As language models (LMs) scale, they develop many novel behaviors, good and bad, exacerbating the need to evaluate how they behave. Prior work creates evaluations with crowdwork (which is time-consuming and expensive) or existing data sources (which are not always available). Here, we automatically generate evaluations with LMs. We explore approaches with varying amounts of human effort, from instructing LMs to write yes/no questions to making complex Winogender schemas with multiple stages of LM-based generation and filtering. Crowdworkers rate the examples as highly relevant and agree with 90-100% of labels, sometimes more so than corresponding human-written datasets. We generate 154 datasets and discover new cases of inverse scaling where LMs get worse with size. Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down. Overall, LM-written evaluations are high-quality and let us quickly discover many novel LM behaviors.

As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels.

Developing safe and useful general-purpose AI systems will require us to make progress on scalable oversight: the problem of supervising systems that potentially outperform us on most skills relevant to the task at hand. Empirical work on this problem is not straightforward, since we do not yet have systems that broadly exceed our abilities. This paper discusses one of the major ways we think about this problem, with a focus on how to turn it into one that can be productively studied empirically. We first present an experimental design centered on choosing tasks for which human specialists succeed but unaided humans and current general AI systems fail. We then present a proof-of-concept experiment following meant to demonstrate a key feature of this experimental design and show its viability with two question-answering tasks: MMLU and time-limited QuALITY. On these tasks, we find that human participants who interact with an unreliable large-language-model dialog assistant through chat -- a trivial baseline strategy for scalable oversight -- substantially outperform both the model alone and their own unaided performance. These results are an encouraging sign that scalable oversight will be tractable to study with present models and bolster recent findings that large language models can productively assist humans with difficult tasks.

对抗性训练遭受了稳健的过度装备，这是一种现象，在训练期间鲁棒测试精度开始减少。在本文中，我们专注于通过使用常见的数据增强方案来减少强大的过度装备。我们证明，与先前的发现相反，当与模型重量平均结合时，数据增强可以显着提高鲁棒精度。此外，我们比较各种增强技术，并观察到空间组合技术适用于对抗性培训。最后，我们评估了我们在Cifar-10上的方法，而不是$ \ ell_ indty $和$ \ ell_2 $ norm-indeded扰动分别为尺寸$ \ epsilon = 8/255 $和$ \ epsilon = 128/255 $。与以前的最先进的方法相比，我们表现出+ 2.93％的绝对改善+ 2.93％，+ 2.16％。特别是，反对$ \ ell_ infty $ norm-indeded扰动尺寸$ \ epsilon = 8/255 $，我们的模型达到60.07％的强劲准确性而不使用任何外部数据。我们还通过这种方法实现了显着的性能提升，同时使用其他架构和数据集如CiFar-100，SVHN和TinyimageNet。

最近的工作认为，强大的培训需要比标准分类所需的数据集大得多。在CiFar-10和CiFar-100上，这转化为仅培训的型号之间的可稳健稳健精度差距，这些型号来自原始训练集的数据，那些从“80万微小图像”数据集（TI-80M）提取的附加数据培训。在本文中，我们探讨了单独培训的生成模型如何利用人为地提高原始训练集的大小，并改善对$ \ ell_p $ norm-inded扰动的对抗鲁棒性。我们确定了包含额外生成数据的充分条件可以改善鲁棒性，并证明可以显着降低具有额外实际数据训练的模型的强大准确性差距。令人惊讶的是，我们甚至表明即使增加了非现实的随机数据（由高斯采样产生）也可以改善鲁棒性。我们在Cifar-10，CiFar-100，SVHN和Tinyimagenet上评估我们的方法，而$ \ ell_ indty $和$ \ ell_2 $ norm-indeded扰动尺寸$ \ epsilon = 8/255 $和$ \ epsilon = 128/255 $分别。与以前的最先进的方法相比，我们以强大的准确性显示出大的绝对改进。反对$ \ ell_ \ infty $ norm-indeded扰动尺寸$ \ epsilon = 8/255 $，我们的车型分别在Cifar-10和Cifar-100上达到66.10％和33.49％（改善状态）最新美术+ 8.96％和+ 3.29％）。反对$ \ ell_2 $ norm-indeded扰动尺寸$ \ epsilon = 128/255 $，我们的型号在Cifar-10（+ 3.81％）上实现78.31％。这些结果击败了使用外部数据的最先前的作品。

现代神经网络Excel在图像分类中，但它们仍然容易受到常见图像损坏，如模糊，斑点噪音或雾。最近的方法关注这个问题，例如Augmix和Deepaulment，引入了在预期运行的防御，以期望图像损坏分布。相比之下，$ \ ell_p $ -norm界限扰动的文献侧重于针对最坏情况损坏的防御。在这项工作中，我们通过提出防范内人来调和两种方法，这是一种优化图像到图像模型的参数来产生对外损坏的增强图像的技术。我们理论上激发了我们的方法，并为其理想化版本的一致性以及大纲领提供了足够的条件。我们的分类机器在预期对CiFar-10-C进行的常见图像腐败基准上提高了最先进的，并改善了CIFAR-10和ImageNet上的$ \ ell_p $ -norm有界扰动的最坏情况性能。

Optical coherence tomography (OCT) captures cross-sectional data and is used for the screening, monitoring, and treatment planning of retinal diseases. Technological developments to increase the speed of acquisition often results in systems with a narrower spectral bandwidth, and hence a lower axial resolution. Traditionally, image-processing-based techniques have been utilized to reconstruct subsampled OCT data and more recently, deep-learning-based methods have been explored. In this study, we simulate reduced axial scan (A-scan) resolution by Gaussian windowing in the spectral domain and investigate the use of a learning-based approach for image feature reconstruction. In anticipation of the reduced resolution that accompanies wide-field OCT systems, we build upon super-resolution techniques to explore methods to better aid clinicians in their decision-making to improve patient outcomes, by reconstructing lost features using a pixel-to-pixel approach with an altered super-resolution generative adversarial network (SRGAN) architecture.

Micro-CT images of the renal arteries of intact rat kidneys, which had their vasculature injected with the contrast agent polymer Microfil, were characterized. Measurement of inter-branch segment properties and the hierarchical structure of the vessel trees were computed by an automated algorithmic approach. The perfusion territories of the different kidneys, as well as the local diameters of the segmented vasculature were mapped onto the representative structures and visually explored. Various parameters were compared in order to outline key geometrical properties, properties which were shown to not have a wide range of inter-specimen variation. It is shown that the fractal scaling in non-symmetric branching reveals itself differently, than in symmetric branching (e.g., in the lung the mean bronchial diameters at each generation are closely related). Also, perfused tissue is shown to have very little inter-specimen variation and therefore could be used in future studies related to characterizing various disease states of tissues and organs based on vascular branching geometry.

We present NusaCrowd, a collaborative initiative to collect and unite existing resources for Indonesian languages, including opening access to previously non-public resources. Through this initiative, we have has brought together 137 datasets and 117 standardized data loaders. The quality of the datasets has been assessed manually and automatically, and their effectiveness has been demonstrated in multiple experiments. NusaCrowd's data collection enables the creation of the first zero-shot benchmarks for natural language understanding and generation in Indonesian and its local languages. Furthermore, NusaCrowd brings the creation of the first multilingual automatic speech recognition benchmark in Indonesian and its local languages. Our work is intended to help advance natural language processing research in under-represented languages.

In a wide variety of fields, analysis of images involves defining a region and measuring its inherent properties. Such measurements include a region's surface area, curvature, volume, average gray and/or color scale, and so on. Furthermore, the subsequent subdivision of these regions is sometimes performed. These subdivisions are then used to measure local information, at even finer scales. However, simple griding or manual editing methods are typically used to subdivide a region into smaller units. The resulting subdivisions can therefore either not relate well to the actual shape or property of the region being studied (i.e., gridding methods), or be time consuming and based on user subjectivity (i.e., manual methods). The method discussed in this work extracts subdivisional units based on a region's general shape information. We present the results of applying our method to the medical image analysis of nested regions-of-interest of myocardial wall, where the subdivisions are used to study temporal and/or spatial heterogeneity of myocardial perfusion. This method is of particular interest for creating subdivision regions-of-interest (SROIs) when no variable intensity or other criteria within a region need be used to separate a particular region into subunits.