthesize, and visualize complex datasets, yet turning their outputs into a coherent narrative aligned with a reader’s expectations remains an open challenge. A single generation pass produces charts that misread the data, claims the visuals do not support, and narratives that lose coherence over a long document. We present a multi-agent system for automated visual data reporting in which every generative stage is paired with a deterministic validator that can reject its output, and failures are repaired in the same context or disclosed rather than shipped. The system organizes the validated panels into a guided narrative and renders it as an interactive report, letting readers follow the analytical thread and then explore the underlying data themselves.

Visual encoding channels are ranked for human observers by how accurately people decode quantities from them, and that ranking underpins decades of design guidance. We ask whether it carries over to multimodal LLMs. We replicate the Cleveland–McGill and Heer–Bostock graphical perception experiments across six models from three capability tiers under six prompting conditions, and extend the task set with two color-based channels. We report three findings. The geometric channels are interchangeable, with no significant pairwise difference, consistently across models and prompts. Luminance and saturation are not usable quantitative channels at all: errors exceed those expected from random responding. A follow-up experiment shows that distractor elements degrade estimation, whereas spatial separation between the compared marks does not, suggesting that for agent-facing charts, the amount of surrounding visual material matters more than the choice of encoding.

Vision-language models (VLMs) have demonstrated significant advances in reasoning with visual data, which is particularly important given that much of human knowledge is encoded visually in scientific articles and other sources. However, a gap remains in understanding how artificial agents, powered by VLMs, perceive and navigate decision-making under uncertainty in the presence of visual data. This study explores this gap by evaluating and analyzing autonomous decision-making processes. Building upon prior work in visualization literacy and decision theory, we conducted two complementary studies: one examining how agents perceive and quantify uncertainty encoded in visualizations, and another investigating how artificial agents make decisions based on visual data in the form of images. We analyzed perception accuracy and decision performance across models and visualization types. Our results show that GPT-4o demonstrates the highest accuracy in perceiving uncertainty in pie and bar charts among the tested models. For decision-making, autonomous agents showed the best alignment with rational decision theory when uncertainty was represented through circle charts, icon arrays, or text.

This report presents two LLM agents developed for the Agentic VIS Challenge 2026: a generator that produces interactive visualizations from a task description and dataset, and an evaluator that reviews visualizations through a browser. Generation separates data analysis from visualization design. An analysis stage formulates questions based on the task and writes executable code to compute their answers. It provides the visualization stage with prepared data files and a compact description of the analysis. The visualization stage uses these outputs to organize and build an interactive HTML page, with browser-based checks supporting revisions to its charts and layout. The evaluator assigns scores using browser screenshots, page information, and changes observed after interacting with the visualization. Both agents use context management, token limits, and error recovery to operate within the competition’s shared token budget. We describe the resulting design and reflect on the tradeoffs observed during its development. Example visualizations and the agent code are available online: Public Task 2, Held-out Task 2, and GitHub.

Benchmarks for visualization agents score the agent alone; the human only supplies reference answers. For high-dimensional data this misses what matters: the knowledge of which variables count is held by on-site experts, not by the model, so the useful question is how well an analyst and an agent do together. We study one specific form of collaboration—the agent proposes a view, the analyst accepts, rejects, or modifies it, and those judgments are kept for later sessions—and propose five measures for it: complementarity, intervention efficiency, appropriate reliance, knowledge retention, and agency. The first four are computed from interaction logs alone. We test the measures, not people: in a testbed for parallel-coordinates design on public manufacturing data (UCI SECOM), with a simulated analyst whose behavior we control, all four are computable and respond to the controlled differences. Whether they reveal anything about real analysts is the next step; we release the testbed for it.

The visualization community has invested substantially in deploying Vision-Language Models (VLMs) for visual analytics. However, the complementary question of how exactly do these models reason about visualizations, and how do we know when they are doing so correctly has received comparatively little rigorous attention. In this position paper, we argue that the interpretability methods currently applied to VLMs in the visualization domain inherit a cascade of structural mismatches. Methods ported from computer vision assume a crisp notion of objectness, spatial locality, and to a large extent, context-independence that natural images possess but data visualizations do not. Human gaze studies, often proposed as ground truth to compare and contrast interpretability methods, carry their own fragilities rooted in the reproducibility crisis in perception research. Finally, and perhaps more importantly, VLMs themselves are not vision models as their visual inputs merely serve as adaptors into a language model’s (linguistic) representational space, fundamentally altering the notion of what “visual attention” means. To probe these questions, we introduce a dual causal-chain account of visualization comprehension, distinguishing the designer’s intended reading path from the reader’s actual comprehension trace, we propose a set of necessary conditions for valid interpretability evaluation in this domain. We posit that the visualization community is uniquely positioned to develop novel interpretability methods which can significantly impact both theoretical and empirical advances in the broad field of visual analytics.

Data scientists are beginning to analyze data alongside visualization agents that both explain their reasoning and produce charts. These two channels can disagree: an agent may acknowledge uncertain evidence in its rationale while emitting a plan whose visual semantics assert certainty. We measure this divergence, without any human evaluation, in rendering-free structured plans from 72 neutral source cases and 24 matched pressure variants, yielding 672 valid outputs from two fixed model case studies, gpt-4o and gpt-5.4-mini. Both models usually acknowledged uncertain evidence yet often did not act on it: this acknowledgment–action mismatch occurred in 41% and 59% of valid uncertain baseline cases, and every action failure violated a prespecified action-consistency rule. Within-call reflection helped one model but not the other. Restricting the action menu removed forbidden action types, yet one model rebuilt the same semantics through allowed parameters. Under explicit pressure to appear decisive, even acknowledgment eroded. For human–agent analysis, the plan, not the rationale, carries the visual claim and is the artifact to check.

Evaluating visualization systems in specialized domains, such as genomics, is challenging due to the scarcity of domain experts and difficulty recruiting relevant user types. LLM-based synthetic personas provide an appealing workaround, but they face skepticism as their claims often lack verifiable grounding. We present Sycamore, an exploratory approach to evaluating visualization systems with grounded synthetic personas whose responses are restricted to interview evidence. Sycamore enforces provenance with participant citations for each claim and an abstention mechanism when retrieval yields no relevant evidence. A protocol runner guides the synthetic personas through an evaluation of the live visualization system, producing multi-turn transcripts for qualitative coding. We instantiate the approach on Geranium, a multimodal genomics visualization search engine, with seven evaluators grounded in prior interviews on genomics visualization authoring. We report themes in the resulting transcripts, highlight an emergent requirement absent from the source data, and provide a preliminary comparison to a prior human user study. Recognizing that our findings derive from an exploratory perspective and that synthetic proxies cannot replace human evaluation, we discuss their implications for integrating synthetic evaluators alongside expert user studies. The protocol, codebook, extended results, transcripts, and demo video are in the Supplementary Materials and at osf.io/kdfr3; the source code and project page are at huyen-nguyen.github.io/sycamore.

Recent LLM-based visualization systems generate charts from natural language, drive visualization tools, prototype analytics interfaces, and score well on visualization benchmarks, but complex scientific visualization tasks still require repeated human steering before they yield a usable application. Much of the work done by visualization researchers calls for automating a larger task: an interactive application that starts from ambiguous domain materials, explores the data, derives task requirements, designs visual representations and coordinated interactions, implements the result, and checks whether it answers the original scientific questions. We argue that agentic visualization systems should automate this application-level design loop, and we study it in a zero-shot setting: one prompt with no human intervention after launch. The domain scientists who need such applications rarely have a visualization expert to consult, so the agent has to get as far as it can from an underspecified request. Each clarifying turn adds design intent that the system is supposed to supply, making the final result hard to attribute. We support the position with a reusable multi-agent harness that acts as a VIS co-scientist for domain experts. We test the harness on IEEE SciVis Contests spanning multiple science and engineering fields. These contests include the pressures we care about: ambiguous requirements, diverse data modalities, design trade-offs, and task-driven validation. We use the runs to map what the current harness can and cannot do, and to identify where visualization expertise should enter agent design.

Students learning visualization design increasingly use GenAI-powered authoring tools, which can produce a chart from a short natural-language description. While these tools lower the barrier to producing visualizations, they also allow learners to produce a satisfactory chart without making the design decisions the exercise was intended to develop, which can give an illusion of competence while leaving design skills underdeveloped. In this position paper, we make a case for AI agents that facilitate cognitive friction as a strategy to support visualization design learning. We define friction for this setting, distinguish it from the incidental cost of a difficult interface, and organize it by the point in the learner–agent loop at which it intervenes, yielding three positions. We then describe five skills an agent needs to deliver friction at these positions. We illustrate both the positions and the skills through a chart-critique exercise suited to introductory visualization teaching. We then discuss what would be needed to test whether an agent with these skills actually supports learners’ design reasoning.

As LLMs take up the role of authoring charts using visualization DSLs, the human constraints that shaped those languages may no longer apply—what is easy for a person is not necessarily easy for a model. To understand how LLMs might work better with DSLs, we explore where and how they fail with current DSL designs. We evaluate 10 JSON-style visualization DSLs with 41 tasks across 3 LLMs, then assess the generated specifications with JSON and rendering checks, and qualitative coding of failed cases. Analyzing how this specification generation process fails, we identify four recurring failure patterns, link each to specific DSL features, and discuss design considerations for future DSL designs.