语言模型既展示了定量的改进，又展示了新的定性功能，随着规模的增加。尽管它们具有潜在的变革性影响，但这些新能力的特征却很差。为了为未来的研究提供信息，为破坏性的新模型能力做准备，并改善社会有害的效果，至关重要的是，我们必须了解目前和近乎未来的能力和语言模型的局限性。为了应对这一挑战，我们介绍了超越模仿游戏基准（Big Bench）。 Big Bench目前由204个任务组成，由132家机构的442位作者贡献。任务主题是多样的，从语言学，儿童发展，数学，常识性推理，生物学，物理学，社会偏见，软件开发等等。 Big-Bench专注于被认为超出当前语言模型的功能的任务。我们评估了OpenAI的GPT型号，Google内部密集变压器体系结构和大型基础上的开关稀疏变压器的行为，跨越了数百万到数十亿个参数。此外，一个人类专家评估者团队执行了所有任务，以提供强大的基准。研究结果包括：模型性能和校准都随规模改善，但绝对的术语（以及与评估者的性能相比）；在模型类中的性能非常相似，尽管带有稀疏性。逐渐和预测的任务通常涉及大量知识或记忆成分，而在临界规模上表现出“突破性”行为的任务通常涉及多个步骤或组成部分或脆性指标；社交偏见通常会随着含糊不清的环境而随着规模而增加，但这可以通过提示来改善。

X射线微型计算机断层扫描（Micro-CT）已被广泛利用，以在地下多孔岩石中表征孔隙尺度几何形状。使用深度学习的超分辨率（SR）方法的最新进程允许在大型空间尺度上进行数字增强低分辨率（LR）图像，从而创建与高分辨率（HR）地理真理相当的SR图像。这避免了传统的解决方案和视野折衷。出色的问题是使用配对（已注册的）LR和HR数据，这些数据通常需要在此类方法的训练步骤中，但难以获得。在这项工作中，我们严格比较两种不同的最先进的SR深度学习技术，使用两者和未配对数据，具有类似于类似的地面真理数据。第一方法需要配对的图像来训练卷积神经网络（CNN），而第二种方法使用未配对的图像来训练生成的对抗网络（GaN）。使用具有复杂的微孔纹理的微型CT碳酸盐岩样品进行比较两种方法。我们实现了基于图像的各种图像和数值验证和实验验证，以定量评估两种方法的物理精度和敏感性。我们的定量结果表明，未配对GaN方法可以将超分辨率图像重建为精确，如配对的CNN方法，具有可比的训练时间和数据集要求。这将使用未配对的深度学习方法解除微型CT图像增强的新应用;数据处理阶段不再需要图像注册。来自数据存储平台的解耦图像可以更有效地利用用于培训SR数字岩体应用的网络。这为异构多孔介质中的多尺度流模拟各种应用开辟了新的途径。

X射线微型计算机断层摄影成像中存在固有的视野和分辨率折衷，这限制了多尺寸多孔系统的表征，分析和模型开发。在本文中，我们通过开发3D增强的深层超分辨率（EDSR）卷积神经网络来克服这些权衡来通过来自低分辨率数据的大型空间尺度创建增强的高分辨率数据。配对高分辨率（HR，2 $ \ MU $ M）和低分辨率（LR，6 $ \ MU $ M）来自Bentheimer Rock样本的图像数据用于培训网络。来自训练样本的未见LR和HR数据以及具有不同微结构的另一个样本，用于验证具有各种度量的网络：文本分析，分段行为和孔网络模型（PNM）多相流模拟。经过验证的EDSR网络用于为每个长度为6-7厘米的全核样品生成约1000个高分辨率转速子图像（总图像大小为约6000x6000x32000体素）。每个子培养物都具有从PNMS预测的不同的岩石物理特性，它们组合以创建每个样本的3D连续级模型。在一系列分数流动下模拟低毛细管数不混溶的流动，并直接在1：1的基础上与实验压力和3D饱和度进行比较。 EDSR产生的模型比在存在异质性存在下预测实验行为的基础LR模型更准确，特别是在遇到孔隙尺寸的广泛分布的流动状态下。该模型通常在预测到在实验重复性和三个数量级的实验重复性和相对渗透率内的饱和度准确。所示的工作流程是一个完全预测的，无需校准，并且打开了在真正的多尺度异构系统中的图像，模拟和分析流动的可能性。

Managing novelty in perception-based human activity recognition (HAR) is critical in realistic settings to improve task performance over time and ensure solution generalization outside of prior seen samples. Novelty manifests in HAR as unseen samples, activities, objects, environments, and sensor changes, among other ways. Novelty may be task-relevant, such as a new class or new features, or task-irrelevant resulting in nuisance novelty, such as never before seen noise, blur, or distorted video recordings. To perform HAR optimally, algorithmic solutions must be tolerant to nuisance novelty, and learn over time in the face of novelty. This paper 1) formalizes the definition of novelty in HAR building upon the prior definition of novelty in classification tasks, 2) proposes an incremental open world learning (OWL) protocol and applies it to the Kinetics datasets to generate a new benchmark KOWL-718, 3) analyzes the performance of current state-of-the-art HAR models when novelty is introduced over time, 4) provides a containerized and packaged pipeline for reproducing the OWL protocol and for modifying for any future updates to Kinetics. The experimental analysis includes an ablation study of how the different models perform under various conditions as annotated by Kinetics-AVA. The protocol as an algorithm for reproducing experiments using the KOWL-718 benchmark will be publicly released with code and containers at https://github.com/prijatelj/human-activity-recognition-in-an-open-world. The code may be used to analyze different annotations and subsets of the Kinetics datasets in an incremental open world fashion, as well as be extended as further updates to Kinetics are released.

Naturally-occurring information-seeking questions often contain questionable assumptions -- assumptions that are false or unverifiable. Questions containing questionable assumptions are challenging because they require a distinct answer strategy that deviates from typical answers to information-seeking questions. For instance, the question "When did Marie Curie discover Uranium?" cannot be answered as a typical when question without addressing the false assumption "Marie Curie discovered Uranium". In this work, we propose (QA)$^2$ (Question Answering with Questionable Assumptions), an open-domain evaluation dataset consisting of naturally-occurring search engine queries that may or may not contain questionable assumptions. To be successful on (QA)$^2$, systems must be able to detect questionable assumptions and also be able to produce adequate responses for both typical information-seeking questions and ones with questionable assumptions. We find that current models do struggle with handling questionable assumptions -- the best performing model achieves 59% human rater acceptability on abstractive QA with (QA)$^2$ questions, leaving substantial headroom for progress.

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.

By transferring knowledge from large, diverse, task-agnostic datasets, modern machine learning models can solve specific downstream tasks either zero-shot or with small task-specific datasets to a high level of performance. While this capability has been demonstrated in other fields such as computer vision, natural language processing or speech recognition, it remains to be shown in robotics, where the generalization capabilities of the models are particularly critical due to the difficulty of collecting real-world robotic data. We argue that one of the keys to the success of such general robotic models lies with open-ended task-agnostic training, combined with high-capacity architectures that can absorb all of the diverse, robotic data. In this paper, we present a model class, dubbed Robotics Transformer, that exhibits promising scalable model properties. We verify our conclusions in a study of different model classes and their ability to generalize as a function of the data size, model size, and data diversity based on a large-scale data collection on real robots performing real-world tasks. The project's website and videos can be found at robotics-transformer.github.io

The simple idea that not all things are equally difficult has surprising implications when applied in a fairness context. In this work we explore how "difficulty" is model-specific, such that different models find different parts of a dataset challenging. When difficulty correlates with group information, we term this difficulty disparity. Drawing a connection with recent work exploring the inductive bias towards simplicity of SGD-trained models, we show that when such a disparity exists, it is further amplified by commonly-used models. We quantify this amplification factor across a range of settings aiming towards a fuller understanding of the role of model bias. We also present a challenge to the simplifying assumption that "fixing" a dataset is sufficient to ensure unbiased performance.

Existing statistical methods can be used to estimate a policy, or a mapping from covariates to decisions, which can then instruct decision makers. There is great interest in using such data-driven policies in healthcare. In healthcare, however, it is often important to explain to the healthcare provider, and to the patient, how a new policy differs from the current standard of care. This end is facilitated if one can pinpoint the aspects (i.e., parameters) of the policy that change most when moving from the standard of care to the new, suggested policy. To this end, we adapt ideas from Trust Region Policy Optimization. In our work, however, unlike in Trust Region Policy Optimization, the difference between the suggested policy and standard of care is required to be sparse, aiding with interpretability. In particular, we trade off between maximizing expected reward and minimizing the $L_1$ norm divergence between the parameters of the two policies. This yields "relative sparsity," where, as a function of a tuning parameter, $\lambda$, we can approximately control the number of parameters in our suggested policy that differ from their counterparts in the standard of care. We develop our methodology for the observational data setting. We propose a problem-specific criterion for selecting $\lambda$, perform simulations, and illustrate our method with a real, observational healthcare dataset, deriving a policy that is easy to explain in the context of the current standard of care. Our work promotes the adoption of data-driven decision aids, which have great potential to improve health outcomes.