Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be resistant to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-limited versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 405B in size. LiveBench is difficult, with top models achieving below 70% accuracy. We release all questions, code, and model answers. Questions are added and updated on a monthly basis, and we release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

Recent claims about the impressive abilities of large language models (LLMs) are often supported by evaluating publicly available benchmarks. Since LLMs train on wide swaths of the internet, this practice raises concerns of data contamination, i.e., evaluating on examples that are explicitly or implicitly included in the training data. Data contamination remains notoriously challenging to measure and mitigate, even with partial attempts like controlled experimentation of training data, canary strings, or embedding similarities. In this work, we conduct the first thorough longitudinal analysis of data contamination in LLMs by using the natural experiment of training cutoffs in GPT models to look at benchmarks released over time. Specifically, we consider two code/mathematical problem-solving datasets, Codeforces and Project Euler, and find statistically significant trends among LLM pass rate vs. GitHub popularity and release date that provide strong evidence of contamination. By open-sourcing our dataset, raw results, and evaluation framework, our work paves the way for rigorous analyses of data contamination in modern models. We conclude with a discussion of best practices and future steps for publicly releasing benchmark in the age of LLMs which train on webscale data.

Neural operators, such as Fourier Neural Operators (FNO), form a principled approach for learning solution operators for partial differential equations (PDE) and other mappings between function spaces. However, many real-world problems require high-resolution training data, and the training time and limited GPU memory pose big barriers. One solution is to train neural operators in mixed precision to reduce the memory requirement and increase training speed. However, existing mixed-precision training techniques are designed for standard neural networks, and we find that their direct application to FNO leads to numerical overflow and poor memory efficiency. Further, at first glance, it may appear that mixed precision in FNO will lead to drastic accuracy degradation since reducing the precision of the Fourier transform yields poor results in classical numerical solvers. We show that this is not the case; in fact, we prove that reducing the precision in FNO still guarantees a good approximation bound, when done in a targeted manner. Specifically, we build on the intuition that neural operator learning inherently induces an approximation error, arising from discretizing the infinite-dimensional ground-truth input function, implying that training in full precision is not needed. We formalize this intuition by rigorously characterizing the approximation and precision errors of FNO and bounding these errors for general input functions. We prove that the precision error is asymptotically comparable to the approximation error. Based on this, we design a simple method to optimize the memory-intensive half-precision tensor contractions by greedily finding the optimal contraction order. Through extensive experiments on different state-of-the-art neural operators, datasets, and GPUs, we demonstrate that our approach reduces GPU memory usage by up to 50% and improves throughput by 58% with little or no reduction in accuracy.

While tabular classification has traditionally relied on from-scratch training, a recent breakthrough called prior-data fitted networks (PFNs) challenges this approach. Similar to large language models, PFNs make use of pretraining and in-context learning to achieve strong performance on new tasks in a single forward pass. However, current PFNs have limitations that prohibit their widespread adoption. Notably, TabPFN achieves very strong performance on small tabular datasets but is not designed to make predictions for datasets of size larger than 1000. In this work, we overcome these limitations and substantially improve the performance of PFNs via context optimization. We introduce TuneTables, a parameter-efficient fine-tuning strategy for PFNs that compresses large datasets into a smaller learned context. We conduct extensive experiments on nineteen algorithms over 98 datasets and find that TuneTables achieves the best performance on average, outperforming boosted trees such as CatBoost, while optimizing fewer than 5\% of TabPFN's parameters. Furthermore, we show that TuneTables can be used as an interpretability tool and can even be used to mitigate biases by optimizing a fairness objective.

Existing neural operator architectures face challenges when solving multiphysics problems with coupled partial differential equations (PDEs) due to complex geometries, interactions between physical variables, and the limited amounts of high-resolution training data. To address these issues, we propose *Codomain Attention Neural Operator* (CoDA-NO), which tokenizes functions along the codomain or channel space, enabling self-supervised learning or pretraining of multiple PDE systems. Specifically, we extend positional encoding, self-attention, and normalization layers to function spaces. CoDA-NO can learn representations of different PDE systems with a single model. We evaluate CoDA-NO's potential as a backbone for learning multiphysics PDEs over multiple systems by considering few-shot learning settings. On complex downstream tasks with limited data, such as fluid flow simulations, fluid-structure interactions, and Rayleigh-Bénard convection, we found CoDA-NO to outperform existing methods by over 36%.

The vast majority of time-series forecasting approaches require a substantial training dataset. However, many real-life forecasting applications have very little initial observations, sometimes just 40 or fewer. Thus, the applicability of most forecasting methods is restricted in data-sparse commercial applications. While there is recent work in the setting of very limited initial data (so-called `zero-shot' forecasting), its performance is inconsistent depending on the data used for pretraining. In this work, we take a different approach and devise ForecastPFN, the first zero-shot forecasting model trained purely on a novel synthetic data distribution. ForecastPFN is a prior-data fitted network, trained to approximate Bayesian inference, which can make predictions on a new time series dataset in a single forward pass. Through extensive experiments, we show that zero-shot predictions made by ForecastPFN are more accurate and faster compared to state-of-the-art forecasting methods, even when the other methods are allowed to train on hundreds of additional in-distribution data points.

Face recognition systems are widely deployed in safety-critical applications, including law enforcement, yet they exhibit bias across a range of socio-demographic dimensions, such as gender and race. Conventional wisdom dictates that model biases arise from biased training data. As a consequence, previous works on bias mitigation largely focused on pre-processing the training data, adding penalties to prevent bias from effecting the model during training, or post-processing predictions to debias them, yet these approaches have shown limited success on hard problems such as face recognition. In our work, we discover that biases are actually inherent to neural network architectures themselves. Following this reframing, we conduct the first neural architecture search for fairness, jointly with a search for hyperparameters. Our search outputs a suite of models which Pareto-dominate all other high-performance architectures and existing bias mitigation methods in terms of accuracy and fairness, often by large margins, on the two most widely used datasets for face identification, CelebA and VGGFace2. Furthermore, these models generalize to other datasets and sensitive attributes. We release our code, models and raw data files at https://github.com/dooleys/FR-NAS.

Tabular data is one of the most commonly used types of data in machine learning. Despite recent advances in neural nets (NNs) for tabular data, there is still an active discussion on whether or not NNs generally outperform gradient-boosted decision trees (GBDTs) on tabular data, with several recent works arguing either that GBDTs consistently outperform NNs on tabular data, or vice versa. In this work, we take a step back and question the importance of this debate. To this end, we conduct the largest tabular data analysis to date, comparing 19 algorithms across 176 datasets, and we find that the 'NN vs. GBDT' debate is overemphasized: for a surprisingly high number of datasets, either the performance difference between GBDTs and NNs is negligible, or light hyperparameter tuning on a GBDT is more important than choosing between NNs and GBDTs. Next, we analyze dozens of metafeatures to determine what \emph{properties} of a dataset make NNs or GBDTs better-suited to perform well. For example, we find that GBDTs are much better than NNs at handling skewed or heavy-tailed feature distributions and other forms of dataset irregularities. Our insights act as a guide for practitioners to determine which techniques may work best on their dataset. Finally, with the goal of accelerating tabular data research, we release the TabZilla Benchmark Suite: a collection of the 36 'hardest' of the datasets we study. Our benchmark suite, codebase, and all raw results are available at https://github.com/naszilla/tabzilla.

Zero-cost proxies (ZC proxies) are a recent architecture performance prediction technique aiming to significantly speed up algorithms for neural architecture search (NAS). Recent work has shown that these techniques show great promise, but certain aspects, such as evaluating and exploiting their complementary strengths, are under-studied. In this work, we create NAS-Bench-Suite: we evaluate 13 ZC proxies across 28 tasks, creating by far the largest dataset (and unified codebase) for ZC proxies, enabling orders-of-magnitude faster experiments on ZC proxies, while avoiding confounding factors stemming from different implementations. To demonstrate the usefulness of NAS-Bench-Suite, we run a large-scale analysis of ZC proxies, including a bias analysis, and the first information-theoretic analysis which concludes that ZC proxies capture substantial complementary information. Motivated by these findings, we present a procedure to improve the performance of ZC proxies by reducing biases such as cell size, and we also show that incorporating all 13 ZC proxies into the surrogate models used by NAS algorithms can improve their predictive performance by up to 42%. Our code and datasets are available at https://github.com/automl/naslib/tree/zerocost.

While other areas of machine learning have seen more and more automation, designing a high-performing recommender system still requires a high level of human effort. Furthermore, recent work has shown that modern recommender system algorithms do not always improve over well-tuned baselines. A natural follow-up question is, "how do we choose the right algorithm for a new dataset and performance metric?" In this work, we start by giving the first large-scale study of recommender system approaches by comparing 24 algorithms and 100 sets of hyperparameters across 85 datasets and 315 metrics. We find that the best algorithms and hyperparameters are highly dependent on the dataset and performance metric. However, there is also a strong correlation between the performance of each algorithm and various meta-features of the datasets. Motivated by these findings, we create RecZilla, a meta-learning approach to recommender systems that uses a model to predict the best algorithm and hyperparameters for new, unseen datasets. By using far more meta-training data than prior work, RecZilla is able to substantially reduce the level of human involvement when faced with a new recommender system application. We not only release our code and pretrained RecZilla models, but also all of our raw experimental results, so that practitioners can train a RecZilla model for their desired performance metric: https://github.com/naszilla/reczilla.

Over the past half-decade, many methods have been considered for neural architecture search (NAS). Bayesian optimization (BO), which has long had success in hyperparameter optimization, has recently emerged as a very promising strategy for NAS when it is coupled with a neural predictor. Recent work has proposed different instantiations of this framework, for example, using Bayesian neural networks or graph convolutional networks as the predictive model within BO. However, the analyses in these papers often focus on the full-fledged NAS algorithm, so it is difficult to tell which individual components of the framework lead to the best performance. In this work, we give a thorough analysis of the "BO + neural predictor framework" by identifying five main components: the architecture encoding, neural predictor, uncertainty calibration method, acquisition function, and acquisition function optimization. We test several different methods for each component and also develop a novel path-based encoding scheme for neural architectures, which we show theoretically and empirically scales better than other encodings. Using all of our analyses, we develop a final algorithm called BANANAS, which achieves state-of-the-art performance on NAS search spaces. We adhere to the NAS research checklist (Lindauer and Hutter 2019) to facilitate best practices, and our code is available at https://github.com/naszilla/naszilla.

Early methods in the rapidly developing field of neural architecture search (NAS) required fully training thousands of neural networks. To reduce this extreme computational cost, dozens of techniques have since been proposed to predict the final performance of neural architectures. Despite the success of such performance prediction methods, it is not well-understood how different families of techniques compare to one another, due to the lack of an agreed-upon evaluation metric and optimization for different constraints on the initialization time and query time. In this work, we give the first large-scale study of performance predictors by analyzing 31 techniques ranging from learning curve extrapolation, to weight-sharing, to supervised learning, to zero-cost proxies. We test a number of correlation- and rank-based performance measures in a variety of settings, as well as the ability of each technique to speed up predictor-based NAS frameworks. Our results act as recommendations for the best predictors to use in different settings, and we show that certain families of predictors can be combined to achieve even better predictive power, opening up promising research directions. We release our code, featuring a library of 31 performance predictors.

While early research in neural architecture search (NAS) required extreme computational resources, the recent releases of tabular and surrogate benchmarks have greatly increased the speed and reproducibility of NAS research. However, two of the most popular benchmarks do not provide the full training information for each architecture. As a result, on these benchmarks it is not possible to evaluate many types of multi-fidelity algorithms, such as learning curve extrapolation, that require evaluating architectures at arbitrary epochs. In this work, we present a method using singular value decomposition and noise modeling to create surrogate benchmarks, NAS-Bench-111, NAS-Bench-311, and NAS-Bench-NLP11, that output the full training information for each architecture, rather than just the final validation accuracy. We demonstrate the power of using the full training information by introducing a learning curve extrapolation framework to modify single-fidelity algorithms, showing that it leads to improvements over popular single-fidelity algorithms which claimed to be state-of-the-art upon release.

Neural architecture search (NAS) has seen a steep rise in interest over the last few years. Many algorithms for NAS consist of searching through a space of architectures by iteratively choosing an architecture, evaluating its performance by training it, and using all prior evaluations to come up with the next choice. The evaluation step is noisy - the final accuracy varies based on the random initialization of the weights. Prior work has focused on devising new search algorithms to handle this noise, rather than quantifying or understanding the level of noise in architecture evaluations. In this work, we show that (1) the simplest hill-climbing algorithm is a powerful baseline for NAS, and (2), when the noise in popular NAS benchmark datasets is reduced to a minimum, the loss landscape becomes near-convex, causing hill-climbing to outperform many popular state-of-the-art algorithms. We further back up this observation by showing that the number of local minima is substantially reduced as the noise decreases and by giving a theoretical characterization of the performance of local search in NAS. Based on our findings, for NAS research we suggest (1) using local search as a baseline, and (2) denoising the training pipeline when possible.

As deep learning models become tasked with more and more decisions that impact human lives, such as criminal recidivism, loan repayment, and face recognition for law enforcement, bias is becoming a growing concern. Debiasing algorithms are typically split into three paradigms: pre-processing, in-processing, and post-processing. However, in computer vision or natural language applications, it is common to start with a large generic model and then fine-tune to a specific use-case. Pre- or in-processing methods would require retraining the entire model from scratch, while post-processing methods only have black-box access to the model, so they do not leverage the weights of the trained model. Creating debiasing algorithms specifically for this fine-tuning use-case has largely been neglected. In this work, we initiate the study of a new paradigm in debiasing research, intra-processing, which sits between in-processing and post-processing methods. Intra-processing methods are designed specifically to debias large models which have been trained on a generic dataset, and fine-tuned on a more specific task. We show how to repurpose existing in-processing methods for this use-case, and we also propose three baseline algorithms: random perturbation, layerwise optimization, and adversarial debiasing. We evaluate these methods across three popular datasets from the AIF360 toolkit, as well as on the CelebA faces dataset. Our code is available at https://github.com/abacusai/intraprocessing_debiasing.

Neural architecture search (NAS) has been extensively studied in the past few years. A popular approach is to represent each neural architecture in the search space as a directed acyclic graph (DAG), and then search over all DAGs by encoding the adjacency matrix and list of operations as a set of hyperparameters. Recent work has demonstrated that even small changes to the way each architecture is encoded can have a significant effect on the performance of NAS algorithms. In this work, we present the first formal study on the effect of architecture encodings for NAS, including a theoretical grounding and an empirical study. First we formally define architecture encodings and give a theoretical characterization on the scalability of the encodings we study. Then we identify the main encoding-dependent subroutines which NAS algorithms employ, running experiments to show which encodings work best with each subroutine for many popular algorithms. The experiments act as an ablation study for prior work, disentangling the algorithmic and encoding-based contributions, as well as a guideline for future work. Our results demonstrate that NAS encodings are an important design decision which can have a significant impact on overall performance. Our code is available at https://github.com/naszilla/naszilla.

The k -center problem is a canonical and long-studied facility location and clustering problem with many applications in both its symmetric and asymmetric forms. Both versions of the problem have tight approximation factors on worst case instances: a 2-approximation for symmetric k -center and an O (log * ( k ))-approximation for the asymmetric version. Therefore, to improve on these ratios, one must go beyond the worst case. In this work, we take this approach and provide strong positive results both for the asymmetric and symmetric k -center problems under a natural input stability (promise) condition called α-perturbation resilience [15], which states that the optimal solution does not change under any α-factor perturbation to the input distances. We provide algorithms that give strong guarantees simultaneously for stable and non-stable instances: Our algorithms always inherit the worst-case guarantees of clustering approximation algorithms and output the optimal solution if the input is 2-perturbation resilient. In particular, we show that if the input is only perturbation resilient on part of the data, our algorithm will return the optimal clusters from the region of the data that is perturbation resilient while achieving the best worst-case approximation guarantee on the remainder of the data. Furthermore, we prove that our result is tight by showing symmetric k -center under (2 − ϵ)-perturbation resilience is hard unless NP = RP . The impact of our results is multifaceted. First, to our knowledge, asymmetric k -center is the first problem that is hard to approximate to any constant factor in the worst case, yet can be optimally solved in polynomial time under perturbation resilience for a constant value of α. This is also the first tight result for any problem under perturbation resilience, i.e., this is the first time the exact value of α for which the problem switches from being NP-hard to efficiently computable has been found. Furthermore, our results illustrate a surprising relationship between symmetric and asymmetric k -center instances under perturbation resilience. Unlike approximation ratio, for which symmetric k -center is easily solved to a factor of 2 but asymmetric k -center cannot be approximated to any constant factor, both symmetric and asymmetric k -center can be solved optimally under resilience to 2-perturbations. Finally, our guarantees in the setting where only part of the data satisfies perturbation resilience make these algorithms more applicable to real-life instances.

In this work, we study the k-median and k-means clustering problems when the data is distributed across many servers and can contain outliers. While there has been a lot of work on these problems for worst-case instances, we focus on gaining a finer understanding through the lens of beyond worst-case analysis. Our main motivation is the following: for many applications such as clustering proteins by function or clustering communities in a social network, there is some unknown target clustering, and the hope is that running a k-median or k-means algorithm will produce clusterings which are close to matching the target clustering. Worst-case results can guarantee constant factor approximations to the optimal k-median or k-means objective value, but not closeness to the target clustering. Our first result is a distributed algorithm which returns a near-optimal clustering assuming a natural notion of stability, namely, approximation stability [Awasthi and Balcan, 2014], even when a constant fraction of the data are outliers. The communication complexity is O~(sk+z) where s is the number of machines, k is the number of clusters, and z is the number of outliers. Next, we show this amount of communication cannot be improved even in the setting when the input satisfies various non-worst-case assumptions. We give a matching Omega(sk+z) lower bound on the communication required both for approximating the optimal k-means or k-median cost up to any constant, and for returning a clustering that is close to the target clustering in Hamming distance. These lower bounds hold even when the data satisfies approximation stability or other common notions of stability, and the cluster sizes are balanced. Therefore, Omega(sk+z) is a communication bottleneck, even for real-world instances.

Algorithms for clustering points in metric spaces is a long-studied area of research. Clustering has seen a multitude of work both theoretically, in understanding the approximation guarantees possible for many objective functions such as k-median and k-means clustering, and experimentally, in finding the fastest algorithms and seeding procedures for Lloyd's algorithm. The performance of a given clustering algorithm depends on the specific application at hand, and this may not be known up front. For example, a "typical instance" may vary depending on the application, and different clustering heuristics perform differently depending on the instance. In this paper, we define an infinite family of algorithms generalizing Lloyd's algorithm, with one parameter controlling the the initialization procedure, and another parameter controlling the local search procedure. This family of algorithms includes the celebrated k-means++ algorithm, as well as the classic farthest-first traversal algorithm. We design efficient learning algorithms which receive samples from an application-specific distribution over clustering instances and learn a near-optimal clustering algorithm from the class. We show the best parameters vary significantly across datasets such as MNIST, CIFAR, and mixtures of Gaussians. Our learned algorithms never perform worse than k-means++, and on some datasets we see significant improvements.

In distributed machine learning, data is dispatched to multiple machines for processing. Motivated by the fact that similar data points often belong to the same or similar classes, and more generally, classification rules of high accuracy tend to be “locally simple but globally complex” (Vapnik and Bottou, 1993), we propose data dependent dispatching that takes advantage of such structure. We present an in-depth analysis of this model, providing new algorithms with provable worst-case guarantees, analysis proving existing scalable heuristics perform well in natural non worst-case conditions, and techniques for extending a dispatching rule from a small sample to the entire distribution. We overcome novel technical challenges to satisfy important conditions for accurate distributed learning, including fault tolerance and balancedness. We empirically compare our approach with baselines based on random partitioning, balanced partition trees, and locality sensitive hashing, showing that we achieve significantly higher accuracy on both synthetic and real world image and advertising datasets. We also demonstrate that our technique strongly scales with the available computing power.

Max-cut, clustering, and many other partitioning problems that are of significant importance to machine learning and other scientific fields are NP-hard, a reality that has motivated researchers to develop a wealth of approximation algorithms and heuristics. Although the best algorithm to use typically depends on the specific application domain, a worst-case analysis is often used to compare algorithms. This may be misleading if worst-case instances occur infrequently, and thus there is a demand for optimization methods which return the algorithm configuration best suited for the given application’s typical inputs. Recently, Gupta and Roughgarden introduced the first learning-theoretic framework to rigorously study this problem, using it to analyze classes of greedy heuristics, parameter tuning in gradient descent, and other problems. We study this algorithm configuration problem for clustering, max-cut, and other partitioning problems, such as integer quadratic programming, by designing computationally efficient and sample efficient learning algorithms which receive samples from an application-specific distribution over problem instances and learn a partitioning algorithm with high expected performance. Our algorithms learn over common integer quadratic programming and clustering algorithm families: SDP rounding algorithms and agglomerative clustering algorithms with dynamic programming. For our sample complexity analysis, we provide tight bounds on the pseudodimension of these algorithm classes, and show that surprisingly, even for classes of algorithms parameterized by a single parameter, the pseudo-dimension is superconstant. In this way, our work both contributes to the foundations of algorithm configuration and pushes the boundaries of learning theory, since the algorithm classes we analyze consist of multi-stage optimization procedures and are significantly more complex than classes typically studied in learning theory.

A large body of work in machine learning has focused on the problem of learning a close approximation to an underlying combinatorial function, given a small set of labeled examples. However, for real-valued functions, cardinal labels might not be accessible, or it may be difficult for an expert to consistently assign real-valued labels over the entire set of examples. For instance, it is notoriously hard for consumers to reliably assign values to bundles of merchandise. Instead, it might be much easier for a consumer to report which of two bundles she likes better. With this motivation in mind, we consider an alternative learning model, wherein the algorithm must learn the underlying function up to pairwise comparisons, from pairwise comparisons. In this model, we present a series of novel algorithms that learn over a wide variety of combinatorial function classes. These range from graph functions to broad classes of valuation functions that are fundamentally important in microeconomic theory, the analysis of social networks, and machine learning, such as coverage, submodular, XOS, and subadditive functions, as well as functions with sparse Fourier support.

The $k$-center problem is a canonical and long-studied facility location and clustering problem with many applications in both its symmetric and asymmetric forms. Both versions of the problem have tight approximation factors on worst case instances. Therefore to improve on these ratios, one must go beyond the worst case. In this work, we take this approach and provide strong positive results both for the asymmetric and symmetric $k$-center problems under a natural input stability (promise) condition called $α$-perturbation resilience [Bilu and Linia 2012], which states that the optimal solution does not change under any alpha-factor perturbation to the input distances. We provide algorithms that give strong guarantees simultaneously for stable and non-stable instances: our algorithms always inherit the worst-case guarantees of clustering approximation algorithms, and output the optimal solution if the input is $2$-perturbation resilient. Furthermore, we prove our result is tight by showing symmetric $k$-center under $(2-ε)$-perturbation resilience is hard unless $NP=RP$. The impact of our results are multifaceted. This is the first tight result for any problem under perturbation resilience. Furthermore, our results illustrate a surprising relationship between symmetric and asymmetric $k$-center instances under perturbation resilience. Unlike approximation ratio, for which symmetric $k$-center is easily solved to a factor of 2 but asymmetric $k$-center cannot be approximated to any constant factor, both symmetric and asymmetric $k$-center can be solved optimally under resilience to 2-perturbations. Finally, our guarantees in the setting where only part of the data satisfies perturbation resilience makes these algorithms more applicable to real-life instances.