Dasol Choi, DongGeon Lee, Brigitta Jesica Kartono, Helena Berndt, Taeyoun Kwon, Joonwon Jang, Haon Park, Hwanjo Yu, Minsuk Kahng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The Transformer architecture underpins modern large language models powering state-of-the-art text generation and AI applications. However, its complexity makes it difficult for non-experts to learn. Existing resources often lack interactivity, rely on static descriptions of simplified architectures, or fail to reflect models’ behavior with real data. To address this gap, we introduce Transformer Explainer, an interactive visualization tool for non-experts to learn Transformers. The tool integrates an overview illustrating the Transformer’s data flow with on-demand explanations that gradually reveal mathematical details. Smooth transitions across abstraction levels highlight the interplay between high-level structures and low-level operations. Running a live GPT-2 instance directly in the browser, Transformer Explainer empowers learners to experiment with custom input and hyperparameters without setup, observing next-token predictions in real time. A 90-participant user study showed that our tool offered significant advantages in improving user understanding and engagement. Transformer Explainer has attracted over 490,000 users.

Large Language Models (LLMs) are increasingly embedded in applications, and people can shape model behavior by editing prompt instructions. Yet encoding subtle, domain-specific policies into prompts is challenging. Although this process often benefits from concrete test cases, test data and prompt instructions are typically developed as separate artifacts, reflecting traditional machine learning practices in which model tuning was slow and test sets were static. We argue that the fast, iterative nature of prompt engineering calls for removing this separation and enabling a new workflow: data-prompt co-evolution, where a living test set and prompt instructions evolve in tandem. We present an interactive system that operationalizes this workflow. It guides application developers to discover edge cases, articulate rationales for desired behavior, and iteratively evaluate revised prompts against a growing test set. A user study shows our workflow helps people refine prompts systematically, better aligning them with their intended policies. This work points toward more robust and responsible LLM applications through human-in-the-loop development.

While large language models (LLMs) have shown remarkable capability to generate convincing text across diverse domains, concerns around its potential risks have highlighted the importance of understanding the rationale behind text generation. We present LLM ATTRIBUTOR, a Python library that provides interactive visualizations for training data attribution of an LLM’s text generation. Our library offers a new way to quickly attribute an LLM’s text generation to training data points to inspect model behaviors, enhance its trustworthiness, and compare model-generated text with user-provided text. Thanks to LLM ATTRIBUTOR’s broad support for computational notebooks, users can easily integrate it into their workflow to interactively visualize attributions of their models.

Large Reasoning Models (LRMs) have become powerful tools for complex problem solving, but their structured reasoning pathways can lead to unsafe outputs when exposed to harmful prompts. Existing safety alignment methods reduce harmful outputs but can degrade reasoning depth, leading to significant trade-offs in complex, multi-step tasks, and remain vulnerable to sophisticated jailbreak attacks. To address this, we introduce SAFEPATH, a lightweight alignment method that fine-tunes LRMs to emit a short, 8-token Safety Primer at the start of their reasoning, in response to harmful prompts, while leaving the rest of the reasoning process unsupervised. Empirical results across multiple benchmarks indicate that SAFEPATH effectively reduces harmful outputs while maintaining reasoning performance. Specifically, SAFEPATH reduces harmful responses by up to 90.0\% and blocks 83.3\% of jailbreak attempts in the DeepSeek-R1-Distill-Llama-8B model, while requiring 295.9x less compute than Direct Refusal and 314.1x less than SafeChain. We further introduce a zero-shot variant that requires no fine-tuning. In addition, we provide a comprehensive analysis of how existing methods in LLMs generalize, or fail, when applied to reasoning-centric models, revealing critical gaps and new directions for safer AI.

Evaluating large language models (LLMs) presents unique challenges. While automatic side-by-side evaluation, also known as LLM-as-a-judge, has become a promising solution, model developers and researchers face difficulties with scalability and interpretability when analyzing these evaluation outcomes. To address these challenges, we introduce LLM Comparator, a new visual analytics tool designed for side-by-side evaluations of LLMs. This tool provides analytical workflows that help users understand when and why one LLM outperforms or underperforms another, and how their responses differ. Through close collaboration with practitioners developing LLMs at Google, we have iteratively designed, developed, and refined the tool. Qualitative feedback from these users highlights that the tool facilitates in-depth analysis of individual examples while enabling users to visually overview and flexibly slice data. This empowers users to identify undesirable patterns, formulate hypotheses about model behavior, and gain insights for model improvement. LLM Comparator has been integrated into Google's LLM evaluation platforms and open-sourced.

Recent work in vision-and-language demonstrates that large-scale pretraining can learn generalizable models that are efficiently transferable to downstream tasks. While this may improve dataset-scale aggregate metrics, analyzing performance around hand-crafted subgroups targeting specific bias dimensions reveals systemic undesirable behaviors. However, this subgroup analysis is frequently stalled by annotation efforts, which require extensive time and resources to collect the necessary data. Prior art attempts to automatically discover subgroups to circumvent these constraints but typically leverages model behavior on existing task-specific annotations and rapidly degrades on more complex inputs beyond "tabular" data, none of which study vision-and-language models. This paper presents VLSlice, an interactive system enabling user-guided discovery of coherent representation-level subgroups with consistent visiolinguistic behavior, denoted as vision-and-language slices, from unlabeled image sets. We show that VLSlice enables users to quickly generate diverse high-coherency slices in a user study (n=22) and release the tool publicly1.

In this paper, we present DendroMap, a novel approach to interactively exploring large-scale image datasets for machine learning (ML). ML practitioners often explore image datasets by generating a grid of images or projecting high-dimensional representations of images into 2-D using dimensionality reduction techniques (e.g., t-SNE). However, neither approach effectively scales to large datasets because images are ineffectively organized and interactions are insufficiently supported. To address these challenges, we develop DendroMap by adapting Treemaps, a well-known visualization technique. DendroMap effectively organizes images by extracting hierarchical cluster structures from high-dimensional representations of images. It enables users to make sense of the overall distributions of datasets and interactively zoom into specific areas of interests at multiple levels of abstraction. Our case studies with widely-used image datasets for deep learning demonstrate that users can discover insights about datasets and trained models by examining the diversity of images, identifying underperforming subgroups, and analyzing classification errors. We conducted a user study that evaluates the effectiveness of DendroMap in grouping and searching tasks by comparing it with a gridified version of t-SNE and found that participants preferred DendroMap. DendroMap is available at https://div-lab.github.io/dendromap/.

Mobile video-based learning attracts many learners with its mobility and ease of access. However, most lectures are designed for desktops. Our formative study reveals mobile learners’ two major needs: more readable content and customizable video design. To support mobile-optimized learning, we present FitVid, a system that provides responsive and customizable video content. Our system consists of (1) an adaptation pipeline that reverse-engineers pixels to retrieve design elements (e.g., text, images) from videos, leveraging deep learning with a custom dataset, which powers (2) a UI that enables resizing, repositioning, and toggling in-video elements. The content adaptation improves the guideline compliance rate by 24% and 8% for word count and font size. The content evaluation study (n=198) shows that the adaptation significantly increases readability and user satisfaction. The user study (n=31) indicates that FitVid significantly improves learning experience, interactivity, and concentration. We discuss design implications for responsive and customizable video adaptation.

Would you allow an AI agent to make decisions on your behalf? If the answer is “not always,” the next question becomes “in what circumstances”? Answering this question requires human users to be able to assess an AI agent—and not just with overall pass/fail assessments or statistics. Here users need to be able to localize an agent’s bugs so that they can determine when they are willing to rely on the agent and when they are not. After-Action Review for AI (AAR/AI), a new AI assessment process for integration with Explainable AI systems, aims to support human users in this endeavor, and in this article we empirically investigate AAR/AI’s effectiveness with domain-knowledgeable users. Our results show that AAR/AI participants not only located significantly more bugs than non-AAR/AI participants did (i.e., showed greater recall) but also located them more precisely (i.e., with greater precision). In fact, AAR/AI participants outperformed non-AAR/AI participants on every bug and were, on average, almost six times as likely as non-AAR/AI participants to find any particular bug. Finally, evidence suggests that incorporating labeling into the AAR/AI process may encourage domain-knowledgeable users to abstract above individual instances of bugs; we hypothesize that doing so may have contributed further to AAR/AI participants’ effectiveness.

Attention maps are popular tools for explaining the decisions of convolutional neural networks (CNNs) for image classification. Typically, for each image of interest, a single attention map is produced, which assigns weights to pixels based on their importance to the classification. We argue that a single attention map provides an incomplete understanding since there are often many other maps that explain a classification equally well. In this paper, we propose to utilize a beam search algorithm to systematically search for multiple explanations for each image. Results show that there are indeed multiple relatively localized explanations for many images. However, naively showing multiple explanations to users can be overwhelming and does not reveal their common and distinct structures. We introduce structured attention graphs (SAGs), which compactly represent sets of attention maps for an image by visualizing how different combinations of image regions impact the confidence of a classifier. An approach to computing a compact and representative SAG for visualization is proposed via diverse sampling. We conduct a user study comparing the use of SAGs to traditional attention maps for answering comparative counterfactual questions about image classifications. Our results show that the users are significantly more accurate when presented with SAGs compared to standard attention map baselines.

Deep learning's great success motivates many practitioners and students to learn about this exciting technology. However, it is often challenging for beginners to take their first step due to the complexity of understanding and applying deep learning. We present CNN Explainer, an interactive visualization tool designed for non-experts to learn and examine convolutional neural networks (CNNs), a foundational deep learning model architecture. Our tool addresses key challenges that novices face while learning about CNNs, which we identify from interviews with instructors and a survey with past students. CNN Explainer tightly integrates a model overview that summarizes a CNN's structure, and on-demand, dynamic visual explanation views that help users understand the underlying components of CNNs. Through smooth transitions across levels of abstraction, our tool enables users to inspect the interplay between low-level mathematical operations and high-level model structures. A qualitative user study shows that CNN Explainer helps users more easily understand the inner workings of CNNs, and is engaging and enjoyable to use. We also derive design lessons from our study. Developed using modern web technologies, CNN Explainer runs locally in users' web browsers without the need for installation or specialized hardware, broadening the public's education access to modern deep learning techniques.

Deep learning has recently seen rapid development and received significant attention due to its state-of-the-art performance on previously-thought hard problems. However, because of the internal complexity and nonlinear structure of deep neural networks, the underlying decision making processes for why these models are achieving such performance are challenging and sometimes mystifying to interpret. As deep learning spreads across domains, it is of paramount importance that we equip users of deep learning with tools for understanding when a model works correctly, when it fails, and ultimately how to improve its performance. Standardized toolkits for building neural networks have helped democratize deep learning; visual analytics systems have now been developed to support model explanation, interpretation, debugging, and improvement. We present a survey of the role of visual analytics in deep learning research, which highlights its short yet impactful history and thoroughly summarizes the state-of-the-art using a human-centered interrogative framework, focusing on the Five W's and How (Why, Who, What, How, When, and Where). We conclude by highlighting research directions and open research problems. This survey helps researchers and practitioners in both visual analytics and deep learning to quickly learn key aspects of this young and rapidly growing body of research, whose impact spans a diverse range of domains.

Recent success in deep learning has generated immense interest among practitioners and students, inspiring many to learn about this new technology. While visual and interactive approaches have been successfully developed to help people more easily learn deep learning, most existing tools focus on simpler models. In this work, we present GAN Lab, the first interactive visualization tool designed for non-experts to learn and experiment with Generative Adversarial Networks (GANs), a popular class of complex deep learning models. With GAN Lab, users can interactively train generative models and visualize the dynamic training process's intermediate results. GAN Lab tightly integrates an model overview graph that summarizes GAN's structure, and a layered distributions view that helps users interpret the interplay between submodels. GAN Lab introduces new interactive experimentation features for learning complex deep learning models, such as step-by-step training at multiple levels of abstraction for understanding intricate training dynamics. Implemented using TensorFlow.js, GAN Lab is accessible to anyone via modern web browsers, without the need for installation or specialized hardware, overcoming a major practical challenge in deploying interactive tools for deep learning.

The advent of mobile health (mHealth) technologies challenges the capabilities of current visualizations, interactive tools, and algorithms. We present Chronodes, an interactive system that unifies data mining and human-centric visualization techniques to support explorative analysis of longitudinal mHealth data. Chronodes extracts and visualizes frequent event sequences that reveal chronological patterns across multiple participant timelines of mHealth data. It then combines novel interaction and visualization techniques to enable multifocus event sequence analysis, which allows health researchers to interactively define, explore, and compare groups of participant behaviors using event sequence combinations. Through summarizing insights gained from a pilot study with 20 behavioral and biomedical health experts, we discuss Chronodes's efficacy and potential impact in the mHealth domain. Ultimately, we outline important open challenges in mHealth, and offer recommendations and design guidelines for future research.

While deep learning models have achieved state-of-the-art accuracies for many prediction tasks, understanding these models remains a challenge. Despite the recent interest in developing visual tools to help users interpret deep learning models, the complexity and wide variety of models deployed in industry, and the large-scale datasets that they used, pose unique design challenges that are inadequately addressed by existing work. Through participatory design sessions with over 15 researchers and engineers at Facebook, we have developed, deployed, and iteratively improved ACTIVIS, an interactive visualization system for interpreting large-scale deep learning models and results. By tightly integrating multiple coordinated views, such as a computation graph overview of the model architecture, and a neuron activation view for pattern discovery and comparison, users can explore complex deep neural network models at both the instance- and subset-level. ACTIVIS has been deployed on Facebook's machine learning platform. We present case studies with Facebook researchers and engineers, and usage scenarios of how ACTIVIS may work with different models.

Visualization is a powerful paradigm for exploratory data analysis. Visualizing large graphs, however, often results in excessive edges crossings and overlapping nodes. We propose a new scalable approach called Facets that helps users adaptively explore large million-node graphs from a local perspective, guiding them to focus on nodes and neighborhoods that are most subjectively interesting to users. We contribute novel ideas to measure this interestingness in terms of how surprising a neighborhood is given the background distribution, as well as how well it matches what the user has chosen to explore. Facets uses Jensen-Shannon divergence over information-theoretically optimized histograms to calculate the subjective user interest and surprise scores. Participants in a user study found Facets easy to use, easy to learn, and exciting to use. Empirical runtime analyses demonstrated Facets's practical scalability on large real-world graphs with up to 5 million edges, returning results in fewer than 1.5 seconds.

Although researchers have devoted considerable attention to helping database users formulate queries, many users still find it challenging to specify queries that involve joining tables. To help users construct join queries for exploring relational databases, we propose ETable , a novel presentation data model that provides users with a presentation-level interactive view. This view compactly presents one-to-many and many-to-many relationships within a single enriched table by allowing a cell to contain a set of entity references . Users can directly interact with this enriched table to incrementally construct complex queries and navigate databases on a conceptual entity-relationship level. In a user study, participants performed a range of database querying tasks faster with ETable than with a commercial graphical query builder. Subjective feedback about ETable was also positive. All participants found that ETable was easier to learn and helpful for exploring databases.

Health care delivery processes consist of complex activity sequences spanning organizational, spatial, and temporal boundaries. Care is human-directed so these processes can have wide variations in cost, quality, and outcome making systemic care process analysis, conformance testing, and improvement challenging. We designed and developed an interactive visual analytic process exploration and discovery tool and used it to explore clinical data from 5784 pediatric asthma emergency department patients.

Mobile devices have become increasingly central to our everyday activities, due to their portability, multi-touch capabilities, and ever-improving computational power. Such attractive features have spurred research interest in leveraging mobile devices for computation. We explore a novel approach that aims to use a single mobile device to perform scalable graph computation on large graphs that do not fit in the device's limited main memory, opening up the possibility of performing on-device analysis of large datasets, without relying on the cloud. Based on the familiar memory mapping capability provided by today's mobile operating systems, our approach to scale up computation is powerful and intentionally kept simple to maximize its applicability across the iOS and Android platforms. Our experiments demonstrate that an iPad mini can perform fast computation on large real graphs with as many as 272 million edges (Google+ social graph), at a speed that is only a few times slower than a 13” Macbook Pro. Through creating a real world iOS app with this technique, we demonstrate the strong potential application for scalable graph computation on a single mobile device using our approach.

Graph computation approaches such as GraphChi and TurboGraph recently demonstrated that a single PC can perform efficient computation on billion-node graphs. To achieve high speed and scalability, they often need sophisticated data structures and memory management strategies. We propose a minimalist approach that forgoes such requirements, by leveraging the fundamental memory mapping (MMap) capability found on operating systems. We contribute: (1) a new insight that MMap is a viable technique for creating fast and scalable graph algorithms that surpasses some of the best techniques; (2) the design and implementation of popular graph algorithms for billion-scale graphs with little code, thanks to memory mapping; (3) extensive experiments on real graphs, including the 6.6 billion edge Yahoo Web graph, and show that this new approach is significantly faster or comparable to the highly-optimized methods (e.g., 9.5X faster than GraphChi for computing PageRank on 1.47B edge Twitter graph). We believe our work provides a new direction in the design and development of scalable algorithms. Our packaged code is available at http://poloclub.gatech.edu/mmap/.

The field of graph visualization has produced a wealth of visualization techniques for accomplishing a variety of analysis tasks. Therefore analysts often rely on a suite of different techniques, and visual graph analysis application builders strive to provide this breadth of techniques. To provide a holistic model for specifying network visualization techniques (as opposed to considering each technique in isolation) we present the Graph-Level Operations (GLO) model. We describe a method for identifying GLOs and apply it to identify five classes of GLOs, which can be flexibly combined to re-create six canonical graph visualization techniques. We discuss advantages of the GLO model, including potentially discovering new, effective network visualization techniques and easing the engineering challenges of building multi-technique graph visualization applications. Finally, we implement the GLOs that we identified into the GLO-STIX prototype system that enables an analyst to interactively explore a graph by applying GLOs.

In this paper, we present a novel random-walk based node ranking measure, PathRank, which is defined on a heterogeneous graph by extending the Personalized PageRank algorithm. Not only can our proposed measure exploit the semantics behind the different types of nodes and edges in a heterogeneous graph, but also it can emulate various recommendation semantics such as collaborative filtering, content-based filtering, and their combinations. The experimental results show that PathRank can produce more various and effective recommendation results compared to existing approaches.

The field of entity search using Semantic Web (RDF) data has gained more interest recently. In this paper, we propose a probabilistic entity retrieval model for RDF graphs using paths in the graph. Unlike previous work which assumes that all descriptions of an entity are directly linked to the entity node, we assume that an entity can be described with any node that can be reached from the entity node by following paths in the RDF graph. Our retrieval model simulates the generation process of query terms from an entity node by traversing the graph. We evaluate our approach using a standard evaluation framework for entity search.

In many applications, flexibility of recommendation, which is the capability of handling multiple dimensions and various recommendation types, is very important. In this paper, we focus on the flexibility of recommendation and propose a graph-based multidimensional recommendation method. We consider the problem as an entity ranking problem on the graph which is constructed using an implicit feedback dataset (e.g. music listening log), and we adapt Personalized PageRank algorithm to rank entities according to a given query that is represented as a set of entities in the graph. Our model has advantages in that not only can it support the flexibility, but also it can take advantage of exploiting indirect relationships in the graph so that it can perform competitively with the other existing recommendation methods without suffering from the sparsity problem.