Self-correction is a highly desirable capability of large language models (LLMs), yet it has consistently been found to be largely ineffective in modern LLMs. Current methods for training self-correction typically depend on either multiple models, a more advanced model, or additional forms of supervision. To address these shortcomings, we develop a multi-turn online reinforcement learning (RL) approach, SCoRe, that significantly improves an LLM's self-correction ability using entirely self-generated data. To build SCoRe, we first show that variants of supervised fine-tuning (SFT) on offline model-generated correction traces are insufficient for instilling self-correction behavior. In particular, we observe that training via SFT either suffers from a distribution mismatch between the training data and the model's own responses or implicitly prefers only a certain mode of correction behavior that is often not effective at test time. SCoRe addresses these challenges by training under the model's own distribution of self-generated correction traces and using appropriate regularization to steer the learning process into learning a self-correction strategy that is effective at test time as opposed to simply fitting high-reward responses for a given prompt. This regularization prescribes running a first phase of RL on a base model to generate a policy initialization that is less susceptible to collapse and then using a reward bonus to amplify self-correction during training. When applied to Gemini 1.0 Pro and 1.5 Flash models, we find that SCoRe achieves state-of-the-art self-correction performance, improving the base models' self-correction by 15.6% and 9.1% respectively on the MATH and HumanEval benchmarks.

As machine learning models become increasingly prevalent in motion forecasting for autonomous vehicles (AVs), it is critical to ensure that model predictions are safe and reliable. In this paper, we examine the robustness of motion forecasting to non-causal perturbations. We construct a new benchmark for evaluating and improving model robustness by applying perturbations to existing data. Specifically, we conduct an extensive labeling effort to identify causal agents, or agents whose presence influences human drivers’ behavior, in the Waymo Open Motion Dataset (WOMD), and we use these labels to perturb the data by deleting non-causal agents from the scene. We evaluate a diverse set of state-of-the-art deep-learning models on our proposed benchmark and find that all evaluated models exhibit large shifts under non-causal perturbation: we observe a surprising 25-38% relative change in minADE as compared to the original. In addition, we investigate techniques to improve model robustness, including increasing the training dataset size and using targeted data augmentations that randomly drop non-causal agents throughout training. Finally, we release the causal agent labels as an extension to WOMD and the robustness benchmarks to aid the community in building more reliable and safe deep-learning models for motion forecasting1.

Imitation learning (IL) is a simple and powerful way to use high-quality human driving data, which can be collected at scale, to produce human-like behavior. However, policies based on imitation learning alone often fail to sufficiently account for safety and reliability concerns. In this paper, we show how imitation learning combined with reinforcement learning using simple rewards can substan-tially improve the safety and reliability of driving policies over those learned from imitation alone. In particular, we train a policy on over lOOk miles of urban driving data, and measure its effectiveness in test scenarios grouped by different levels of collision likelihood. Our analysis shows that while imitation can perform well in low-difficulty scenarios that are well-covered by the demonstration data, our proposed approach significantly improves robustness on the most challenging scenarios (over 38 % reduction in failures). To our knowledge, this is the first application of a combined imitation and reinforcement learning approach in autonomous driving that utilizes large amounts of real- world human driving data.

Simulation is an essential tool to develop and benchmark autonomous vehicle planning software in a safe and cost-effective manner. However, realistic simulation requires accurate modeling of multi-agent interactive behaviors to be trustworthy, behaviors which can be highly nuanced and complex. To address these challenges, we introduce Waymax, a new data-driven simulator for autonomous driving in multi-agent scenes, designed for large-scale simulation and testing. Waymax uses publicly-released, real-world driving data (e.g., the Waymo Open Motion Dataset) to initialize or play back a diverse set of multi-agent simulated scenarios. It runs entirely on hardware accelerators such as TPUs/GPUs and supports in-graph simulation for training, making it suitable for modern large-scale, distributed machine learning workflows. To support online training and evaluation, Waymax includes several learned and hard-coded behavior models that allow for realistic interaction within simulation. To supplement Waymax, we benchmark a suite of popular imitation and reinforcement learning algorithms with ablation studies on different design decisions, where we highlight the effectiveness of routes as guidance for planning agents and the ability of RL to overfit against simulated agents.

For an AI system to be reliable, the confidence it expresses in its decisions must match its accuracy. To assess the degree of match, examples are typically binned by confidence and the per-bin mean confidence and accuracy are compared. Most research in calibration focuses on techniques to reduce this empirical measure of calibration error, ECE_bin. We instead focus on assessing statistical bias in this empirical measure, and we identify better estimators. We propose a framework through which we can compute the bias of a particular estimator for an evaluation data set of a given size. The framework involves synthesizing model outputs that have the same statistics as common neural architectures on popular data sets. We find that binning-based estimators with bins of equal mass (number of instances) have lower bias than estimators with bins of equal width. Our results indicate two reliable calibration-error estimators: the debiased estimator (Brocker, 2012; Ferro and Fricker, 2012) and a method we propose, ECE_sweep, which uses equal-mass bins and chooses the number of bins to be as large as possible while preserving monotonicity in the calibration function. With these estimators, we observe improvements in the effectiveness of recalibration methods and in the detection of model miscalibration.

Large pre-trained models such as CLIP or ALIGN offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning methods substantially improve accuracy on a given target distribution, they often reduce robustness to distribution shifts. We address this tension by introducing a simple and effective method for improving robustness while fine-tuning: ensembling the weights of the zero-shot and fine-tuned models (WiSE-FT). Compared to standard fine-tuning, WiSE-FT provides large accuracy improvements under distribution shift, while preserving high accuracy on the target distribution. On ImageNet and five derived distribution shifts, WiSE-FT improves accuracy under distribution shift by 4 to 6 percentage points (pp) over prior work while increasing ImageNet accuracy by 1.6 pp. WiSE-FT achieves similarly large robustness gains (2 to 23 pp) on a diverse set of six further distribution shifts, and accuracy gains of 0.8 to 3.3 pp compared to standard fine-tuning on commonly used transfer learning datasets. These improvements come at no additional computational cost during fine-tuning or inference.

The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we may average many models without incurring any additional inference or memory costs—we call the results “model soups.” When fine-tuning large pre-trained models such as CLIP, ALIGN, and a ViT-G pre-trained on JFT, our soup recipe provides significant improvements over the best model in a hyperparameter sweep on ImageNet. The resulting ViT-G model, which attains 90.94% top-1 accuracy on ImageNet, achieved a new state of the art. Furthermore, we show that the model soup approach extends to multiple image classification and natural language processing tasks, improves out-of-distribution performance, and improves zero-shot performance on new downstream tasks. Finally, we analytically relate the performance similarity of weight-averaging and logit-ensembling to flatness of the loss and confidence of the predictions, and validate this relation empirically. Code is available at https://github.com/mlfoundations/model-soups.

Despite their ability to represent highly expressive functions, deep learning models seem to find simple solutions that generalize surprisingly well. Spectral bias -- the tendency of neural networks to prioritize learning low frequency functions -- is one possible explanation for this phenomenon, but so far spectral bias has primarily been observed in theoretical models and simplified experiments. In this work, we propose methodologies for measuring spectral bias in modern image classification networks on CIFAR-10 and ImageNet. We find that these networks indeed exhibit spectral bias, and that interventions that improve test accuracy on CIFAR-10 tend to produce learned functions that have higher frequencies overall but lower frequencies in the vicinity of examples from each class. This trend holds across variation in training time, model architecture, number of training examples, data augmentation, and self-distillation. We also explore the connections between function frequency and image frequency and find that spectral bias is sensitive to the low frequencies prevalent in natural images. On ImageNet, we find that learned function frequency also varies with internal class diversity, with higher frequencies on more diverse classes. Our work enables measuring and ultimately influencing the spectral behavior of neural networks used for image classification, and is a step towards understanding why deep models generalize well.

Image classification accuracy on the ImageNet dataset has been a barometer for progress in computer vision over the last decade. Several recent papers have questioned the degree to which the benchmark remains useful to the community, yet innovations continue to contribute gains to performance, with today's largest models achieving 90%+ top-1 accuracy. To help contextualize progress on ImageNet and provide a more meaningful evaluation for today's state-of-the-art models, we manually review and categorize every remaining mistake that a few top models make in order to provide insight into the long-tail of errors on one of the most benchmarked datasets in computer vision. We focus on the multi-label subset evaluation of ImageNet, where today's best models achieve upwards of 97% top-1 accuracy. Our analysis reveals that nearly half of the supposed mistakes are not mistakes at all, and we uncover new valid multi-labels, demonstrating that, without careful review, we are significantly underestimating the performance of these models. On the other hand, we also find that today's best models still make a significant number of mistakes (40%) that are obviously wrong to human reviewers. To calibrate future progress on ImageNet, we provide an updated multi-label evaluation set, and we curate ImageNet-Major: a 68-example "major error" slice of the obvious mistakes made by today's top models -- a slice where models should achieve near perfection, but today are far from doing so.

Vision models notoriously flicker when applied to videos: they correctly recognize objects in some frames, but fail on perceptually similar, nearby frames. In this work, we systematically analyze the robustness of image classifiers to such temporal perturbations in videos. To do so, we construct two new datasets, ImageNet-Vid-Robust and YTBB-Robust, containing a total of 57,897 images grouped into 3,139 sets of perceptually similar images. Our datasets were derived from ImageNet-Vid and Youtube-BB, respectively, and thoroughly re-annotated by human experts for image similarity. We evaluate a diverse array of classifiers pre-trained on ImageNet and show a median classification accuracy drop of 16 and 10 points, respectively, on our two datasets. Additionally, we evaluate three detection models and show that natural perturbations induce both classification as well as localization errors, leading to a median drop in detection mAP of 14 points. Our analysis demonstrates that perturbations occurring naturally in videos pose a substantial and realistic challenge to deploying convolutional neural networks in environments that require both reliable and low-latency predictions.

Optimal decision making requires that classifiers produce uncertainty estimates consistent with their empirical accuracy. However, deep neural networks are often under- or over-confident in their predictions. Consequently, methods have been developed to improve the calibration of their predictive uncertainty both during training and post-hoc. In this work, we propose differentiable losses to improve calibration based on a soft (continuous) version of the binning operation underlying popular calibration-error estimators. When incorporated into training, these soft calibration losses achieve state-of-the-art single-model ECE across multiple datasets with less than 1% decrease in accuracy. For instance, we observe an 82% reduction in ECE (70% relative to the post-hoc rescaled ECE) in exchange for a 0.7% relative decrease in accuracy relative to the cross entropy baseline on CIFAR-100.When incorporated post-training, the soft-binning-based calibration error objective improves upon temperature scaling, a popular recalibration method. Overall, experiments across losses and datasets demonstrate that using calibration-sensitive procedures yield better uncertainty estimates under dataset shift than the standard practice of using a cross entropy loss and post-hoc recalibration methods.

We evaluate a wide range of ImageNet models with five trained human labelers. In our year-long experiment, trained humans first annotated 40,000 images from the ImageNet and ImageNetV2 test sets with multi-class labels to enable a semantically coherent evaluation. Then we measured the classification accuracy of the five trained humans on the full task with 1,000 classes. Only the latest models from 2020 are on par with our best human labeler, and human accuracy on the 590 object classes is still 4% and 10% higher than the best model on ImageNet and ImageNetV2, respectively. Moreover, humans achieve the same accuracy on ImageNet and ImageNetV2, while all models see a consistent accuracy drop. Overall, our results show that there is still substantial room for improvement on ImageNet and direct accuracy comparisons between humans and machines may overstate machine performance.

We build new test sets for the CIFAR-10 and ImageNet datasets. Both benchmarks have been the focus of intense research for almost a decade, raising the danger of overfitting to excessively re-used test sets. By closely following the original dataset creation processes, we test to what extent current classification models generalize to new data. We evaluate a broad range of models and find accuracy drops of 3% - 15% on CIFAR-10 and 11% - 14% on ImageNet. However, accuracy gains on the original test sets translate to larger gains on the new test sets. Our results suggest that the accuracy drops are not caused by adaptivity, but by the models’ inability to generalize to slightly "harder" images than those found in the original test sets.

We conduct the first large meta-analysis of overfitting due to test set reuse in the machine learning community. Our analysis is based on over one hundred machine learning competitions hosted on the Kaggle platform over the course of several years. In each competition, numerous practitioners repeatedly evaluated their progress against a holdout set that forms the basis of a public ranking available throughout the competition. Performance on a separate test set used only once determined the final ranking. By systematically comparing the public ranking with the final ranking, we assess how much participants adapted to the holdout set over the course of a competition. Our study shows, somewhat surprisingly, little evidence of substantial overfitting. These findings speak to the robustness of the holdout method across different data domains, loss functions, model classes, and human analysts.

Adaptive optimization methods, which perform local optimization with a metric constructed from the history of iterates, are becoming increasingly popular for training deep neural networks. Examples include AdaGrad, RMSProp, and Adam. We show that for simple overparameterized problems, adaptive methods often find drastically different solutions than gradient descent (GD) or stochastic gradient descent (SGD). We construct an illustrative binary classification problem where the data is linearly separable, GD and SGD achieve zero test error, and AdaGrad, Adam, and RMSProp attain test errors arbitrarily close to half. We additionally study the empirical generalization capability of adaptive methods on several state-of-the-art deep learning models. We observe that the solutions found by adaptive methods generalize worse (often significantly worse) than SGD, even when these solutions have better training performance. These results suggest that practitioners should reconsider the use of adaptive methods to train neural networks.