Modern Large Language Models (LLMs) are safety trained using Supervised Fine Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) to mitigate harmful, undesirable, or otherwise disallowed outputs [2]-[4]. Despite ongoing advances in these methods, today's models still exhibit safety shortcomings: they can be tricked into revealing harmful content, often refuse legitimate requests, and remain vulnerable to jailbreak attacks [5]-[8]. We argue that many of these failures arise from two limitations in modern safety training. First, LLMs must respond instantly to user requests using a fixed amount of compute, without deliberation even for complex safety scenarios. Second, LLMs must infer underlying safety standards indirectly from large sets of labeled examples, rather than directly learning the safety specifications that govern them. This reliance on implicit, pattern-based learning leads to poor data efficiency and makes it challenging for models to generalize when facing unfamiliar scenarios or adversarial attacks. We propose deliberative alignment, a training approach that teaches LLMs to explicitly reason through safety specifications before producing an answer. By applying this method to OpenAI's o-series models [1], we enable them to use chain-of-thought (CoT) reasoning to examine user prompts, identify relevant policy guidelines, and generate safer responses (e.g., Figure 1).

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.

An open problem in artificial intelligence is how to train language models that produce responses that are factually correct. Current frontier models sometimes produce false outputs or answers that are not substantiated by evidence, a problem known as "hallucinations." Such hallucinations are one of the major barriers for broader adoption of general forms artificial intelligence like large language models. Factuality is a complicated topic because it is hard to measure--evaluating the factuality of any given arbitrary claim can be challenging, and language models often generate long completions that contain dozens of factual claims. In this work we will sidestep the open-endedness of language models by considering only short, fact-seeking questions with a single answer. This reduction of scope is important because it makes measuring factuality much more tractable, albeit at the cost of leaving open research questions such as whether improved behavior on short-form factuality generalizes to long-form factuality. We present a benchmark called SimpleQA, which contains 4,326 short, fact-seeking questions. SimpleQA was designed with a few important properties in mind: High correctness. Reference answers to questions are determined by two independent AI trainers, and questions were written in such a way that the predicted answers are easily gradable.

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.

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.

Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world. In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt. Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity.AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers. Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future work, we release FreshQA at github.com/freshllms/freshqa and commit to updating it at regular intervals.

The strong performance (Chowdhery et al., 2022; Nostalgebraist, 2022; OpenAI, 2023; Google, 2023), and emergent abilities (Wei et al., 2022) of modern language models (LMs) depend on self-supervised pretraining on massive text datasets. All model developers implicitly or explicitly decide the composition of these datasets: what data sources to include, whether to filter for attributes such as quality and toxicity, and when to gather new documents. While many of the most prominent models do not document their curation procedures (OpenAI, 2023; Google, 2023), or only document which procedures they used (Brown et al., 2020; Nostalgebraist, 2022; Scao et al., 2022; Touvron et al., 2023), they rarely document why they chose those protocols or what effect they had. This documentation debt leaves practitioners to be guided by intuitions and precedents, neither thoroughly evaluated (Bandy and Vincent, 2021; Sambasivan et al., 2021). Given the outsized and fundamental role of pretraining data in modern LMs, we believe this neglectful practice has detracted from responsible data use and hampered effective model development (Rogers, 2021; Gebru et al., 2021; Bender and Friedman, 2018). Among the small number of general-purpose LMs dominating community use and discussion, the prevailing focus has been on the scale of pretraining data and number of optimization steps (Brown et al., 2020; Nostalgebraist, 2022; Google, 2023). In this work, we systematically test how common data design decisions affect model performance--specifically: the time of collection, content filtering strategy (toxicity/quality), and domain composition. We study the impacts in two ways.

We present COVID-Q, a set of 1,690 questions about COVID-19 from 13 sources, which we annotate into 15 question categories and 207 question clusters. The most common questions in our dataset asked about transmission, prevention, and societal effects of COVID, and we found that many questions that appeared in multiple sources were not answered by any FAQ websites of reputable organizations such as the CDC and FDA. We post our dataset publicly at https://github.com/JerryWeiAI/COVID-Q. For classifying questions into 15 categories, a BERT baseline scored 58.1% accuracy when trained on 20 examples per category, and for a question clustering task, a BERT + triplet loss baseline achieved 49.5% accuracy. We hope COVID-Q can help either for direct use in developing applied systems or as a domain-specific resource for model evaluation.

Sparse Mixture-of-Experts (MoE) is a neural architecture design that can be utilized to add learnable parameters to Large Language Models (LLMs) without increasing inference cost. Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE models benefit more from instruction tuning than dense models. In particular, we conduct empirical studies across three experimental setups: (i) Direct finetuning on individual downstream tasks devoid of instruction tuning; (ii) Instructiontuning followed by in-context few-shot or zero-shot generalization on downstream tasks; and (iii) Instruction tuning supplemented by further finetuning on individual downstream tasks. In the first scenario, MoE models overall underperform dense models of identical computational capacity. This narrative, however, dramatically changes with the introduction of instruction tuning (second and third scenario), used independently or in conjunction with task-specific finetuning. Our most powerful model, FLAN-MOE-32B, surpasses the performance of FLAN-PALM-62B on four benchmark tasks, while using only a third of the FLOPs. The advancements embodied byFLAN-MOE inspire a reevaluation of the design principles of large-scale, high-performance language models in the framework of task-agnostic learning.

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.