When designing and analyzing a study, researchers must navigate a large space of methodological decisions, or “researcher degrees of freedom.” If these choices are not preregistered or transparently reported, they can increase the risk of inflated false-positive rates and exaggerated effect sizes, undermining scientific credibility. Drawing on psychology research that characterizes these degrees of freedom, we create a protocol for scoring how hypotheses are reported in the HCI literature (ReportDoF). We manually apply ReportDoF to 100 hypotheses from HCI texts authored between 2015-2025, including both preregistrations and papers. Based on this experience, we contribute an LLM workflow and proof-of-concept interactive interface (RDoFlow) that applies ReportDoF to new texts, enabling large-scale analysis of the composition and quality of reported analysis specifications. For example, RDoFlow reveals that HCI research more frequently tests multiple dependent variables for a single hypothesis than psychology research does—a practice that increases the risk of false positives.

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.

Decisions about how to responsibly collect, use and document data often rely upon understanding how people are represented in data. Yet, the unlabeled nature and scale of data used in foundation model development poses a direct challenge to systematic analyses of downstream risks, such as representational harms. We provide a framework designed to help RAI practitioners more easily plan and structure analyses of how people are represented in unstructured data and identify downstream risks. The framework is organized into groups of analyses that map to 3 basic questions: 1) Who is represented in the data, 2) What content is in the data, and 3) How are the two associated. We use the framework to analyze human representation in two commonly used datasets: the Common Crawl web corpus (C4) of 356 billion tokens, and the LAION-400M dataset of 400 million text-image pairs, both developed in the English language. We illustrate how the framework informs action steps for hypothetical teams faced with data use, development, and documentation decisions. Ultimately, the framework structures human representation analyses and maps out analysis planning considerations, goals, and risk mitigation actions at different stages of dataset and model development.

Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, Daphne Ippolito. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Studies show that safety-tuned models may nevertheless divulge harmful information. In this work, we show that whether they do so depends significantly on who they are talking to, which we refer to as *user persona*. In fact, we find manipulating user persona to be more effective for eliciting harmful content than certain more direct attempts to control model refusal. We study both natural language prompting and activation steering as intervention methods and show that activation steering is significantly more effective at bypassing safety filters. We shed light on the mechanics of this phenomenon by showing that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. We also show we can predict a persona’s effect on refusal given only the geometry of its steering vector. Finally, we show that certain user personas induce the model to form more charitable interpretations of otherwise dangerous queries.

Large language models achieve high performance on many but not all downstream tasks.The interaction between pretraining data and task data is commonly assumed to determine this variance: a task with data that is more similar to a model's pretraining data is assumed to be easier for that model.We test whether distributional and example-specific similarity measures (embedding-, token-and model-based) correlate with language model performance through a large-scale comparison of the Pile and C4 pretraining datasets with downstream benchmarks.Similarity correlates with performance for multilingual datasets, but in other benchmarks, we surprisingly find that similarity metrics are not correlated with accuracy or even each other.This suggests that the relationship between pretraining data and downstream tasks is more complex than often assumed.

Large language models have been shown to achieve remarkable performance across a variety of natural language tasks using few-shot learning, which drastically reduces the number of task-specific training examples needed to adapt the model to a particular application. To further our understanding of the impact of scale on few-shot learning, we trained a 540-billion parameter, densely activated, Transformer language model, which we call Pathways Language Model (PaLM). We trained PaLM on 6144 TPU v4 chips using Pathways, a new ML system which enables highly efficient training across multiple TPU Pods. We demonstrate continued benefits of scaling by achieving state-of-the-art few-shot learning results on hundreds of language understanding and generation benchmarks. On a number of these tasks, PaLM 540B achieves breakthrough performance, outperforming the finetuned state-of-the-art on a suite of multi-step reasoning tasks, and outperforming average human performance on the recently released BIG-bench benchmark. A significant number of BIG-bench tasks showed discontinuous improvements from model scale, meaning that performance steeply increased as we scaled to our largest model. PaLM also has strong capabilities in multilingual tasks and source code generation, which we demonstrate on a wide array of benchmarks. We additionally provide a comprehensive analysis on bias and toxicity, and study the extent of training data memorization with respect to model scale. Finally, we discuss the ethical considerations related to large language models and discuss potential mitigation strategies.

The latest generation of large neural language models such as GPT-3 have achieved new levels of performance on benchmarks for language understanding and generation. These models have even demonstrated an ability to perform arbitrary tasks without explicit training. In this work, we sought to learn how people might use such models in the process of creative writing. We built Wordcraft, a text editor in which users collaborate with a generative language model to write a story. We evaluated Wordcraft with a user study in which participants wrote short stories with and without the tool. Our results show that large language models enable novel co-writing experiences. For example, the language model is able to engage in open-ended conversation about the story, respond to writers’ custom requests expressed in natural language (such as ”rewrite this text to be more Dickensian”), and generate suggestions that serve to unblock writers in the creative process. Based on these results, we discuss design implications for future human-AI co-writing systems.

Interpretability of machine learning models is critical to scientific understanding, AI safety, as well as debugging. Attribution is one approach to interpretability, which highlights input dimensions that are influential to a neural network’s prediction. Evaluation of these methods is largely qualitative for image and text models, because acquiring ground truth attributions requires expensive and unreliable human judgment. Attribution has been little studied for graph neural networks (GNNs), a model class of growing importance that makes predictions on arbitrarily-sized graphs. In this work we adapt commonly-used attribution methods for GNNs and quantitatively evaluate them using computable ground-truths that are objective and challenging to learn. We make concrete recommendations for which attribution methods to use, and provide the data and code for our benchmarking suite. Rigorous and open source benchmarking of attribution methods in graphs could enable new methods development and broader use of attribution in real-world ML tasks.

Machine learning (ML) is increasingly being used in image retrieval systems for medical decision making. One application of ML is to retrieve visually similar medical images from past patients (e.g. tissue from biopsies) to reference when making a medical decision with a new patient. However, no algorithm can perfectly capture an expert's ideal notion of similarity for every case: an image that is algorithmically determined to be similar may not be medically relevant to a doctor's specific diagnostic needs. In this paper, we identified the needs of pathologists when searching for similar images retrieved using a deep learning algorithm, and developed tools that empower users to cope with the search algorithm on-the-fly, communicating what types of similarity are most important at different moments in time. In two evaluations with pathologists, we found that these tools increased the diagnostic utility of images found and increased user trust in the algorithm. The tools were preferred over a traditional interface, without a loss in diagnostic accuracy. We also observed that users adopted new strategies when using refinement tools, re-purposing them to test and understand the underlying algorithm and to disambiguate ML errors from their own errors. Taken together, these findings inform future human-ML collaborative systems for expert decision-making.

Transformer architectures show significant promise for natural language processing. Given that a single pretrained model can be fine-tuned to perform well on many different tasks, these networks appear to extract generally useful linguistic features. A natural question is how such networks represent this information internally. This paper describes qualitative and quantitative investigations of one particularly effective model, BERT. At a high level, linguistic features seem to be represented in separate semantic and syntactic subspaces. We find evidence of a fine-grained geometric representation of word senses. We also present empirical descriptions of syntactic representations in both attention matrices and individual word embeddings, as well as a mathematical argument to explain the geometry of these representations.

Solar panels sustainably harvest energy from the sun. To improve performance, panels are often equipped with a tracking mechanism that computes the sun’s position in the sky throughout the day. Based on the tracker’s estimate of the sun’s location, a controller orients the panel to minimize the angle of incidence between solar radiant energy and the photovoltaic cells on the surface of the panel, increasing total energy harvested. Prior work has developed efficient tracking algorithms that accurately compute the sun’s location to facilitate solar tracking and control. However, always pointing a panel directly at the sun does not account for diffuse irradiance in the sky, reflected irradiance from the ground and surrounding surfaces, power required to reorient the panel, shading effects from neighboring panels and foliage, or changing weather conditions (such as clouds), all of which are contributing factors to the total energy harvested by a fleet of solar panels. In this work, we show that a bandit-based approach can increase the total energy harvested by solar panels by learning to dynamically account for such other factors. Our contribution is threefold: (1) the development of a test bed based on typical solar and irradiance models for experimenting with solar panel control using a variety of learning methods, (2) simulated validation that bandit algorithms can effectively learn to control solar panels, and (3) the design and construction of an intelligent solar panel prototype that learns to angle itself using bandit algorithms.