Marvin Zhang

Aaron Hurst, Adam Lerer, Adam P. Goucher et al. · openai

GPT-4o is an autoregressive omni model that accepts as input any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs. It's trained end-to-end across text, vision, and audio, meaning all inputs and outputs are processed by the same neural network. GPT-4o can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds, which is similar to human response time in conversation. It matches GPT-4 Turbo performance on text in English and code, with significant improvement on text in non-English languages, while also being much faster and 50\% cheaper in the API. GPT-4o is especially better at vision and audio understanding compared to existing models. In line with our commitment to building AI safely and consistent with our voluntary commitments to the White House, we are sharing the GPT-4o System Card, which includes our Preparedness Framework evaluations. In this System Card, we provide a detailed look at GPT-4o's capabilities, limitations, and safety evaluations across multiple categories, focusing on speech-to-speech while also evaluating text and image capabilities, and measures we've implemented to ensure the model is safe and aligned. We also include third-party assessments on dangerous capabilities, as well as discussion of potential societal impacts of GPT-4o's text and vision capabilities.

Josh Achiam, Steven Adler, Sandhini Agarwal et al. · berkeley, deepmind

We report the development of GPT-4, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers. GPT-4 is a Transformer-based model pre-trained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior. A core component of this project was developing infrastructure and optimization methods that behave predictably across a wide range of scales. This allowed us to accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4.

Aaditya Singh, Adam Fry, Adam Perelman et al. · berkeley, mila

This is the system card published alongside the OpenAI GPT-5 launch, August 2025. GPT-5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say 'think hard about this' in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time. Once usage limits are reached, a mini version of each model handles remaining queries. This system card focuses primarily on gpt-5-thinking and gpt-5-main, while evaluations for other models are available in the appendix. The GPT-5 system not only outperforms previous models on benchmarks and answers questions more quickly, but -- more importantly -- is more useful for real-world queries. We've made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy, and have leveled up GPT-5's performance in three of ChatGPT's most common uses: writing, coding, and health. All of the GPT-5 models additionally feature safe-completions, our latest approach to safety training to prevent disallowed content. Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards. While we do not have definitive evidence that this model could meaningfully help a novice to create severe biological harm -- our defined threshold for High capability -- we have chosen to take a precautionary approach.

Pang Wei Koh, Shiori Sagawa, Henrik Marklund et al.

Distribution shifts -- where the training distribution differs from the test distribution -- can substantially degrade the accuracy of machine learning (ML) systems deployed in the wild. Despite their ubiquity in the real-world deployments, these distribution shifts are under-represented in the datasets widely used in the ML community today. To address this gap, we present WILDS, a curated benchmark of 10 datasets reflecting a diverse range of distribution shifts that naturally arise in real-world applications, such as shifts across hospitals for tumor identification; across camera traps for wildlife monitoring; and across time and location in satellite imaging and poverty mapping. On each dataset, we show that standard training yields substantially lower out-of-distribution than in-distribution performance. This gap remains even with models trained by existing methods for tackling distribution shifts, underscoring the need for new methods for training models that are more robust to the types of distribution shifts that arise in practice. To facilitate method development, we provide an open-source package that automates dataset loading, contains default model architectures and hyperparameters, and standardizes evaluations. Code and leaderboards are available at https://wilds.stanford.edu.

Marvin Zhang, Sergey Levine, Chelsea Finn

While deep neural networks can attain good accuracy on in-distribution test points, many applications require robustness even in the face of unexpected perturbations in the input, changes in the domain, or other sources of distribution shift. We study the problem of test time robustification, i.e., using the test input to improve model robustness. Recent prior works have proposed methods for test time adaptation, however, they each introduce additional assumptions, such as access to multiple test points, that prevent widespread adoption. In this work, we aim to study and devise methods that make no assumptions about the model training process and are broadly applicable at test time. We propose a simple approach that can be used in any test setting where the model is probabilistic and adaptable: when presented with a test example, perform different data augmentations on the data point, and then adapt (all of) the model parameters by minimizing the entropy of the model's average, or marginal, output distribution across the augmentations. Intuitively, this objective encourages the model to make the same prediction across different augmentations, thus enforcing the invariances encoded in these augmentations, while also maintaining confidence in its predictions. In our experiments, we evaluate two baseline ResNet models, two robust ResNet-50 models, and a robust vision transformer model, and we demonstrate that this approach achieves accuracy gains of 1-8\% over standard model evaluation and also generally outperforms prior augmentation and adaptation strategies. For the setting in which only one test point is available, we achieve state-of-the-art results on the ImageNet-C, ImageNet-R, and, among ResNet-50 models, ImageNet-A distribution shift benchmarks.

Marvin Zhang, Henrik Marklund, Nikita Dhawan et al.

A fundamental assumption of most machine learning algorithms is that the training and test data are drawn from the same underlying distribution. However, this assumption is violated in almost all practical applications: machine learning systems are regularly tested under distribution shift, due to changing temporal correlations, atypical end users, or other factors. In this work, we consider the problem setting of domain generalization, where the training data are structured into domains and there may be multiple test time shifts, corresponding to new domains or domain distributions. Most prior methods aim to learn a single robust model or invariant feature space that performs well on all domains. In contrast, we aim to learn models that adapt at test time to domain shift using unlabeled test points. Our primary contribution is to introduce the framework of adaptive risk minimization (ARM), in which models are directly optimized for effective adaptation to shift by learning to adapt on the training domains. Compared to prior methods for robustness, invariance, and adaptation, ARM methods provide performance gains of 1-4% test accuracy on a number of image classification problems exhibiting domain shift.

Laura Smith, Nikita Dhawan, Marvin Zhang et al.

Robotic reinforcement learning (RL) holds the promise of enabling robots to learn complex behaviors through experience. However, realizing this promise for long-horizon tasks in the real world requires mechanisms to reduce human burden in terms of defining the task and scaffolding the learning process. In this paper, we study how these challenges can be alleviated with an automated robotic learning framework, in which multi-stage tasks are defined simply by providing videos of a human demonstrator and then learned autonomously by the robot from raw image observations. A central challenge in imitating human videos is the difference in appearance between the human and robot, which typically requires manual correspondence. We instead take an automated approach and perform pixel-level image translation via CycleGAN to convert the human demonstration into a video of a robot, which can then be used to construct a reward function for a model-based RL algorithm. The robot then learns the task one stage at a time, automatically learning how to reset each stage to retry it multiple times without human-provided resets. This makes the learning process largely automatic, from intuitive task specification via a video to automated training with minimal human intervention. We demonstrate that our approach is capable of learning complex tasks, such as operating a coffee machine, directly from raw image observations, requiring only 20 minutes to provide human demonstrations and about 180 minutes of robot interaction.

Michael Janner, Justin Fu, Marvin Zhang et al.

Designing effective model-based reinforcement learning algorithms is difficult because the ease of data generation must be weighed against the bias of model-generated data. In this paper, we study the role of model usage in policy optimization both theoretically and empirically. We first formulate and analyze a model-based reinforcement learning algorithm with a guarantee of monotonic improvement at each step. In practice, this analysis is overly pessimistic and suggests that real off-policy data is always preferable to model-generated on-policy data, but we show that an empirical estimate of model generalization can be incorporated into such analysis to justify model usage. Motivated by this analysis, we then demonstrate that a simple procedure of using short model-generated rollouts branched from real data has the benefits of more complicated model-based algorithms without the usual pitfalls. In particular, this approach surpasses the sample efficiency of prior model-based methods, matches the asymptotic performance of the best model-free algorithms, and scales to horizons that cause other model-based methods to fail entirely.

Marvin Zhang, Sharad Vikram, Laura Smith et al.

Model-based reinforcement learning (RL) has proven to be a data efficient approach for learning control tasks but is difficult to utilize in domains with complex observations such as images. In this paper, we present a method for learning representations that are suitable for iterative model-based policy improvement, even when the underlying dynamical system has complex dynamics and image observations, in that these representations are optimized for inferring simple dynamics and cost models given data from the current policy. This enables a model-based RL method based on the linear-quadratic regulator (LQR) to be used for systems with image observations. We evaluate our approach on a range of robotics tasks, including manipulation with a real-world robotic arm directly from images. We find that our method produces substantially better final performance than other model-based RL methods while being significantly more efficient than model-free RL.

Yevgen Chebotar, Karol Hausman, Marvin Zhang et al.

Reinforcement learning (RL) algorithms for real-world robotic applications need a data-efficient learning process and the ability to handle complex, unknown dynamical systems. These requirements are handled well by model-based and model-free RL approaches, respectively. In this work, we aim to combine the advantages of these two types of methods in a principled manner. By focusing on time-varying linear-Gaussian policies, we enable a model-based algorithm based on the linear quadratic regulator (LQR) that can be integrated into the model-free framework of path integral policy improvement (PI2). We can further combine our method with guided policy search (GPS) to train arbitrary parameterized policies such as deep neural networks. Our simulation and real-world experiments demonstrate that this method can solve challenging manipulation tasks with comparable or better performance than model-free methods while maintaining the sample efficiency of model-based methods. A video presenting our results is available at https://sites.google.com/site/icml17pilqr

Marvin Zhang, Xinyang Geng, Jonathan Bruce et al.

Tensegrity robots, composed of rigid rods connected by elastic cables, have a number of unique properties that make them appealing for use as planetary exploration rovers. However, control of tensegrity robots remains a difficult problem due to their unusual structures and complex dynamics. In this work, we show how locomotion gaits can be learned automatically using a novel extension of mirror descent guided policy search (MDGPS) applied to periodic locomotion movements, and we demonstrate the effectiveness of our approach on tensegrity robot locomotion. We evaluate our method with real-world and simulated experiments on the SUPERball tensegrity robot, showing that the learned policies generalize to changes in system parameters, unreliable sensor measurements, and variation in environmental conditions, including varied terrains and a range of different gravities. Our experiments demonstrate that our method not only learns fast, power-efficient feedback policies for rolling gaits, but that these policies can succeed with only the limited onboard sensing provided by SUPERball's accelerometers. We compare the learned feedback policies to learned open-loop policies and hand-engineered controllers, and demonstrate that the learned policy enables the first continuous, reliable locomotion gait for the real SUPERball robot. Our code and other supplementary materials are available from http://rll.berkeley.edu/drl_tensegrity

Marvin Zhang, Zoe McCarthy, Chelsea Finn et al.

Policy learning for partially observed control tasks requires policies that can remember salient information from past observations. In this paper, we present a method for learning policies with internal memory for high-dimensional, continuous systems, such as robotic manipulators. Our approach consists of augmenting the state and action space of the system with continuous-valued memory states that the policy can read from and write to. Learning general-purpose policies with this type of memory representation directly is difficult, because the policy must automatically figure out the most salient information to memorize at each time step. We show that, by decomposing this policy search problem into a trajectory optimization phase and a supervised learning phase through a method called guided policy search, we can acquire policies with effective memorization and recall strategies. Intuitively, the trajectory optimization phase chooses the values of the memory states that will make it easier for the policy to produce the right action in future states, while the supervised learning phase encourages the policy to use memorization actions to produce those memory states. We evaluate our method on tasks involving continuous control in manipulation and navigation settings, and show that our method can learn complex policies that successfully complete a range of tasks that require memory.