Abstract Recent advances in Large Language Models have led to Large Reasoning Models, which produce step‐by‐step reasoning traces. Such traces may offer insight into how models think, improving explainability and clarifying the underlying process. These traces, however, are often verbose and complex, making them cognitively demanding to comprehend. With this in mind, we propose R e T race , an interactive system that structures and visualizes textual reasoning traces to support understanding. We use a validated reasoning taxonomy to produce structured reasoning data and investigate two types of interactive visualizations thereof. In a controlled human‐subject study, the visualizations provided more accurate comprehension of the model's reasoning and less perceived effort than a raw text baseline. The results of this study could have design implications for making long and complex machine‐generated reasoning processes more usable and transparent, an important step in AI explainability.

Abstract Dimensionality reduction (DR) techniques are often characterized by whether they preserve global, high‐level structures in the data or local, neighborhood structures. This distinction matters in visualization: global methods can obscure clusters while local methods can over‐emphasize them. Yet, even when clusters appear distinct, their relative arrangement in the projection may be arbitrary or misleading, a common issue in techniques such as t‐SNE and UMAP. Existing cluster quality metrics either only measure cluster separability or assume spherical, globular clusters in the original space. We introduce the Class Angular Distortion Index (CADI), a metric that uses internal angles among point triples to determine the faithfulness of cluster organization in a projection. We show cases on both real and synthetic data where existing cluster metrics fail, but CADI provides an interpretable result. Since it relies on computing angles, CADI is also differentiable, enabling optimization. We demonstrate this with a CADI‐based DR technique.

Abstract Edge bundling is widely used for reducing visual clutter in large 2D network and trajectory visualizations. Various edge bundling methods have been proposed, each producing qualitatively distinct outputs for the same data; however, few quantitative metrics exist for systematic evaluation. In this paper, we propose a set of quantitative metrics at the edge level (e.g., geometric properties of the resulting curves), the bundle level (e.g., thickness and number of bundles), and the global level (e.g., ambiguity and clustering). We propose a benchmark of 115 representative datasets and evaluate five representative edge bundling techniques that cover a broad range of methodological approaches. We also conduct a correlation analysis between bundling metrics and network drawing properties. To facilitate further analysis and comparison, we provide an interactive dashboard that includes all methods, metrics, and datasets, enabling side‐by‐side exploration of edge bundling effects. All supplemental materials and a link to our dashboard application are available at OSF.

In this paper, we test whether Multimodal Large Language Models (MLLMs) can match human-subject performance in tasks involving the perception of properties in network layouts. Specifically, we replicate a human-subject experiment about perceiving quality (namely stress) in network layouts using GPT-4o, Gemini-2.5 and Qwen2.5. Our experiments show that giving MLLMs the same study information as trained human participants yields performance comparable to that of human experts and exceeds that of untrained non-experts. Additionally, we show that prompt engineering that deviates from the human-subject experiment can lead to better-than-human performance in some settings. Interestingly, like human subjects, the MLLMs seem to rely on visual proxies rather than computing the actual value of stress, indicating some sense or facsimile of perception. Explanations from the models are similar to those used by the human participants (e.g., an even distribution of nodes and uniform edge lengths).

Complex, high-dimensional data is ubiquitous across many scientific disciplines, including machine learning, biology, and the social sciences. One of the primary methods of visualizing these datasets is with two-dimensional scatter plots that visually capture some properties of the data. Because visually determining the accuracy of these plots is challenging, researchers often use quality metrics to measure the projection's accuracy and faithfulness to the original data. One of the most commonly employed metrics, normalized stress, is sensitive to uniform scaling (stretching, shrinking) of the projection, despite this act not meaningfully changing anything about the projection. Another quality metric, the Kullback-Leibler (KL) divergence used in the popular t-Distributed Stochastic Neighbor Embedding (t-SNE) technique, is also susceptible to this scale sensitivity. We investigate the effect of scaling on stress and KL divergence analytically and empirically by showing just how much the values change and how this affects dimension reduction technique evaluations. We introduce a simple technique to make both metrics scale-invariant and show that it accurately captures expected behavior on a small benchmark.

A strengthened version of Harborth’s well-known conjecture - known as Kleber’s conjecture - states that every planar graph admits a planar straight-line drawing where every edge has integer length and each vertex is restricted to the integer grid. Positive results for Kleber’s conjecture are known for planar 3-regular graphs, for planar graphs that have maximum degree 4, and for planar 3-trees. However, all but one of the existing results are existential and do not provide bounds on the required grid size. We provide polynomial-time algorithms for computing crossing-free straight-line drawings of trees and cactus graphs with integer edge lengths and integer vertex position on polynomial-size integer grids. We also give an historic overview of planar straight-line graph drawing results.

Stress in a graph drawing has been a popular layout principle for more than two decades. Low stress drawings exhibit the property that the geometric distances between all pairs of nodes correlate with the shortest paths between them. The assumption has always been that low stress drawings are "nicer" and better support human perception and comprehension than high stress drawings. In this paper, we put these assumptions to the test. We use a normalised scale-independent and rotation-independent metric for stress; this is necessary to ensure strict controls on our experimental stimuli. We report on three experiments, exploring human perception of stress, preference for stress, and the effect of stress on a graph performance task. We conclude that people can see stress in a graph drawing, that they prefer low stress drawings, and that their performance in a shortest path task improves as stress decreases - thus empirically confirming long-standing assumptions.

Abstract Bipartite graphs are commonly used to visualize objects and their features. An object may possess several features and several objects may share a common feature. The standard visualization of bipartite graphs, with objects and features on two (say horizontal) parallel lines at integer coordinates and edges drawn as line segments, can often be difficult to work with. A common task in visualization of such graphs is to consider one object and all its features. This naturally defines a drawing window, defined as the smallest interval that contains the x-coordinates of the object and all its features. We show that if both objects and features can be reordered, minimizing the average window size is NP-hard. However, if the features are fixed, then we provide an efficient polynomial-time algorithm for arranging the objects, so as to minimize the average window size. Finally, we introduce a different way of visualizing the bipartite graph, by placing the nodes of the two parts on two concentric circles. For this setting we also show NP-hardness for the general case and a polynomial-time algorithm when the features are fixed.

Abstract We investigate the usability of Euclidean, spherical and hyperbolic geometries for network visualization. Several techniques have been proposed for both spherical and hyperbolic network visualization tools, based on the fact that some networks admit lower embedding error (distortion) in such non‐Euclidean geometries. However, it is not yet known whether a lower embedding error translates to human subject benefits, e.g., better task accuracy or lower task completion time. We design, implement, conduct, and analyze a human subjects study to compare Euclidean, spherical and hyperbolic network visualizations using tasks that span the network task taxonomy. While in some cases accuracy and response times are negatively impacted when using non‐Euclidean visualizations, the evaluation shows that differences in accuracy for hyperbolic and spherical visualizations are not statistically significant when compared to Euclidean visualizations. Additionally, differences in response times for spherical visualizations are not statistically significant compared to Euclidean visualizations.

Most of the common graph layout principles (a.k.a. "aesthetics") on which many graph drawing algorithms are based are easy to define and to perceive. For example, the number of pairs of edges that cross each other, how symmetric a drawing looks, the aspect ratio of the bounding box, or the angular resolution at the nodes. The extent to which a graph drawing conforms to these principles can be determined by looking at how it is drawn - that is, by looking at the marks on the page - without consideration for the underlying structure of the graph. A key layout principle is that of optimising "stress", the basis for many algorithms such as the popular Kamada & Kawai algorithm and several force-directed algorithms. The stress of a graph drawing is, loosely speaking, the extent to which the geometric distance between each pair of nodes is proportional to the shortest path between them - over the whole graph drawing. The definition of stress therefore relies on the underlying structure of the graph (the "paths") in a way that other layout principles do not, making stress difficult to describe to novices unfamiliar with graph drawing principles, and, we believe, difficult to perceive. We conducted an experiment to see whether people (novices as well as experts) can see stress in graph drawings, and found that it is possible to train novices to "see" stress - even if their perception strategies are not based on the definitional concepts.

When visualizing a high-dimensional dataset, dimension reduction techniques are commonly employed which provide a single 2 dimensional view of the data. We describe ENS-t-SNE: an algorithm for Embedding Neighborhoods Simultaneously that generalizes the t-Stochastic Neighborhood Embedding approach. By using different viewpoints in ENS-t-SNE’s 3D embedding, one can visualize different types of clusters within the same high-dimensional dataset. This enables the viewer to see and keep track of the different types of clusters, which is harder to do when providing multiple 2D embeddings, where corresponding points cannot be easily identified. We illustrate the utility of ENS-t-SNE with real-world applications and provide an extensive quantitative evaluation with datasets of different types and sizes.

Many real-world networks contain structurally-equivalent nodes. These are defined as vertices that share the same set of neighboring nodes, making them interchangeable with a traditional graph layout approach. However, many real-world graphs also have properties associated with nodes, adding additional meaning to them. We present an approach for swapping locations of structurally-equivalent nodes in graph layout so that those with more similar properties have closer proximity to each other. This improves the usefulness of the visualization from an attribute perspective without negatively impacting the visualization from a structural perspective. We include an algorithm for finding these sets of nodes in linear time, as well as methodologies for ordering nodes based on their attribute similarity, which works for scalar, ordinal, multidimensional, and categorical data.

Abstract Visualizing graphs and networks in non‐Euclidean space can have benefits such as natural focus+context in hyperbolic space and the familiarity of interactions in spherical space. Despite work on these topics going back to the mid 1990s, there is no survey, or a part of a survey for this area of research. In this paper we review and categorize over 60 relevant papers and analyze them by geometry, (e.g., spherical, hyperbolic, torus), by contribution (e.g., technique, evaluation, proof, application), and by graph class (e.g., tree, planar, complex).

Hyperbolic geometry offers a natural ‘focus+context’ for data visualization and has been shown to underlie real-world complex networks. However, current hyperbolic network visualization approaches are limited to special types of networks and do not scale to large datasets. With this in mind, we designed, implemented, and analyzed three methods for hyperbolic visualization of networks in the browser based on inverse projections, generalized force-directed algorithms, and hyperbolic multi-dimensional scaling (H-MDS). A comparison with Euclidean MDS shows that H - MDS produces embeddings with lower distortion for several types of networks. All three methods can handle node-link representations and are available in fully functional web-based systems.