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.

LLM leaderboards are widely used to compare models and guide deployment decisions. However, leaderboard rankings are shaped by evaluation priorities set by benchmark designers, rather than by the diverse goals and constraints of actual users and organizations. A single aggregate score often obscures how models behave across different prompt types and compositions. In this work, we conduct an in-depth analysis of the dataset used in the LMArena (formerly Chatbot Arena) benchmark and investigate this evaluation challenge by designing an interactive visualization interface as a design probe. Our analysis reveals that the dataset is heavily skewed toward certain topics, that model rankings vary across prompt slices, and that preference-based judgments are used in ways that blur their intended scope. Building on this analysis, we introduce a visualization interface that allows users to define their own evaluation priorities by selecting and weighting prompt slices and to explore how rankings change accordingly. A qualitative study suggests that this interactive approach improves transparency and supports more context-specific model evaluation, pointing toward alternative ways to design and use LLM leaderboards.

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.

As large language models are deployed in high-stakes enterprise applications, from healthcare to finance, ensuring adherence to organization-specific policies has become essential. Yet existing safety evaluations focus exclusively on universal harms. We present COMPASS (Company/Organization Policy Alignment Assessment), the first systematic framework for evaluating whether LLMs comply with organizational allowlist and denylist policies. We apply COMPASS to eight diverse industry scenarios, generating and validating 5,920 queries that test both routine compliance and adversarial robustness through strategically designed edge cases. Evaluating seven state-of-the-art models, we uncover a fundamental asymmetry: models reliably handle legitimate requests (>95% accuracy) but catastrophically fail at enforcing prohibitions, refusing only 13–40% of adversarial denylist violations. These results demonstrate that current LLMs lack the robustness required for policy-critical deployments, establishing COMPASS as an essential evaluation framework for organizational AI safety.

The tendency of users to anthropomorphise large language models (LLMs) is of growing interest to AI developers, researchers, and policy-makers. Here, we present a novel method for empirically evaluating anthropomorphic LLM behaviours in realistic and varied settings. Going beyond single-turn static benchmarks, we contribute three methodological advances in state-of-the-art (SOTA) LLM evaluation. First, we develop a multi-turn evaluation of 14 anthropomorphic behaviours. Second, we present a scalable, automated approach by employing simulations of user interactions. Third, we conduct an interactive, large-scale human subject study (N=1101) to validate that the model behaviours we measure predict real users' anthropomorphic perceptions. We find that all SOTA LLMs evaluated exhibit similar behaviours, characterised by relationship-building (e.g., empathy and validation) and first-person pronoun use, and that the majority of behaviours only first occur after multiple turns. Our work lays an empirical foundation for investigating how design choices influence anthropomorphic model behaviours and for progressing the ethical debate on the desirability of these behaviours. It also showcases the necessity of multi-turn evaluations for complex social phenomena in human-AI interaction.

Despite the global deployment of text-to-image (T2I) models, their safety frameworks are largely calibrated to a Western-centric default, creating significant vulnerabilities for the rest of the world. To embrace cultural pluralism and bring historically under-represented perspectives in T2I safety, we conduct localised community-centered red teaming studies in the Global South. Our two-fold approach prioritizes localization and participation, by focusing on secondary urban centers in these regions, and conducting community engagement and training workshops to contextualize local norms. As a result, we present PLACES, a dataset comprising over 26,000 examples of T2I model failures collected in partnership with universities in Ghana, Nigeria, and two regions of India (Karnataka and Punjab). Analysis of prompts collected reveals a wide-ranging diversity in socio-cultural and linguistic attributes, when compared to existing geography-agnostic crowdsourced red-teaming data. We observe unique adversarial patterns enabled by local cultural and linguistic nuances, and distinct clusters within region around specific themes, such as religion in India. Moreover, we uncover structural contextual gaps in existing safety frameworks by identifying novel harms showing normative dissonance (e.g., violating religious norms, ignoring local customs, and ominous symbolism). This work argues that expanding T2I safety requires moving beyond mere scale to incorporate deeply localised, participatory methodologies for data collection and contextualization. Content warning: This paper includes examples containing potentially harmful or offensive content.

Current AI safety evaluation and benchmarking frameworks predominantly rely on Western-centric culture-agnostic defaults that mask critical regional laws, socio-linguistic nuances, and cultural taboos, leaving Vision-Language Models (VLMs) vulnerable in global deployments. We introduce Pluralis v0.1: a novel multimodal, multi-regional, and multilingual dataset built from a culture-first perspective. Spanning 6,448 prompts across six Asia-Pacific countries (Bangladesh, India, Korea, Pakistan, Singapore, Taiwan) and eight languages, Pluralis diverges from prior work by natively sourcing localized safety hazards rather than adapting Western datasets. Crucially, it introduces a multimodal evaluation paradigm: user text (e.g., "Should I gift this?") and an image referring to "this" (e.g., a clock) - both innocuous in isolation, but synergistically triggering specific legal or cultural violations. Pluralis disentangles universal safety violations from localized cultural appropriateness, establishing the latter as a first-class evaluation axis. To operationalize this, we present Judge-Pluralis, an agreement-gated LLM-as-a-Judge ensemble trained on examples classified in an empirically derived cultural taxonomy. Observing VLM behavior on a subset of the Pluralis surfaces recurring, locale-specific failure modes such as image misidentifications with downstream harm, missed item-context-locale interactions, and inadequate refusals. These failure modes vary systematically across locales and languages, exposing blind spots that globally averaged metrics conceal. Ultimately, Pluralis is not presented as a solved evaluation framework for cultural alignment, but rather as a first step and catalyst for future innovation. We call upon the research community to utilize this foundation to advance the science of multilingual, multicultural evaluation to better support AI cultural alignment globally.

While prior research on text-to-image generation has predominantly focused on biases in human depictions, we investigate a more subtle yet pervasive phenomenon: demographic bias in generated objects (e.g., cars). We introduce SODA (Stereotyped Object Diagnostic Audit), a novel framework for systematically measuring such biases. Our approach compares visual attributes of objects generated with demographic cues (e.g., "for young people”) to those from neutral prompts, across 2,700 images produced by three state-of-the-art models (GPT Image-1, Imagen 4, and Stable Diffusion) in five object categories. Through a comprehensive analysis, we uncover strong associations between specific demographic groups and visual attributes, such as recurring color patterns prompted by gender or ethnicity cues. These patterns reflect and reinforce not only well-known stereotypes but also more subtle and unintuitive biases. We also observe that some models generate less diverse outputs, which in turn amplifies the visual disparities compared to neutral prompts. Our proposed auditing framework offers a practical approach for testing, revealing how stereotypes still remain embedded in today's generative models. We see this as an essential step toward more systematic and responsible AI 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.

Text-to-image (T2I) models have become prevalent across numerous applications, making their robust evaluation against adversarial attacks a critical priority. Continuous access to new and challenging adversarial prompts across diverse domains is essential for stress-testing these models for resilience against novel attacks from multiple vectors. Current techniques for generating such prompts are either entirely authored by humans or synthetically generated. On the one hand, datasets of human-crafted adversarial prompts are often too small in size and imbalanced in their cultural and contextual representation. On the other hand, datasets of synthetically-generated prompts achieve scale, but typically lack the realistic nuances and creative adversarial strategies found in human-crafted prompts. To combine the strengths of both human and machine approaches, we propose Seed2Harvest, a hybrid red-teaming method for guided expansion of culturally diverse, human-crafted adversarial prompt seeds. The resulting prompts preserve the characteristics and attack patterns of human prompts while maintaining comparable average attack success rates (0.31 NudeNet, 0.36 SD NSFW, 0.12 Q16). Our expanded dataset achieves substantially higher diversity with 535 unique geographic locations and a Shannon entropy of 7.48, compared to 58 locations and 5.28 entropy in the original dataset. Our work demonstrates the importance of human-machine collaboration in leveraging human creativity and machine computational capacity to achieve comprehensive, scalable red-teaming for continuous T2I model safety evaluation.

As large language models (LLMs) grow more proficient at processing unstructured text data, they offer new opportunities to enhance data curation workflows. This paper presents findings from a user study involving 12 industry practitioners from various roles and organizations across a large technology company (N=12). The study examines their data curation workflows before and after LLM adoption, using two custom design probes that integrate LLMs into existing tools. Our study reveals a shift from heuristics-driven, bottom-up curation to insights-driven, top-down workflows supported by LLMs. To navigate increasingly complex data landscapes, practitioners supplement traditional subject-expert-created "golden datasets" with LLM-generated "silver" datasets and rigorously validated "super golden" datasets curated by diverse experts. This research highlights the transformative potential of LLMs in large-scale analysis of unstructured data and highlights opportunities for further tool development.

Automatic side-by-side evaluation has emerged as a promising approach to evaluating the quality of responses from large language models (LLMs). However, analyzing the results from this evaluation approach raises scalability and interpretability challenges. In this paper, we present LLM Comparator, a novel visual analytics tool for interactively analyzing results from automatic side-by-side evaluation. The tool supports interactive workflows for users to understand when and why a model performs better or worse than a baseline model, and how the responses from two models are qualitatively different. We iteratively designed and developed the tool by closely working with researchers and engineers at Google. This paper details the user challenges we identified, the design and development of the tool, and an observational study with participants who regularly evaluate their models.

As large language models (LLMs) grow increasingly adept at processing unstructured text data, they offer new opportunities to enhance data curation workflows. This paper explores the evolution of LLM adoption among practitioners at a large technology company, evaluating the impact of LLMs in data curation tasks through participants' perceptions, integration strategies, and reported usage scenarios. Through a series of surveys, interviews, and user studies, we provide a timely snapshot of how organizations are navigating a pivotal moment in LLM evolution. In Q2 2023, we conducted a survey to assess LLM adoption in industry for development tasks (N=84), and facilitated expert interviews to assess evolving data needs (N=10) in Q3 2023. In Q2 2024, we explored practitioners' current and anticipated LLM usage through a user study involving two LLM-based prototypes (N=12). While each study addressed distinct research goals, they revealed a broader narrative about evolving LLM usage in aggregate. We discovered an emerging shift in data understanding from heuristic-first, bottom-up approaches to insights-first, top-down workflows supported by LLMs. Furthermore, to respond to a more complex data landscape, data practitioners now supplement traditional subject-expert-created 'golden datasets' with LLM-generated 'silver' datasets and rigorously validated 'super golden' datasets curated by diverse experts. This research sheds light on the transformative role of LLMs in large-scale analysis of unstructured data and highlights opportunities for further tool development.

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.

We present Sequence Salience, a visual tool for interactive prompt debugging with input salience methods. Sequence Salience builds on widely used salience methods for text classification and single-token prediction, and extends this to a system tailored for debugging complex LLM prompts. Our system is well-suited for long texts, and expands on previous work by 1) providing controllable aggregation of token-level salience to the word, sentence, or paragraph level, making salience over long inputs tractable; and 2) supporting rapid iteration where practitioners can act on salience results, refine prompts, and run salience on the new output. We include case studies showing how Sequence Salience can help practitioners work with several complex prompting strategies, including few-shot, chain-of-thought, and constitutional principles. Sequence Salience is built on the Learning Interpretability Tool, an open-source platform for ML model visualizations, and code, notebooks, and tutorials are available at http://goo.gle/sequence-salience.

With text-to-image (T2I) generative AI models reaching wide audiences, it is critical to evaluate model robustness against non-obvious attacks to mitigate the generation of offensive images. By focusing on “implicitly adversarial” prompts (those that trigger T2I models to generate unsafe images for non-obvious reasons), we isolate a set of difficult safety issues that human creativity is well-suited to uncover. To this end, we built the Adversarial Nibbler Challenge, a red-teaming methodology for crowdsourcing a diverse set of implicitly adversarial prompts. We have assembled a suite of state-of-the-art T2I models, employed a simple user interface to identify and annotate harms, and engaged diverse populations to capture long-tail safety issues that may be overlooked in standard testing. We present an in-depth account of our methodology, a systematic study of novel attack strategies and safety failures, and a visualization tool for easy exploration of the dataset. The first challenge round resulted in over 10k prompt-image pairs with machine annotations for safety. A subset of 1.5k samples contains rich human annotations of harm types and attack styles. Our findings emphasize the necessity of continual auditing and adaptation as new vulnerabilities emerge. This work will enable proactive, iterative safety assessments and promote responsible development of T2I models.

As large language models (LLMs) become more advanced and impactful, it is increasingly important to scrutinize the data that they rely upon and produce. What is it to be a dataset practitioner doing this work? We approach this in two parts: first, we define the role of "dataset practitioners" by performing a retrospective analysis on the responsibilities of teams contributing to LLM development at a technology company, Google. Then, we conduct semi-structured interviews with a cross-section of these practitioners (N=10). We find that although data quality is a top priority, there is little consensus around what data quality is and how to evaluate it. Consequently, practitioners either rely on their own intuition or write custom code to evaluate their data. We discuss potential reasons for this phenomenon and opportunities for alignment.

Making sense of unstructured text datasets is perennially difficult, yet increasingly relevant with Large Language Models. Data practitioners often rely on dataset summaries, especially distributions of various derived features. Some features, like toxicity or topics, are relevant to many datasets, but many interesting features are domain specific: instruments and genres for a music dataset, or diseases and symptoms for a medical dataset. Accordingly, data practitioners often run custom analyses for each dataset, which is cumbersome and difficult, or use unsupervised methods. We present AutoHistograms, a visualization tool leveraging LLMs. AutoHistograms automatically identifies relevant entity-based features, visualizes them, and allows the user to interactively query the dataset for new categories of entities. In a user study with (n=10) data practitioners, we observe that participants were able to quickly onboard to AutoHistograms, use the tool to identify actionable insights, and conceptualize a broad range of applicable use cases. Together, this tool and user study contribute to the growing field of LLM-assisted sensemaking tools.

In this work, we introduce Gemma 2, a new addition to the Gemma family of lightweight, state-of-the-art open models, ranging in scale from 2 billion to 27 billion parameters. In this new version, we apply several known technical modifications to the Transformer architecture, such as interleaving local-global attentions (Beltagy et al., 2020a) and group-query attention (Ainslie et al., 2023). We also train the 2B and 9B models with knowledge distillation (Hinton et al., 2015) instead of next token prediction. The resulting models deliver the best performance for their size, and even offer competitive alternatives to models that are 2-3 times bigger. We release all our models to the community.