James Wexler

Savvas Petridis, Ben Wedin, James Wexler et al.

Large language model (LLM) prompting is a promising new approach for users to create and customize their own chatbots. However, current methods for steering a chatbot's outputs, such as prompt engineering and fine-tuning, do not support users in converting their natural feedback on the model's outputs to changes in the prompt or model. In this work, we explore how to enable users to interactively refine model outputs through their feedback, by helping them convert their feedback into a set of principles (i.e. a constitution) that dictate the model's behavior. From a formative study, we (1) found that users needed support converting their feedback into principles for the chatbot and (2) classified the different principle types desired by users. Inspired by these findings, we developed ConstitutionMaker, an interactive tool for converting user feedback into principles, to steer LLM-based chatbots. With ConstitutionMaker, users can provide either positive or negative feedback in natural language, select auto-generated feedback, or rewrite the chatbot's response; each mode of feedback automatically generates a principle that is inserted into the chatbot's prompt. In a user study with 14 participants, we compare ConstitutionMaker to an ablated version, where users write their own principles. With ConstitutionMaker, participants felt that their principles could better guide the chatbot, that they could more easily convert their feedback into principles, and that they could write principles more efficiently, with less mental demand. ConstitutionMaker helped users identify ways to improve the chatbot, formulate their intuitive responses to the model into feedback, and convert this feedback into specific and clear principles. Together, these findings inform future tools that support the interactive critiquing of LLM outputs.

Kehang Zhu, Nithum Thain, Vivian Tsai et al.

As AI usage becomes more prevalent in social contexts, understanding agent-user interaction is critical to designing systems that improve both individual and group outcomes. We present an online behavioral experiment (N = 243) in which participants play three multi-turn bargaining games in groups of three. Each game, presented in randomized order, grants access to a single LLM assistance modality: proactive recommendations from an Advisor, reactive feedback from a Coach, or autonomous execution by a Delegate; all modalities are powered by an underlying LLM that achieves superhuman performance in an all-agent environment. On each turn, participants privately decide whether to act manually or use the AI modality available in that game. Despite preferring the Advisor modality, participants achieve the highest mean individual gains with the Delegate, demonstrating a preference-performance misalignment. Moreover, delegation generates positive externalities; even non-adopting users in access-to-delegate treatment groups benefit by receiving higher-quality offers. Mechanism analysis reveals that the Delegate agent acts as a market maker, injecting rational, Pareto-improving proposals that restructure the trading environment. Our research reveals a gap between agent capabilities and realized group welfare. While autonomous agents can exhibit super-human strategic performance, their impact on realized welfare gains can be constrained by interfaces, user perceptions, and adoption barriers. Assistance modalities should be designed as mechanisms with endogenous participation; adoption-compatible interaction rules are a prerequisite to improving human welfare with automated assistance.

Ian Tenney, James Wexler, Jasmijn Bastings et al.

We present the Language Interpretability Tool (LIT), an open-source platform for visualization and understanding of NLP models. We focus on core questions about model behavior: Why did my model make this prediction? When does it perform poorly? What happens under a controlled change in the input? LIT integrates local explanations, aggregate analysis, and counterfactual generation into a streamlined, browser-based interface to enable rapid exploration and error analysis. We include case studies for a diverse set of workflows, including exploring counterfactuals for sentiment analysis, measuring gender bias in coreference systems, and exploring local behavior in text generation. LIT supports a wide range of models--including classification, seq2seq, and structured prediction--and is highly extensible through a declarative, framework-agnostic API. LIT is under active development, with code and full documentation available at https://github.com/pair-code/lit.

James Wexler, Mahima Pushkarna, Tolga Bolukbasi et al.

A key challenge in developing and deploying Machine Learning (ML) systems is understanding their performance across a wide range of inputs. To address this challenge, we created the What-If Tool, an open-source application that allows practitioners to probe, visualize, and analyze ML systems, with minimal coding. The What-If Tool lets practitioners test performance in hypothetical situations, analyze the importance of different data features, and visualize model behavior across multiple models and subsets of input data. It also lets practitioners measure systems according to multiple ML fairness metrics. We describe the design of the tool, and report on real-life usage at different organizations.

Marzyeh Ghassemi, Mahima Pushkarna, James Wexler et al.

Making decisions about what clinical tasks to prepare for is multi-factored, and especially challenging in intensive care environments where resources must be balanced with patient needs. Electronic health records (EHRs) are a rich data source, but are task-agnostic and can be difficult to use as summarizations of patient needs for a specific task, such as "could this patient need a ventilator tomorrow?" In this paper, we introduce ClinicalVis, an open-source EHR visualization-based prototype system for task-focused design evaluation of interactions between healthcare providers (HCPs) and EHRs. We situate ClinicalVis in a task-focused proof-of-concept design study targeting these interactions with real patient data. We conduct an empirical study of 14 HCPs, and discuss our findings on usability, accuracy, preference, and confidence in treatment decisions. We also present design implications that our findings suggest for future EHR interfaces, the presentation of clinical data for task-based planning, and evaluating task-focused HCP/EHR interactions in practice.

Minsuk Kahng, Ian Tenney, Mahima Pushkarna et al. · deepmind

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 a large technology company. 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.

Emily Reif, Crystal Qian, James Wexler et al.

Making sense of unstructured text datasets is perennially difficult, yet increasingly relevant with Large Language Models. Data workers 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 workers often run custom analyses for each dataset, which is cumbersome and difficult. We present AutoHistograms, a visualization tool leveragingLLMs. AutoHistograms automatically identifies relevant features, visualizes them with histograms, and allows the user to interactively query the dataset for categories of entities and create new histograms. In a user study with 10 data workers (n=10), we observe that participants can quickly identify insights and explore the data using AutoHistograms, and conceptualize a broad range of applicable use cases. Together, this tool and user study contributeto the growing field of LLM-assisted sensemaking tools.

Savvas Petridis, Ben Wedin, Ann Yuan et al.

Large language models (LLMs) are highly capable at a variety of tasks given the right prompt, but writing one is still a difficult and tedious process. In this work, we introduce ConstitutionalExperts, a method for learning a prompt consisting of constitutional principles (i.e. rules), given a training dataset. Unlike prior methods that optimize the prompt as a single entity, our method incrementally improves the prompt by surgically editing individual principles. We also show that we can improve overall performance by learning unique prompts for different semantic regions of the training data and using a mixture-of-experts (MoE) architecture to route inputs at inference time. We compare our method to other state of the art prompt-optimization techniques across six benchmark datasets. We also investigate whether MoE improves these other techniques. Our results suggest that ConstitutionalExperts outperforms other prompt optimization techniques by 10.9% (F1) and that mixture-of-experts improves all techniques, suggesting its broad applicability.

Crystal Qian, Michael Xieyang Liu, Emily Reif et al.

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.

Crystal Qian, Kehang Zhu, John Horton et al.

As large language models (LLMs) are increasingly embedded in collaborative human activities such as business negotiations and group coordination, it becomes critical to evaluate both the performance gains they can achieve and how they interact in dynamic, multi-agent environments. Unlike traditional statistical agents such as Bayesian models, which may excel under well-specified conditions, large language models (LLMs) can generalize across diverse, real-world scenarios, raising new questions about how their strategies and behaviors compare to those of humans and other agent types. In this work, we compare outcomes and behavioral dynamics across humans (N = 216), LLMs (GPT-4o, Gemini 1.5 Pro), and Bayesian agents in a dynamic negotiation setting under identical conditions. Bayesian agents extract the highest surplus through aggressive optimization, at the cost of frequent trade rejections. Humans and LLMs achieve similar overall surplus, but through distinct behaviors: LLMs favor conservative, concessionary trades with few rejections, while humans employ more strategic, risk-taking, and fairness-oriented behaviors. Thus, we find that performance parity -- a common benchmark in agent evaluation -- can conceal fundamental differences in process and alignment, which are critical for practical deployment in real-world coordination tasks. By establishing foundational behavioral baselines under matched conditions, this work provides a baseline for future studies in more applied, variable-rich environments.

Leon Sixt, Evan Zheran Liu, Marie Pellat et al.

Machine Learning has been successfully applied in systems applications such as memory prefetching and caching, where learned models have been shown to outperform heuristics. However, the lack of understanding the inner workings of these models -- interpretability -- remains a major obstacle for adoption in real-world deployments. Understanding a model's behavior can help system administrators and developers gain confidence in the model, understand risks, and debug unexpected behavior in production. Interpretability for models used in computer systems poses a particular challenge: Unlike ML models trained on images or text, the input domain (e.g., memory access patterns, program counters) is not immediately interpretable. A major challenge is therefore to explain the model in terms of concepts that are approachable to a human practitioner. By analyzing a state-of-the-art caching model, we provide evidence that the model has learned concepts beyond simple statistics that can be leveraged for explanations. Our work provides a first step towards explanability of system ML models and highlights both promises and challenges of this emerging research area.

Jessica Schrouff, Sebastien Baur, Shaobo Hou et al.

Interpretability techniques aim to provide the rationale behind a model's decision, typically by explaining either an individual prediction (local explanation, e.g. 'why is this patient diagnosed with this condition') or a class of predictions (global explanation, e.g. 'why is this set of patients diagnosed with this condition in general'). While there are many methods focused on either one, few frameworks can provide both local and global explanations in a consistent manner. In this work, we combine two powerful existing techniques, one local (Integrated Gradients, IG) and one global (Testing with Concept Activation Vectors), to provide local and global concept-based explanations. We first sanity check our idea using two synthetic datasets with a known ground truth, and further demonstrate with a benchmark natural image dataset. We test our method with various concepts, target classes, model architectures and IG parameters (e.g. baselines). We show that our method improves global explanations over vanilla TCAV when compared to ground truth, and provides useful local insights. Finally, a user study demonstrates the usefulness of the method compared to no or global explanations only. We hope our work provides a step towards building bridges between many existing local and global methods to get the best of both worlds.

Cheng-Zhi Anna Huang, Curtis Hawthorne, Adam Roberts et al.

To make music composition more approachable, we designed the first AI-powered Google Doodle, the Bach Doodle, where users can create their own melody and have it harmonized by a machine learning model Coconet (Huang et al., 2017) in the style of Bach. For users to input melodies, we designed a simplified sheet-music based interface. To support an interactive experience at scale, we re-implemented Coconet in TensorFlow.js (Smilkov et al., 2019) to run in the browser and reduced its runtime from 40s to 2s by adopting dilated depth-wise separable convolutions and fusing operations. We also reduced the model download size to approximately 400KB through post-training weight quantization. We calibrated a speed test based on partial model evaluation time to determine if the harmonization request should be performed locally or sent to remote TPU servers. In three days, people spent 350 years worth of time playing with the Bach Doodle, and Coconet received more than 55 million queries. Users could choose to rate their compositions and contribute them to a public dataset, which we are releasing with this paper. We hope that the community finds this dataset useful for applications ranging from ethnomusicological studies, to music education, to improving machine learning models.

Amirata Ghorbani, James Wexler, James Zou et al.

Interpretability has become an important topic of research as more machine learning (ML) models are deployed and widely used to make important decisions. Most of the current explanation methods provide explanations through feature importance scores, which identify features that are important for each individual input. However, how to systematically summarize and interpret such per sample feature importance scores itself is challenging. In this work, we propose principles and desiderata for \emph{concept} based explanation, which goes beyond per-sample features to identify higher-level human-understandable concepts that apply across the entire dataset. We develop a new algorithm, ACE, to automatically extract visual concepts. Our systematic experiments demonstrate that \alg discovers concepts that are human-meaningful, coherent and important for the neural network's predictions.

Alvin Rajkomar, Eyal Oren, Kai Chen et al.

Predictive modeling with electronic health record (EHR) data is anticipated to drive personalized medicine and improve healthcare quality. Constructing predictive statistical models typically requires extraction of curated predictor variables from normalized EHR data, a labor-intensive process that discards the vast majority of information in each patient's record. We propose a representation of patients' entire, raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format. We demonstrate that deep learning methods using this representation are capable of accurately predicting multiple medical events from multiple centers without site-specific data harmonization. We validated our approach using de-identified EHR data from two U.S. academic medical centers with 216,221 adult patients hospitalized for at least 24 hours. In the sequential format we propose, this volume of EHR data unrolled into a total of 46,864,534,945 data points, including clinical notes. Deep learning models achieved high accuracy for tasks such as predicting in-hospital mortality (AUROC across sites 0.93-0.94), 30-day unplanned readmission (AUROC 0.75-0.76), prolonged length of stay (AUROC 0.85-0.86), and all of a patient's final discharge diagnoses (frequency-weighted AUROC 0.90). These models outperformed state-of-the-art traditional predictive models in all cases. We also present a case-study of a neural-network attribution system, which illustrates how clinicians can gain some transparency into the predictions. We believe that this approach can be used to create accurate and scalable predictions for a variety of clinical scenarios, complete with explanations that directly highlight evidence in the patient's chart.

Been Kim, Martin Wattenberg, Justin Gilmer et al.

The interpretation of deep learning models is a challenge due to their size, complexity, and often opaque internal state. In addition, many systems, such as image classifiers, operate on low-level features rather than high-level concepts. To address these challenges, we introduce Concept Activation Vectors (CAVs), which provide an interpretation of a neural net's internal state in terms of human-friendly concepts. The key idea is to view the high-dimensional internal state of a neural net as an aid, not an obstacle. We show how to use CAVs as part of a technique, Testing with CAVs (TCAV), that uses directional derivatives to quantify the degree to which a user-defined concept is important to a classification result--for example, how sensitive a prediction of "zebra" is to the presence of stripes. Using the domain of image classification as a testing ground, we describe how CAVs may be used to explore hypotheses and generate insights for a standard image classification network as well as a medical application.