Moritz Hardt

Ali Shirali, Rediet Abebe, Moritz Hardt · berkeley

Dynamic benchmarks interweave model fitting and data collection in an attempt to mitigate the limitations of static benchmarks. In contrast to an extensive theoretical and empirical study of the static setting, the dynamic counterpart lags behind due to limited empirical studies and no apparent theoretical foundation to date. Responding to this deficit, we initiate a theoretical study of dynamic benchmarking. We examine two realizations, one capturing current practice and the other modeling more complex settings. In the first model, where data collection and model fitting alternate sequentially, we prove that model performance improves initially but can stall after only three rounds. Label noise arising from, for instance, annotator disagreement leads to even stronger negative results. Our second model generalizes the first to the case where data collection and model fitting have a hierarchical dependency structure. We show that this design guarantees strictly more progress than the first, albeit at a significant increase in complexity. We support our theoretical analysis by simulating dynamic benchmarks on two popular datasets. These results illuminate the benefits and practical limitations of dynamic benchmarking, providing both a theoretical foundation and a causal explanation for observed bottlenecks in empirical work.

Ali Shirali, Moritz Hardt · berkeley

ImageNet was famously created from Flickr image search results. What if we recreated ImageNet instead by searching the massive LAION dataset based on image captions alone? In this work, we carry out this counterfactual investigation. We find that the resulting ImageNet recreation, which we call LAIONet, looks distinctly unlike the original. Specifically, the intra-class similarity of images in the original ImageNet is dramatically higher than it is for LAIONet. Consequently, models trained on ImageNet perform significantly worse on LAIONet. We propose a rigorous explanation for the discrepancy in terms of a subtle, yet important, difference in two plausible causal data-generating processes for the respective datasets, that we support with systematic experimentation. In a nutshell, searching based on an image caption alone creates an information bottleneck that mitigates the selection bias otherwise present in image-based filtering. Our explanation formalizes a long-held intuition in the community that ImageNet images are stereotypical, unnatural, and overly simple representations of the class category. At the same time, it provides a simple and actionable takeaway for future dataset creation efforts.

Ricardo Dominguez-Olmedo, Vedant Nanda, Rediet Abebe et al.

Annotation and classification of legal text are central components of empirical legal research. Traditionally, these tasks are often delegated to trained research assistants. Motivated by the advances in language modeling, empirical legal scholars are increasingly turning to prompting commercial models, hoping that it will alleviate the significant cost of human annotation. Despite growing use, our understanding of how to best utilize large language models for legal annotation remains limited. To bridge this gap, we introduce CaselawQA, a benchmark comprising 260 legal annotation tasks, nearly all new to the machine learning community. We demonstrate that commercial models, such as GPT-4.5 and Claude 3.7 Sonnet, achieve non-trivial yet highly variable accuracy, generally falling short of the performance required for legal work. We then demonstrate that small, lightly fine-tuned models outperform commercial models. A few hundred to a thousand labeled examples are usually enough to achieve higher accuracy. Our work points to a viable alternative to the predominant practice of prompting commercial models. For concrete legal annotation tasks with some available labeled data, researchers are likely better off using a fine-tuned open-source model.

Ricardo Dominguez-Olmedo, Moritz Hardt, Celestine Mendler-Dünner

Surveys have recently gained popularity as a tool to study large language models. By comparing survey responses of models to those of human reference populations, researchers aim to infer the demographics, political opinions, or values best represented by current language models. In this work, we critically examine this methodology on the basis of the well-established American Community Survey by the U.S. Census Bureau. Evaluating 43 different language models using de-facto standard prompting methodologies, we establish two dominant patterns. First, models' responses are governed by ordering and labeling biases, for example, towards survey responses labeled with the letter "A". Second, when adjusting for these systematic biases through randomized answer ordering, models across the board trend towards uniformly random survey responses, irrespective of model size or pre-training data. As a result, in contrast to conjectures from prior work, survey-derived alignment measures often permit a simple explanation: models consistently appear to better represent subgroups whose aggregate statistics are closest to uniform for any survey under consideration.

Moritz Hardt, Celestine Mendler-Dünner

Predictions in the social world generally influence the target of prediction, a phenomenon known as performativity. Self-fulfilling and self-negating predictions are examples of performativity. Of fundamental importance to economics, finance, and the social sciences, the notion has been absent from the development of machine learning that builds on the static perspective of pattern recognition. In machine learning applications, however, performativity often surfaces as distribution shift. A predictive model deployed on a digital platform, for example, influences behavior and thereby changes the data-generating distribution. We discuss the recently founded area of performative prediction that provides a definition and conceptual framework to study performativity in machine learning. A key element of performative prediction is a natural equilibrium notion that gives rise to new optimization challenges. What emerges is a distinction between learning and steering, two mechanisms at play in performative prediction. Steering is in turn intimately related to questions of power in digital markets. The notion of performative power that we review gives an answer to the question how much a platform can steer participants through its predictions. We end on a discussion of future directions, such as the role that performativity plays in contesting algorithmic systems.

Juan C. Perdomo, Tolani Britton, Moritz Hardt et al.

Early warning systems (EWS) are predictive tools at the center of recent efforts to improve graduation rates in public schools across the United States. These systems assist in targeting interventions to individual students by predicting which students are at risk of dropping out. Despite significant investments in their widespread adoption, there remain large gaps in our understanding of the efficacy of EWS, and the role of statistical risk scores in education. In this work, we draw on nearly a decade's worth of data from a system used throughout Wisconsin to provide the first large-scale evaluation of the long-term impact of EWS on graduation outcomes. We present empirical evidence that the prediction system accurately sorts students by their dropout risk. We also find that it may have caused a single-digit percentage increase in graduation rates, though our empirical analyses cannot reliably rule out that there has been no positive treatment effect. Going beyond a retrospective evaluation of DEWS, we draw attention to a central question at the heart of the use of EWS: Are individual risk scores necessary for effectively targeting interventions? We propose a simple mechanism that only uses information about students' environments -- such as their schools, and districts -- and argue that this mechanism can target interventions just as efficiently as the individual risk score-based mechanism. Our argument holds even if individual predictions are highly accurate and effective interventions exist. In addition to motivating this simple targeting mechanism, our work provides a novel empirical backbone for the robust qualitative understanding among education researchers that dropout is structurally determined. Combined, our insights call into question the marginal value of individual predictions in settings where outcomes are driven by high levels of inequality.

Moritz Hardt, Eric Mazumdar, Celestine Mendler-Dünner et al.

We initiate a principled study of algorithmic collective action on digital platforms that deploy machine learning algorithms. We propose a simple theoretical model of a collective interacting with a firm's learning algorithm. The collective pools the data of participating individuals and executes an algorithmic strategy by instructing participants how to modify their own data to achieve a collective goal. We investigate the consequences of this model in three fundamental learning-theoretic settings: the case of a nonparametric optimal learning algorithm, a parametric risk minimizer, and gradient-based optimization. In each setting, we come up with coordinated algorithmic strategies and characterize natural success criteria as a function of the collective's size. Complementing our theory, we conduct systematic experiments on a skill classification task involving tens of thousands of resumes from a gig platform for freelancers. Through more than two thousand model training runs of a BERT-like language model, we see a striking correspondence emerge between our empirical observations and the predictions made by our theory. Taken together, our theory and experiments broadly support the conclusion that algorithmic collectives of exceedingly small fractional size can exert significant control over a platform's learning algorithm.

André F. Cruz, Moritz Hardt, Celestine Mendler-Dünner

Current question-answering benchmarks predominantly focus on accuracy in realizable prediction tasks. Conditioned on a question and answer-key, does the most likely token match the ground truth? Such benchmarks necessarily fail to evaluate LLMs' ability to quantify ground-truth outcome uncertainty. In this work, we focus on the use of LLMs as risk scores for unrealizable prediction tasks. We introduce folktexts, a software package to systematically generate risk scores using LLMs, and evaluate them against US Census data products. A flexible API enables the use of different prompting schemes, local or web-hosted models, and diverse census columns that can be used to compose custom prediction tasks. We evaluate 17 recent LLMs across five proposed benchmark tasks. We find that zero-shot risk scores produced by multiple-choice question-answering have high predictive signal but are widely miscalibrated. Base models consistently overestimate outcome uncertainty, while instruction-tuned models underestimate uncertainty and produce over-confident risk scores. In fact, instruction-tuning polarizes answer distribution regardless of true underlying data uncertainty. This reveals a general inability of instruction-tuned LLMs to express data uncertainty using multiple-choice answers. A separate experiment using verbalized chat-style risk queries yields substantially improved calibration across instruction-tuned models. These differences in ability to quantify data uncertainty cannot be revealed in realizable settings, and highlight a blind-spot in the current evaluation ecosystem that folktexts covers.

André F. Cruz, Moritz Hardt

Seven years ago, researchers proposed a postprocessing method to equalize the error rates of a model across different demographic groups. The work launched hundreds of papers purporting to improve over the postprocessing baseline. We empirically evaluate these claims through thousands of model evaluations on several tabular datasets. We find that the fairness-accuracy Pareto frontier achieved by postprocessing contains all other methods we were feasibly able to evaluate. In doing so, we address two common methodological errors that have confounded previous observations. One relates to the comparison of methods with different unconstrained base models. The other concerns methods achieving different levels of constraint relaxation. At the heart of our study is a simple idea we call unprocessing that roughly corresponds to the inverse of postprocessing. Unprocessing allows for a direct comparison of methods using different underlying models and levels of relaxation.

Moritz Hardt, Michael P. Kim

When does a machine learning model predict the future of individuals and when does it recite patterns that predate the individuals? In this work, we propose a distinction between these two pathways of prediction, supported by theoretical, empirical, and normative arguments. At the center of our proposal is a family of simple and efficient statistical tests, called backward baselines, that demonstrate if, and to what extent, a model recounts the past. Our statistical theory provides guidance for interpreting backward baselines, establishing equivalences between different baselines and familiar statistical concepts. Concretely, we derive a meaningful backward baseline for auditing a prediction system as a black box, given only background variables and the system's predictions. Empirically, we evaluate the framework on different prediction tasks derived from longitudinal panel surveys, demonstrating the ease and effectiveness of incorporating backward baselines into the practice of machine learning.

Rediet Abebe, Moritz Hardt, Angela Jin et al.

The U.S. criminal legal system increasingly relies on software output to convict and incarcerate people. In a large number of cases each year, the government makes these consequential decisions based on evidence from statistical software -- such as probabilistic genotyping, environmental audio detection, and toolmark analysis tools -- that defense counsel cannot fully cross-examine or scrutinize. This undermines the commitments of the adversarial criminal legal system, which relies on the defense's ability to probe and test the prosecution's case to safeguard individual rights. Responding to this need to adversarially scrutinize output from such software, we propose robust adversarial testing as an audit framework to examine the validity of evidentiary statistical software. We define and operationalize this notion of robust adversarial testing for defense use by drawing on a large body of recent work in robust machine learning and algorithmic fairness. We demonstrate how this framework both standardizes the process for scrutinizing such tools and empowers defense lawyers to examine their validity for instances most relevant to the case at hand. We further discuss existing structural and institutional challenges within the U.S. criminal legal system that may create barriers for implementing this and other such audit frameworks and close with a discussion on policy changes that could help address these concerns.

Nikhil Chandak, Shashwat Goel, Ameya Prabhu et al.

High-stakes decision making involves reasoning under uncertainty about the future. In this work, we train language models to make predictions on open-ended forecasting questions. To scale up training data, we synthesize novel forecasting questions from global events reported in daily news, using a fully automated, careful curation recipe. We train the Qwen3 thinking models on our dataset, OpenForesight. To prevent leakage of future information during training and evaluation, we use an offline news corpus, both for data generation and retrieval in our forecasting system. Guided by a small validation set, we show the benefits of retrieval, and an improved reward function for reinforcement learning (RL). Once we obtain our final forecasting system, we perform held-out testing between May to August 2025. Our specialized model, OpenForecaster 8B, matches much larger proprietary models, with our training improving the accuracy, calibration, and consistency of predictions. We find calibration improvements from forecasting training generalize across popular benchmarks. We open-source all our models, code, and data to make research on language model forecasting broadly accessible.

Gary Cheng, Moritz Hardt, Celestine Mendler-Dünner

Regulators and academics are increasingly interested in the causal effect that algorithmic actions of a digital platform have on consumption. We introduce a general causal inference problem we call the steerability of consumption that abstracts many settings of interest. Focusing on observational designs and exploiting the structure of the problem, we exhibit a set of assumptions for causal identifiability that significantly weaken the often unrealistic overlap assumptions of standard designs. The key novelty of our approach is to explicitly model the dynamics of consumption over time, viewing the platform as a controller acting on a dynamical system. From this dynamical systems perspective, we are able to show that exogenous variation in consumption and appropriately responsive algorithmic control actions are sufficient for identifying steerability of consumption. Our results illustrate the fruitful interplay of control theory and causal inference, which we illustrate with examples from econometrics, macroeconomics, and machine learning.

Mina Remeli, Moritz Hardt, Robert C. Williamson

Our paper studies the predictability of online speech -- that is, how well language models learn to model the distribution of user generated content on X (previously Twitter). We define predictability as a measure of the model's uncertainty, i.e. its negative log-likelihood. As the basis of our study, we collect 10M tweets for ``tweet-tuning'' base models and a further 6.25M posts from more than five thousand X (previously Twitter) users and their peers. In our study involving more than 5000 subjects, we find that predicting posts of individual users remains surprisingly hard. Moreover, it matters greatly what context is used: models using the users' own history significantly outperform models using posts from their social circle. We validate these results across four large language models ranging in size from 1.5 billion to 70 billion parameters. Moreover, our results replicate if instead of prompting the model with additional context, we finetune on it. We follow up with a detailed investigation on what is learned in-context and a demographic analysis. Up to 20\% of what is learned in-context is the use of @-mentions and hashtags. Our main results hold across the demographic groups we studied.

Ricardo Dominguez-Olmedo, Florian E. Dorner, Moritz Hardt

We study a fundamental problem in the evaluation of large language models that we call training on the test task. Unlike wrongful practices like training on the test data, leakage, or data contamination, training on the test task is not a malpractice. Rather, the term describes a growing set of practices that utilize knowledge about evaluation tasks at training time. We demonstrate that training on the test task confounds both relative model evaluations and claims about emergent capabilities. We argue that the seeming superiority of one model family over another may be explained by a different degree of training on the test task. To this end, we propose an effective method to adjust for the effect of training on the test task on benchmark evaluations. Put simply, to fine-tune each model under comparison on the same task-relevant data prior to evaluation. We then show that instances of emergent behavior disappear gradually as models train on the test task. Our work promotes a new perspective on the evaluation of large language models, with broad implications for benchmarking and the study of emergent capabilities.

Ana-Andreea Stoica, Celestine Mendler-Duenner, Moritz Hardt

Digital labor platforms are increasingly used to procure human input, ranging from annotating data and red-teaming AI models, to ride-sharing and food delivery. A central concern in such markets is the ability of platforms to suppress wages by exploiting the abundance of low-cost labor. To study this exploitation pattern, we introduce a novel posted-price procurement model with coverage objectives. A platform seeks to complete M tasks by posting prices to sequentially arriving workers, each of whom accepts a task if it exceeds their private cost. First, we show that under natural assumptions on the workers' estimated cost, there exists a simple pricing strategy for the platform to cover all M tasks with wait time O(M), while paying only a O(log(M)/M) fraction of the total cost of labor. This result highlights how platforms can exploit workers' uncertainty about the cost of labor to effectively suppress wages. Then, we study collective action as a lever to increase wages and promote welfare in digital labor markets. In particular, we show how a small coalition of targeted low-cost workers who commit to a price floor forces the platform's total spending from logarithmic to linear in M. In contrast, a randomly sampled coalition of equal size remains largely ineffective. We complement our theory with synthetic experiments, showcasing the benefits of collective action across different market regimes.

Shashwat Goel, Nikhil Chandak, Arvindh Arun et al.

AI agents are being increasingly deployed in dynamic, open-ended environments that require adapting to new information as it arrives. To efficiently measure this capability for realistic use-cases, we propose building grounded simulations that replay real-world events in the order they occurred. We build FutureSim, where agents forecast world events beyond their knowledge cutoff while interacting with a chronological replay of the world: real news articles arriving and questions resolving over the simulated period. We evaluate frontier agents in their native harness, testing their ability to predict world events over a three-month period from January to March 2026. FutureSim reveals a clear separation in their capabilities, with the best agent's accuracy being 25%, and many having worse Brier skill score than making no prediction at all. Through careful ablations, we show how FutureSim offers a realistic setting to study emerging research directions like long-horizon test-time adaptation, search, memory, and reasoning about uncertainty. Overall, we hope our benchmark design paves the way to measure AI progress on open-ended adaptation spanning long time-horizons in the real world.

Frances Ding, Moritz Hardt, John Miller et al.

Although the fairness community has recognized the importance of data, researchers in the area primarily rely on UCI Adult when it comes to tabular data. Derived from a 1994 US Census survey, this dataset has appeared in hundreds of research papers where it served as the basis for the development and comparison of many algorithmic fairness interventions. We reconstruct a superset of the UCI Adult data from available US Census sources and reveal idiosyncrasies of the UCI Adult dataset that limit its external validity. Our primary contribution is a suite of new datasets derived from US Census surveys that extend the existing data ecosystem for research on fair machine learning. We create prediction tasks relating to income, employment, health, transportation, and housing. The data span multiple years and all states of the United States, allowing researchers to study temporal shift and geographic variation. We highlight a broad initial sweep of new empirical insights relating to trade-offs between fairness criteria, performance of algorithmic interventions, and the role of distribution shift based on our new datasets. Our findings inform ongoing debates, challenge some existing narratives, and point to future research directions. Our datasets are available at https://github.com/zykls/folktables.

Ricardo Olmedo, Bernhard Schölkopf, Moritz Hardt

Consider a market of competing model providers selling query access to models with varying costs and capabilities. Customers submit problem instances and are willing to pay up to a budget for a verifiable solution. An arbitrageur efficiently allocates inference budget across providers to undercut the market, thus creating a competitive offering with no model-development risk. In this work, we initiate the study of arbitrage in AI model markets, empirically demonstrating the viability of arbitrage and illustrating its economic consequences. We conduct an in-depth case study of SWE-bench GitHub issue resolution using two representative models, GPT-5 mini and DeepSeek v3.2. In this verifiable domain, simple arbitrage strategies generate net profit margins of up to 40%. Robust arbitrage strategies that generalize across different domains remain profitable. Distillation further creates strong arbitrage opportunities, potentially at the expense of the teacher model's revenue. Multiple competing arbitrageurs drive down consumer prices, reducing the marginal revenue of model providers. At the same time, arbitrage reduces market segmentation and facilitates market entry for smaller model providers by enabling earlier revenue capture. Our results suggest that arbitrage can be a powerful force in AI model markets with implications for model development, distillation, and deployment.

Florian E. Dorner, Vivian Y. Nastl, Moritz Hardt

High quality annotations are increasingly a bottleneck in the explosively growing machine learning ecosystem. Scalable evaluation methods that avoid costly annotation have therefore become an important research ambition. Many hope to use strong existing models in lieu of costly labels to provide cheap model evaluations. Unfortunately, this method of using models as judges introduces biases, such as self-preferencing, that can distort model comparisons. An emerging family of debiasing tools promises to fix these issues by using a few high quality labels to debias a large number of model judgments. In this paper, we study how far such debiasing methods, in principle, can go. Our main result shows that when the judge is no more accurate than the evaluated model, no debiasing method can decrease the required amount of ground truth labels by more than half. Our result speaks to the severe limitations of the LLM-as-a-judge paradigm at the evaluation frontier where the goal is to assess newly released models that are possibly better than the judge. Through an empirical evaluation, we demonstrate that the sample size savings achievable in practice are even more modest than what our theoretical limit suggests. Along the way, our work provides new observations about debiasing methods for model evaluation, and points out promising avenues for future work.

Guanhua Zhang, Moritz Hardt

We examine multi-task benchmarks in machine learning through the lens of social choice theory. We draw an analogy between benchmarks and electoral systems, where models are candidates and tasks are voters. This suggests a distinction between cardinal and ordinal benchmark systems. The former aggregate numerical scores into one model ranking; the latter aggregate rankings for each task. We apply Arrow's impossibility theorem to ordinal benchmarks to highlight the inherent limitations of ordinal systems, particularly their sensitivity to the inclusion of irrelevant models. Inspired by Arrow's theorem, we empirically demonstrate a strong trade-off between diversity and sensitivity to irrelevant changes in existing multi-task benchmarks. Our result is based on new quantitative measures of diversity and sensitivity that we introduce. Sensitivity quantifies the impact that irrelevant changes to tasks have on a benchmark. Diversity captures the degree of disagreement in model rankings across tasks. We develop efficient approximation algorithms for both measures, as exact computation is computationally challenging. Through extensive experiments on seven cardinal benchmarks and eleven ordinal benchmarks, we demonstrate a clear trade-off between diversity and stability: The more diverse a multi-task benchmark, the more sensitive to trivial changes it is. Additionally, we show that the aggregated rankings of existing benchmarks are highly unstable under irrelevant changes. The codes and data are available at https://socialfoundations.github.io/benchbench/.

Vivian Y. Nastl, Moritz Hardt

We study how well machine learning models trained on causal features generalize across domains. We consider 16 prediction tasks on tabular datasets covering applications in health, employment, education, social benefits, and politics. Each dataset comes with multiple domains, allowing us to test how well a model trained in one domain performs in another. For each prediction task, we select features that have a causal influence on the target of prediction. Our goal is to test the hypothesis that models trained on causal features generalize better across domains. Without exception, we find that predictors using all available features, regardless of causality, have better in-domain and out-of-domain accuracy than predictors using causal features. Moreover, even the absolute drop in accuracy from one domain to the other is no better for causal predictors than for models that use all features. In addition, we show that recent causal machine learning methods for domain generalization do not perform better in our evaluation than standard predictors trained on the set of causal features. Likewise, causal discovery algorithms either fail to run or select causal variables that perform no better than our selection. Extensive robustness checks confirm that our findings are stable under variable misclassification.

Olawale Salaudeen, Moritz Hardt

We introduce ImageNot, a dataset designed to match the scale of ImageNet while differing drastically in other aspects. We show that key model architectures developed for ImageNet over the years rank identically when trained and evaluated on ImageNot to how they rank on ImageNet. This is true when training models from scratch or fine-tuning them. Moreover, the relative improvements of each model over earlier models strongly correlate in both datasets. We further give evidence that ImageNot has a similar utility as ImageNet for transfer learning purposes. Our work demonstrates a surprising degree of external validity in the relative performance of image classification models. This stands in contrast with absolute accuracy numbers that typically drop sharply even under small changes to a dataset.

Nikhil Chandak, Shashwat Goel, Ameya Prabhu et al.

Multiple choice benchmarks have long been the workhorse of language model evaluation because grading multiple choice is objective and easy to automate. However, we show multiple choice questions from popular benchmarks can often be answered without even seeing the question. These shortcuts arise from a fundamental limitation of discriminative evaluation not shared by evaluations of the model's free-form, generative answers. Until recently, there appeared to be no viable, scalable alternative to multiple choice--but, we show that this has changed. We consider generative evaluation via what we call answer matching: Give the candidate model the question without the options, have it generate a free-form response, then use a modern language model with the reference answer to determine if the response matches the reference. To compare the validity of different evaluation strategies, we annotate MMLU-Pro and GPQA-Diamond to obtain human grading data, and measure the agreement of each evaluation approach. We find answer matching using recent models--even small ones--achieves near-perfect agreement, in the range of inter-annotator agreement. In contrast, both multiple choice evaluation and using LLM-as-a-judge without reference answers aligns poorly with human grading. Improving evaluations via answer matching is not merely a conceptual concern: the rankings of several models change significantly when evaluating their free-form responses with answer matching. In light of these findings, we discuss how to move the evaluation ecosystem from multiple choice to answer matching.

Florian E. Dorner, Moritz Hardt

We study how to best spend a budget of noisy labels to compare the accuracy of two binary classifiers. It's common practice to collect and aggregate multiple noisy labels for a given data point into a less noisy label via a majority vote. We prove a theorem that runs counter to conventional wisdom. If the goal is to identify the better of two classifiers, we show it's best to spend the budget on collecting a single label for more samples. Our result follows from a non-trivial application of Cramér's theorem, a staple in the theory of large deviations. We discuss the implications of our work for the design of machine learning benchmarks, where they overturn some time-honored recommendations. In addition, our results provide sample size bounds superior to what follows from Hoeffding's bound.

Sílvia Casacuberta, Moritz Hardt

Conditional average treatment effect (CATE) estimation is the de facto gold standard for targeting a treatment to a heterogeneous population. The method estimates treatment effects up to an error $ε> 0$ in each of $M$ different strata of the population, targeting individuals in decreasing order of estimated treatment effect until the budget runs out. In general, this method requires $O(M/ε^2)$ samples. This is best possible if the goal is to estimate all treatment effects up to an $ε$ error. In this work, we show how to achieve the same total treatment effect as CATE with only $O(M/ε)$ samples for natural distributions of treatment effects. The key insight is that coarse estimates suffice for near-optimal treatment allocations. In addition, we show that budget flexibility can further reduce the sample complexity of allocation. Finally, we evaluate our algorithm on various real-world RCT datasets. In all cases, it finds nearly optimal treatment allocations with surprisingly few samples. Our work highlights the fundamental distinction between treatment effect estimation and treatment allocation: the latter requires far fewer samples.

Guanhua Zhang, Ricardo Dominguez-Olmedo, Moritz Hardt

Existing language model benchmarks provide contradictory model rankings, even for benchmarks that aim to capture similar skills. This dilemma of conflicting rankings hampers model selection, clouds model comparisons, and adds confusion to a growing ecosystem of competing models. In this paper, we take a different perspective on model comparison: instead of relying on out-of-the-box performance via direct evaluation, we compare model potential by providing each model with identical benchmark-specific fine-tuning before evaluation. We call this approach train-before-test. Our primary contribution is a comprehensive empirical evaluation of model potential across 24 benchmarks and 61 models. First, we demonstrate that model potential rankings obtained through train-before-test exhibit remarkable consistency across all benchmarks. Whereas traditional rankings demonstrate little external validity under direct evaluation, they enjoy a significant degree of external validity when applying train-before-test: model potential rankings transfer gracefully from one benchmark to another. Second, train-before-test restores the connection between perplexity and downstream task performance, lost under direct evaluation. Remarkably, even pre-finetuning perplexity of a base model predicts post-finetuning downstream performance, suggesting that ranking consistency reflects inherent model potential rather than fine-tuning artifacts. Finally, train-before-test reduces the model-score matrix to essentially rank one, indicating that model potential is dominated by one latent factor, uncovered by train-before-test. Our work supports the recommendation to make train-before-test a default component of LLM benchmarking.

Guanhua Zhang, Florian E. Dorner, Moritz Hardt

Large language model (LLM) evaluation is increasingly costly, prompting interest in methods that speed up evaluation by shrinking benchmark datasets. Benchmark prediction (also called efficient LLM evaluation) aims to select a small subset of evaluation points and predict overall benchmark performance from that subset. In this paper, we systematically assess the strengths and limitations of 11 benchmark prediction methods across 19 diverse benchmarks. First, we identify a highly competitive baseline: Take a random sample and fit a regression model on the sample to predict missing entries. Outperforming most existing methods, this baseline challenges the assumption that careful subset selection is necessary for benchmark prediction. Second, we discover that all existing methods crucially depend on model similarity. They work best when interpolating scores among similar models. The effectiveness of benchmark prediction sharply declines when new models have higher accuracy than previously seen models. In this setting of extrapolation, none of the previous methods consistently beat a simple average over random samples. To improve over the sample average, we introduce a new method inspired by augmented inverse propensity weighting. This method consistently outperforms the random sample average even for extrapolation. However, its performance still relies on model similarity and the gains are modest in general. This shows that benchmark prediction fails just when it is most needed: at the evaluation frontier, where the goal is to evaluate new models of unknown capabilities.

Jonas Hübotter, Leander Diaz-Bone, Ido Hakimi et al.

Humans are good at learning on the job: We learn how to solve the tasks we face as we go along. Can a model do the same? We propose an agent that assembles a task-specific curriculum, called test-time curriculum (TTC-RL), and applies reinforcement learning to continue training the model for its target task. The test-time curriculum avoids time-consuming human curation of datasets by automatically selecting the most task-relevant data from a large pool of available training data. Our experiments demonstrate that reinforcement learning on a test-time curriculum consistently improves the model on its target tasks, across a variety of evaluations and models. Notably, on challenging math and coding benchmarks, TTC-RL improves the pass@1 of Qwen3-8B by approximately 1.8x on AIME25 and 2.1x on CodeElo. Moreover, we find that TTC-RL significantly raises the performance ceiling compared to the initial model, increasing pass@8 on AIME25 from 40% to 62% and on CodeElo from 28% to 43%. Our findings show the potential of test-time curricula in extending the test-time scaling paradigm to continual training on thousands of task-relevant experiences during test-time.

Ali Shirali, Rediet Abebe, Moritz Hardt

Algorithmic predictions are emerging as a promising solution concept for efficiently allocating societal resources. Fueling their use is an underlying assumption that such systems are necessary to identify individuals for interventions. We propose a principled framework for assessing this assumption: Using a simple mathematical model, we evaluate the efficacy of prediction-based allocations in settings where individuals belong to larger units such as hospitals, neighborhoods, or schools. We find that prediction-based allocations outperform baseline methods using aggregate unit-level statistics only when between-unit inequality is low and the intervention budget is high. Our results hold for a wide range of settings for the price of prediction, treatment effect heterogeneity, and unit-level statistics' learnability. Combined, we highlight the potential limits to improving the efficacy of interventions through prediction.

Moritz Hardt, Yu Sun

Many recent efforts augment language models with retrieval, by adding retrieved data to the input context. For this approach to succeed, the retrieved data must be added at both training and test time. Moreover, as input length grows linearly with the size of retrieved data, cost in computation and memory grows quadratically for modern Transformers. To avoid these complications, we simply fine-tune the model on retrieved data at test time, using its standard training setup. We build a large-scale distributed index based on text embeddings of the Pile dataset. For each test input, our system retrieves its neighbors and fine-tunes the model on their text. Surprisingly, retrieving and training on as few as 20 neighbors, each for only one gradient iteration, drastically improves performance across more than 20 language modeling tasks in the Pile. For example, test-time training with nearest neighbors significantly narrows the performance gap between a small GPT-2 and a GPT-Neo model more than 10 times larger. Sufficient index quality and size, however, are necessary. Our work establishes a first baseline of test-time training for language modeling.

Moritz Hardt, Meena Jagadeesan, Celestine Mendler-Dünner

We introduce the notion of performative power, which measures the ability of a firm operating an algorithmic system, such as a digital content recommendation platform, to cause change in a population of participants. We relate performative power to the economic study of competition in digital economies. Traditional economic concepts struggle with identifying anti-competitive patterns in digital platforms not least due to the complexity of market definition. In contrast, performative power is a causal notion that is identifiable with minimal knowledge of the market, its internals, participants, products, or prices. Low performative power implies that a firm can do no better than to optimize their objective on current data. In contrast, firms of high performative power stand to benefit from steering the population towards more profitable behavior. We confirm in a simple theoretical model that monopolies maximize performative power. A firm's ability to personalize increases performative power, while competition and outside options decrease performative power. On the empirical side, we propose an observational causal design to identify performative power from discontinuities in how digital platforms display content. This allows to repurpose causal effects from various studies about digital platforms as lower bounds on performative power. Finally, we speculate about the role that performative power might play in competition policy and antitrust enforcement in digital marketplaces.

Smitha Milli, Luca Belli, Moritz Hardt

Online platforms regularly conduct randomized experiments to understand how changes to the platform causally affect various outcomes of interest. However, experimentation on online platforms has been criticized for having, among other issues, a lack of meaningful oversight and user consent. As platforms give users greater agency, it becomes possible to conduct observational studies in which users self-select into the treatment of interest as an alternative to experiments in which the platform controls whether the user receives treatment or not. In this paper, we conduct four large-scale within-study comparisons on Twitter aimed at assessing the effectiveness of observational studies derived from user self-selection on online platforms. In a within-study comparison, treatment effects from an observational study are assessed based on how effectively they replicate results from a randomized experiment with the same target population. We test the naive difference in group means estimator, exact matching, regression adjustment, and inverse probability of treatment weighting while controlling for plausible confounding variables. In all cases, all observational estimates perform poorly at recovering the ground-truth estimate from the analogous randomized experiments. In all cases except one, the observational estimates have the opposite sign of the randomized estimate. Our results suggest that observational studies derived from user self-selection are a poor alternative to randomized experimentation on online platforms. In discussing our results, we postulate a "Catch-22" that suggests that the success of causal inference in these settings may be at odds with the original motivations for providing users with greater agency.

Meena Jagadeesan, Celestine Mendler-Dünner, Moritz Hardt

When reasoning about strategic behavior in a machine learning context it is tempting to combine standard microfoundations of rational agents with the statistical decision theory underlying classification. In this work, we argue that a direct combination of these standard ingredients leads to brittle solution concepts of limited descriptive and prescriptive value. First, we show that rational agents with perfect information produce discontinuities in the aggregate response to a decision rule that we often do not observe empirically. Second, when any positive fraction of agents is not perfectly strategic, desirable stable points -- where the classifier is optimal for the data it entails -- cease to exist. Third, optimal decision rules under standard microfoundations maximize a measure of negative externality known as social burden within a broad class of possible assumptions about agent behavior. Recognizing these limitations we explore alternatives to standard microfoundations for binary classification. We start by describing a set of desiderata that help navigate the space of possible assumptions about how agents respond to a decision rule. In particular, we analyze a natural constraint on feature manipulations, and discuss properties that are sufficient to guarantee the robust existence of stable points. Building on these insights, we then propose the noisy response model. Inspired by smoothed analysis and empirical observations, noisy response incorporates imperfection in the agent responses, which we show mitigates the limitations of standard microfoundations. Our model retains analytical tractability, leads to more robust insights about stable points, and imposes a lower social burden at optimality.

Moritz Hardt, Benjamin Recht

This graduate textbook on machine learning tells a story of how patterns in data support predictions and consequential actions. Starting with the foundations of decision making, we cover representation, optimization, and generalization as the constituents of supervised learning. A chapter on datasets as benchmarks examines their histories and scientific bases. Self-contained introductions to causality, the practice of causal inference, sequential decision making, and reinforcement learning equip the reader with concepts and tools to reason about actions and their consequences. Throughout, the text discusses historical context and societal impact. We invite readers from all backgrounds; some experience with probability, calculus, and linear algebra suffices.

Chloe Ching-Yun Hsu, Celestine Mendler-Dünner, Moritz Hardt

Proximal Policy Optimization (PPO) is a popular deep policy gradient algorithm. In standard implementations, PPO regularizes policy updates with clipped probability ratios, and parameterizes policies with either continuous Gaussian distributions or discrete Softmax distributions. These design choices are widely accepted, and motivated by empirical performance comparisons on MuJoCo and Atari benchmarks. We revisit these practices outside the regime of current benchmarks, and expose three failure modes of standard PPO. We explain why standard design choices are problematic in these cases, and show that alternative choices of surrogate objectives and policy parameterizations can prevent the failure modes. We hope that our work serves as a reminder that many algorithmic design choices in reinforcement learning are tied to specific simulation environments. We should not implicitly accept these choices as a standard part of a more general algorithm.

Smitha Milli, Luca Belli, Moritz Hardt

Most recommendation engines today are based on predicting user engagement, e.g. predicting whether a user will click on an item or not. However, there is potentially a large gap between engagement signals and a desired notion of "value" that is worth optimizing for. We use the framework of measurement theory to (a) confront the designer with a normative question about what the designer values, (b) provide a general latent variable model approach that can be used to operationalize the target construct and directly optimize for it, and (c) guide the designer in evaluating and revising their operationalization. We implement our approach on the Twitter platform on millions of users. In line with established approaches to assessing the validity of measurements, we perform a qualitative evaluation of how well our model captures a desired notion of "value".

Celestine Mendler-Dünner, Juan C. Perdomo, Tijana Zrnic et al.

In performative prediction, the choice of a model influences the distribution of future data, typically through actions taken based on the model's predictions. We initiate the study of stochastic optimization for performative prediction. What sets this setting apart from traditional stochastic optimization is the difference between merely updating model parameters and deploying the new model. The latter triggers a shift in the distribution that affects future data, while the former keeps the distribution as is. Assuming smoothness and strong convexity, we prove rates of convergence for both greedily deploying models after each stochastic update (greedy deploy) as well as for taking several updates before redeploying (lazy deploy). In both cases, our bounds smoothly recover the optimal $O(1/k)$ rate as the strength of performativity decreases. Furthermore, they illustrate how depending on the strength of performative effects, there exists a regime where either approach outperforms the other. We experimentally explore the trade-off on both synthetic data and a strategic classification simulator.

Esther Rolf, Max Simchowitz, Sarah Dean et al.

While real-world decisions involve many competing objectives, algorithmic decisions are often evaluated with a single objective function. In this paper, we study algorithmic policies which explicitly trade off between a private objective (such as profit) and a public objective (such as social welfare). We analyze a natural class of policies which trace an empirical Pareto frontier based on learned scores, and focus on how such decisions can be made in noisy or data-limited regimes. Our theoretical results characterize the optimal strategies in this class, bound the Pareto errors due to inaccuracies in the scores, and show an equivalence between optimal strategies and a rich class of fairness-constrained profit-maximizing policies. We then present empirical results in two different contexts -- online content recommendation and sustainable abalone fisheries -- to underscore the applicability of our approach to a wide range of practical decisions. Taken together, these results shed light on inherent trade-offs in using machine learning for decisions that impact social welfare.

Juan C. Perdomo, Tijana Zrnic, Celestine Mendler-Dünner et al.

When predictions support decisions they may influence the outcome they aim to predict. We call such predictions performative; the prediction influences the target. Performativity is a well-studied phenomenon in policy-making that has so far been neglected in supervised learning. When ignored, performativity surfaces as undesirable distribution shift, routinely addressed with retraining. We develop a risk minimization framework for performative prediction bringing together concepts from statistics, game theory, and causality. A conceptual novelty is an equilibrium notion we call performative stability. Performative stability implies that the predictions are calibrated not against past outcomes, but against the future outcomes that manifest from acting on the prediction. Our main results are necessary and sufficient conditions for the convergence of retraining to a performatively stable point of nearly minimal loss. In full generality, performative prediction strictly subsumes the setting known as strategic classification. We thus also give the first sufficient conditions for retraining to overcome strategic feedback effects.

John Miller, Smitha Milli, Moritz Hardt

Consequential decision-making incentivizes individuals to strategically adapt their behavior to the specifics of the decision rule. While a long line of work has viewed strategic adaptation as gaming and attempted to mitigate its effects, recent work has instead sought to design classifiers that incentivize individuals to improve a desired quality. Key to both accounts is a cost function that dictates which adaptations are rational to undertake. In this work, we develop a causal framework for strategic adaptation. Our causal perspective clearly distinguishes between gaming and improvement and reveals an important obstacle to incentive design. We prove any procedure for designing classifiers that incentivize improvement must inevitably solve a non-trivial causal inference problem. Moreover, we show a similar result holds for designing cost functions that satisfy the requirements of previous work. With the benefit of hindsight, our results show much of the prior work on strategic classification is causal modeling in disguise.

Yu Sun, Xiaolong Wang, Zhuang Liu et al.

In this paper, we propose Test-Time Training, a general approach for improving the performance of predictive models when training and test data come from different distributions. We turn a single unlabeled test sample into a self-supervised learning problem, on which we update the model parameters before making a prediction. This also extends naturally to data in an online stream. Our simple approach leads to improvements on diverse image classification benchmarks aimed at evaluating robustness to distribution shifts.

Chloe Ching-Yun Hsu, Michaela Hardt, Moritz Hardt

Linear dynamical systems are a fundamental and powerful parametric model class. However, identifying the parameters of a linear dynamical system is a venerable task, permitting provably efficient solutions only in special cases. This work shows that the eigenspectrum of unknown linear dynamics can be identified without full system identification. We analyze a computationally efficient and provably convergent algorithm to estimate the eigenvalues of the state-transition matrix in a linear dynamical system. When applied to time series clustering, our algorithm can efficiently cluster multi-dimensional time series with temporal offsets and varying lengths, under the assumption that the time series are generated from linear dynamical systems. Evaluating our algorithm on both synthetic data and real electrocardiogram (ECG) signals, we see improvements in clustering quality over existing baselines.

Michaela Hardt, Alvin Rajkomar, Gerardo Flores et al.

Much work aims to explain a model's prediction on a static input. We consider explanations in a temporal setting where a stateful dynamical model produces a sequence of risk estimates given an input at each time step. When the estimated risk increases, the goal of the explanation is to attribute the increase to a few relevant inputs from the past. While our formal setup and techniques are general, we carry out an in-depth case study in a clinical setting. The goal here is to alert a clinician when a patient's risk of deterioration rises. The clinician then has to decide whether to intervene and adjust the treatment. Given a potentially long sequence of new events since she last saw the patient, a concise explanation helps her to quickly triage the alert. We develop methods to lift static attribution techniques to the dynamical setting, where we identify and address challenges specific to dynamics. We then experimentally assess the utility of different explanations of clinical alerts through expert evaluation.

Horia Mania, John Miller, Ludwig Schmidt et al.

Excessive reuse of test data has become commonplace in today's machine learning workflows. Popular benchmarks, competitions, industrial scale tuning, among other applications, all involve test data reuse beyond guidance by statistical confidence bounds. Nonetheless, recent replication studies give evidence that popular benchmarks continue to support progress despite years of extensive reuse. We proffer a new explanation for the apparent longevity of test data: Many proposed models are similar in their predictions and we prove that this similarity mitigates overfitting. Specifically, we show empirically that models proposed for the ImageNet ILSVRC benchmark agree in their predictions well beyond what we can conclude from their accuracy levels alone. Likewise, models created by large scale hyperparameter search enjoy high levels of similarity. Motivated by these empirical observations, we give a non-asymptotic generalization bound that takes similarity into account, leading to meaningful confidence bounds in practical settings.

Vitaly Feldman, Roy Frostig, Moritz Hardt

Excessive reuse of holdout data can lead to overfitting. However, there is little concrete evidence of significant overfitting due to holdout reuse in popular multiclass benchmarks today. Known results show that, in the worst-case, revealing the accuracy of $k$ adaptively chosen classifiers on a data set of size $n$ allows to create a classifier with bias of $Θ(\sqrt{k/n})$ for any binary prediction problem. We show a new upper bound of $\tilde O(\max\{\sqrt{k\log(n)/(mn)},k/n\})$ on the worst-case bias that any attack can achieve in a prediction problem with $m$ classes. Moreover, we present an efficient attack that achieve a bias of $Ω(\sqrt{k/(m^2 n)})$ and improves on previous work for the binary setting ($m=2$). We also present an inefficient attack that achieves a bias of $\tildeΩ(k/n)$. Complementing our theoretical work, we give new practical attacks to stress-test multiclass benchmarks by aiming to create as large a bias as possible with a given number of queries. Our experiments show that the additional uncertainty of prediction with a large number of classes indeed mitigates the effect of our best attacks. Our work extends developments in understanding overfitting due to adaptive data analysis to multiclass prediction problems. It also bears out the surprising fact that multiclass prediction problems are significantly more robust to overfitting when reusing a test (or holdout) dataset. This offers an explanation as to why popular multiclass prediction benchmarks, such as ImageNet, may enjoy a longer lifespan than what intuition from literature on binary classification suggests.

Chiyuan Zhang, Samy Bengio, Moritz Hardt et al.

We study the interplay between memorization and generalization of overparameterized networks in the extreme case of a single training example and an identity-mapping task. We examine fully-connected and convolutional networks (FCN and CNN), both linear and nonlinear, initialized randomly and then trained to minimize the reconstruction error. The trained networks stereotypically take one of two forms: the constant function (memorization) and the identity function (generalization). We formally characterize generalization in single-layer FCNs and CNNs. We show empirically that different architectures exhibit strikingly different inductive biases. For example, CNNs of up to 10 layers are able to generalize from a single example, whereas FCNs cannot learn the identity function reliably from 60k examples. Deeper CNNs often fail, but nonetheless do astonishing work to memorize the training output: because CNN biases are location invariant, the model must progressively grow an output pattern from the image boundaries via the coordination of many layers. Our work helps to quantify and visualize the sensitivity of inductive biases to architectural choices such as depth, kernel width, and number of channels.

Tijana Zrnic, Moritz Hardt

Adaptive data analysis is frequently criticized for its pessimistic generalization guarantees. The source of these pessimistic bounds is a model that permits arbitrary, possibly adversarial analysts that optimally use information to bias results. While being a central issue in the field, still lacking are notions of natural analysts that allow for more optimistic bounds faithful to the reality that typical analysts aren't adversarial. In this work, we propose notions of natural analysts that smoothly interpolate between the optimal non-adaptive bounds and the best-known adaptive generalization bounds. To accomplish this, we model the analyst's knowledge as evolving according to the rules of an unknown dynamical system that takes in revealed information and outputs new statistical queries to the data. This allows us to restrict the analyst through different natural control-theoretic notions. One such notion corresponds to a recency bias, formalizing an inability to arbitrarily use distant information. Another complementary notion formalizes an anchoring bias, a tendency to weight initial information more strongly. Both notions come with quantitative parameters that smoothly interpolate between the non-adaptive case and the fully adaptive case, allowing for a rich spectrum of intermediate analysts that are neither non-adaptive nor adversarial. Natural not only from a cognitive perspective, we show that our notions also capture standard optimization methods, like gradient descent in various settings. This gives a new interpretation to the fact that gradient descent tends to overfit much less than its adaptive nature might suggest.

Liam Li, Kevin Jamieson, Afshin Rostamizadeh et al.

Modern learning models are characterized by large hyperparameter spaces and long training times. These properties, coupled with the rise of parallel computing and the growing demand to productionize machine learning workloads, motivate the need to develop mature hyperparameter optimization functionality in distributed computing settings. We address this challenge by first introducing a simple and robust hyperparameter optimization algorithm called ASHA, which exploits parallelism and aggressive early-stopping to tackle large-scale hyperparameter optimization problems. Our extensive empirical results show that ASHA outperforms existing state-of-the-art hyperparameter optimization methods; scales linearly with the number of workers in distributed settings; and is suitable for massive parallelism, as demonstrated on a task with 500 workers. We then describe several design decisions we encountered, along with our associated solutions, when integrating ASHA in Determined AI's end-to-end production-quality machine learning system that offers hyperparameter tuning as a service.

Julius Adebayo, Justin Gilmer, Michael Muelly et al.

Saliency methods have emerged as a popular tool to highlight features in an input deemed relevant for the prediction of a learned model. Several saliency methods have been proposed, often guided by visual appeal on image data. In this work, we propose an actionable methodology to evaluate what kinds of explanations a given method can and cannot provide. We find that reliance, solely, on visual assessment can be misleading. Through extensive experiments we show that some existing saliency methods are independent both of the model and of the data generating process. Consequently, methods that fail the proposed tests are inadequate for tasks that are sensitive to either data or model, such as, finding outliers in the data, explaining the relationship between inputs and outputs that the model learned, and debugging the model. We interpret our findings through an analogy with edge detection in images, a technique that requires neither training data nor model. Theory in the case of a linear model and a single-layer convolutional neural network supports our experimental findings.