Sebastien Bubeck

Sandhini Agarwal, Lama Ahmad, Jason Ai et al. · openai

We present gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models that push the frontier of accuracy and inference cost. The models use an efficient mixture-of-expert transformer architecture and are trained using large-scale distillation and reinforcement learning. We optimize the models to have strong agentic capabilities (deep research browsing, python tool use, and support for developer-provided functions), all while using a rendered chat format that enables clear instruction following and role delineation. Both models achieve strong results on benchmarks ranging from mathematics, coding, and safety. We release the model weights, inference implementations, tool environments, and tokenizers under an Apache 2.0 license to enable broad use and further research.

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.

Mojan Javaheripi, Gustavo H. de Rosa, Subhabrata Mukherjee et al. · microsoft-research

The Transformer architecture is ubiquitously used as the building block of large-scale autoregressive language models. However, finding architectures with the optimal trade-off between task performance (perplexity) and hardware constraints like peak memory utilization and latency is non-trivial. This is exacerbated by the proliferation of various hardware. We leverage the somewhat surprising empirical observation that the number of decoder parameters in autoregressive Transformers has a high rank correlation with task performance, irrespective of the architecture topology. This observation organically induces a simple Neural Architecture Search (NAS) algorithm that uses decoder parameters as a proxy for perplexity without need for any model training. The search phase of our training-free algorithm, dubbed Lightweight Transformer Search (LTS), can be run directly on target devices since it does not require GPUs. Using on-target-device measurements, LTS extracts the Pareto-frontier of perplexity versus any hardware performance cost. We evaluate LTS on diverse devices from ARM CPUs to NVIDIA GPUs and two popular autoregressive Transformer backbones: GPT-2 and Transformer-XL. Results show that the perplexity of 16-layer GPT-2 and Transformer-XL can be achieved with up to 1.5x, 2.5x faster runtime and 1.2x, 2.0x lower peak memory utilization. When evaluated in zero and one-shot settings, LTS Pareto-frontier models achieve higher average accuracy compared to the 350M parameter OPT across 14 tasks, with up to 1.6x lower latency. LTS extracts the Pareto-frontier in under 3 hours while running on a commodity laptop. We effectively remove the carbon footprint of hundreds of GPU hours of training during search, offering a strong simple baseline for future NAS methods in autoregressive language modeling.

Ganesh Jawahar, Subhabrata Mukherjee, Xiaodong Liu et al. · microsoft-research

Mixture-of-Expert (MoE) models have obtained state-of-the-art performance in Neural Machine Translation (NMT) tasks. Existing works in MoE mostly consider a homogeneous design where the same number of experts of the same size are placed uniformly throughout the network. Furthermore, existing MoE works do not consider computational constraints (e.g., FLOPs, latency) to guide their design. To this end, we develop AutoMoE -- a framework for designing heterogeneous MoE's under computational constraints. AutoMoE leverages Neural Architecture Search (NAS) to obtain efficient sparse MoE sub-transformers with 4x inference speedup (CPU) and FLOPs reduction over manually designed Transformers, with parity in BLEU score over dense Transformer and within 1 BLEU point of MoE SwitchTransformer, on aggregate over benchmark datasets for NMT. Heterogeneous search space with dense and sparsely activated Transformer modules (e.g., how many experts? where to place them? what should be their sizes?) allows for adaptive compute -- where different amounts of computations are used for different tokens in the input. Adaptivity comes naturally from routing decisions which send tokens to experts of different sizes. AutoMoE code, data, and trained models are available at https://aka.ms/AutoMoE.

Ananya Kumar, Ruoqi Shen, Sebastien Bubeck et al. · microsoft-research

SGD and AdamW are the two most used optimizers for fine-tuning large neural networks in computer vision. When the two methods perform the same, SGD is preferable because it uses less memory (12 bytes/parameter with momentum and 8 bytes/parameter without) than AdamW (16 bytes/parameter). However, on a suite of downstream tasks, especially those with distribution shifts, we find that fine-tuning with AdamW performs substantially better than SGD on modern Vision Transformer and ConvNeXt models. We find that large gaps in performance between SGD and AdamW occur when the fine-tuning gradients in the first "embedding" layer are much larger than in the rest of the model. Our analysis suggests an easy fix that works consistently across datasets and models: freezing the embedding layer (less than 1% of the parameters) leads to SGD with or without momentum performing slightly better than AdamW while using less memory (e.g., on ViT-L, SGD uses 33% less GPU memory). Our insights result in state-of-the-art accuracies on five popular distribution shift benchmarks: WILDS-FMoW, WILDS-Camelyon, BREEDS-Living-17, Waterbirds, and DomainNet.

Hadi Salman, Greg Yang, Jerry Li et al.

Recent works have shown the effectiveness of randomized smoothing as a scalable technique for building neural network-based classifiers that are provably robust to $\ell_2$-norm adversarial perturbations. In this paper, we employ adversarial training to improve the performance of randomized smoothing. We design an adapted attack for smoothed classifiers, and we show how this attack can be used in an adversarial training setting to boost the provable robustness of smoothed classifiers. We demonstrate through extensive experimentation that our method consistently outperforms all existing provably $\ell_2$-robust classifiers by a significant margin on ImageNet and CIFAR-10, establishing the state-of-the-art for provable $\ell_2$-defenses. Moreover, we find that pre-training and semi-supervised learning boost adversarially trained smoothed classifiers even further. Our code and trained models are available at http://github.com/Hadisalman/smoothing-adversarial .

Bingbin Liu, Sebastien Bubeck, Ronen Eldan et al.

Small-scale models offer various computational advantages, and yet to which extent size is critical for problem-solving abilities remains an open question. Specifically for solving grade school math, the smallest model size so far required to break the 80\% barrier on the GSM8K benchmark remains to be 34B. Our work studies how high-quality datasets may be the key for small language models to acquire mathematical reasoning. We introduce \texttt{TinyGSM}, a synthetic dataset of 12.3M grade school math problems paired with Python solutions, generated fully by GPT-3.5. After finetuning on \texttt{TinyGSM}, we find that a duo of a 1.3B generation model and a 1.3B verifier model can achieve 81.5\% accuracy, outperforming existing models that are orders of magnitude larger. This also rivals the performance of the GPT-3.5 ``teacher'' model (77.4\%), from which our model's training data is generated. Our approach is simple and has two key components: 1) the high-quality dataset \texttt{TinyGSM}, 2) the use of a verifier, which selects the final outputs from multiple candidate generations.

Debadeepta Dey, Shital Shah, Sebastien Bubeck

The fundamental problem in Neural Architecture Search (NAS) is to efficiently find high-performing architectures from a given search space. We propose a simple but powerful method which we call FEAR, for ranking architectures in any search space. FEAR leverages the viewpoint that neural networks are powerful non-linear feature extractors. First, we train different architectures in the search space to the same training or validation error. Then, we compare the usefulness of the features extracted by each architecture. We do so with a quick training keeping most of the architecture frozen. This gives fast estimates of the relative performance. We validate FEAR on Natsbench topology search space on three different datasets against competing baselines and show strong ranking correlation especially compared to recently proposed zero-cost methods. FEAR particularly excels at ranking high-performance architectures in the search space. When used in the inner loop of discrete search algorithms like random search, FEAR can cut down the search time by approximately 2.4X without losing accuracy. We additionally empirically study very recently proposed zero-cost measures for ranking and find that they breakdown in ranking performance as training proceeds and also that data-agnostic ranking scores which ignore the dataset do not generalize across dissimilar datasets.

Chi Jin, Zeyuan Allen-Zhu, Sebastien Bubeck et al.

Model-free reinforcement learning (RL) algorithms, such as Q-learning, directly parameterize and update value functions or policies without explicitly modeling the environment. They are typically simpler, more flexible to use, and thus more prevalent in modern deep RL than model-based approaches. However, empirical work has suggested that model-free algorithms may require more samples to learn [Deisenroth and Rasmussen 2011, Schulman et al. 2015]. The theoretical question of "whether model-free algorithms can be made sample efficient" is one of the most fundamental questions in RL, and remains unsolved even in the basic scenario with finitely many states and actions. We prove that, in an episodic MDP setting, Q-learning with UCB exploration achieves regret $\tilde{O}(\sqrt{H^3 SAT})$, where $S$ and $A$ are the numbers of states and actions, $H$ is the number of steps per episode, and $T$ is the total number of steps. This sample efficiency matches the optimal regret that can be achieved by any model-based approach, up to a single $\sqrt{H}$ factor. To the best of our knowledge, this is the first analysis in the model-free setting that establishes $\sqrt{T}$ regret without requiring access to a "simulator."

Kevin Jamieson, Matthew Malloy, Robert Nowak et al.

Sampling from distributions to find the one with the largest mean arises in a broad range of applications, and it can be mathematically modeled as a multi-armed bandit problem in which each distribution is associated with an arm. This paper studies the sample complexity of identifying the best arm (largest mean) in a multi-armed bandit problem. Motivated by large-scale applications, we are especially interested in identifying situations where the total number of samples that are necessary and sufficient to find the best arm scale linearly with the number of arms. We present a single-parameter multi-armed bandit model that spans the range from linear to superlinear sample complexity. We also give a new algorithm for best arm identification, called PRISM, with linear sample complexity for a wide range of mean distributions. The algorithm, like most exploration procedures for multi-armed bandits, is adaptive in the sense that the next arms to sample are selected based on previous samples. We compare the sample complexity of adaptive procedures with simpler non-adaptive procedures using new lower bounds. For many problem instances, the increased sample complexity required by non-adaptive procedures is a polynomial factor of the number of arms.

Sebastien Bubeck, Damien Ernst, Aurelien Garivier

We consider an original problem that arises from the issue of security analysis of a power system and that we name optimal discovery with probabilistic expert advice. We address it with an algorithm based on the optimistic paradigm and on the Good-Turing missing mass estimator. We prove two different regret bounds on the performance of this algorithm under weak assumptions on the probabilistic experts. Under more restrictive hypotheses, we also prove a macroscopic optimality result, comparing the algorithm both with an oracle strategy and with uniform sampling. Finally, we provide numerical experiments illustrating these theoretical findings.

Sebastien Bubeck, Aleksandrs Slivkins

We present a new bandit algorithm, SAO (Stochastic and Adversarial Optimal), whose regret is, essentially, optimal both for adversarial rewards and for stochastic rewards. Specifically, SAO combines the square-root worst-case regret of Exp3 (Auer et al., SIAM J. on Computing 2002) and the (poly)logarithmic regret of UCB1 (Auer et al., Machine Learning 2002) for stochastic rewards. Adversarial rewards and stochastic rewards are the two main settings in the literature on (non-Bayesian) multi-armed bandits. Prior work on multi-armed bandits treats them separately, and does not attempt to jointly optimize for both. Our result falls into a general theme of achieving good worst-case performance while also taking advantage of "nice" problem instances, an important issue in the design of algorithms with partially known inputs.