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.

神经网络经常将许多无关的概念包装到一个神经元中 - 一种令人困惑的现象被称为“多疾病”，这使解释性更具挑战性。本文提供了一个玩具模型，可以完全理解多义，这是由于模型在“叠加”中存储其他稀疏特征的结果。我们证明了相变的存在，与均匀多型的几何形状的令人惊讶的联系以及与对抗性例子联系的证据。我们还讨论了对机械解释性的潜在影响。

我们研究语言模型是否可以评估自己主张的有效性，并预测他们能够正确回答的问题。我们首先表明，当以正确的格式提供时，较大的模型在多样化的多项选择和True/False问题上进行了很好的校准。因此，我们可以通过要求模型首先提出答案，然后评估其答案正确的概率“ p（true）”来对开放式采样任务进行自我评估。我们发现在各种任务中，P（true）的表现，校准和缩放令人鼓舞。当我们允许模型考虑自己的许多样本之前，在预测一种特定可能性的有效性之前，自我评估的性能进一步改善。接下来，我们研究是否可以培训模型来预测“ P（ik）”，即“我知道”问题的概率，而无需参考任何特定提出的答案。模型在预测P（IK）方面表现良好，并且在跨任务中部分概括，尽管它们在新任务上的P（IK）校准方面遇到了困难。预测的p（IK）概率在存在相关的原始材料的情况下以及对数学单词问题解决方案的提示也适当增加。我们希望这些观察结果为培训更诚实的模型提供了基础，并研究了诚实对模型模仿人类写作以外的其他目标培训的案例的普遍性。

Relation extraction (RE), which has relied on structurally annotated corpora for model training, has been particularly challenging in low-resource scenarios and domains. Recent literature has tackled low-resource RE by self-supervised learning, where the solution involves pretraining the relation embedding by RE-based objective and finetuning on labeled data by classification-based objective. However, a critical challenge to this approach is the gap in objectives, which prevents the RE model from fully utilizing the knowledge in pretrained representations. In this paper, we aim at bridging the gap and propose to pretrain and finetune the RE model using consistent objectives of contrastive learning. Since in this kind of representation learning paradigm, one relation may easily form multiple clusters in the representation space, we further propose a multi-center contrastive loss that allows one relation to form multiple clusters to better align with pretraining. Experiments on two document-level RE datasets, BioRED and Re-DocRED, demonstrate the effectiveness of our method. Particularly, when using 1% end-task training data, our method outperforms PLM-based RE classifier by 10.5% and 5.8% on the two datasets, respectively.

Recent methods in self-supervised learning have demonstrated that masking-based pretext tasks extend beyond NLP, serving as useful pretraining objectives in computer vision. However, existing approaches apply random or ad hoc masking strategies that limit the difficulty of the reconstruction task and, consequently, the strength of the learnt representations. We improve upon current state-of-the-art work in learning adversarial masks by proposing a new framework that generates masks in a sequential fashion with different constraints on the adversary. This leads to improvements in performance on various downstream tasks, such as classification on ImageNet100, STL10, and CIFAR10/100 and segmentation on Pascal VOC. Our results further demonstrate the promising capabilities of masking-based approaches for SSL in computer vision.

Convolutional neural networks (CNNs) are currently among the most widely-used neural networks available and achieve state-of-the-art performance for many problems. While originally applied to computer vision tasks, CNNs work well with any data with a spatial relationship, besides images, and have been applied to different fields. However, recent works have highlighted how CNNs, like other deep learning models, are sensitive to noise injection which can jeopardise their performance. This paper quantifies the numerical uncertainty of the floating point arithmetic inaccuracies of the inference stage of DeepGOPlus, a CNN that predicts protein function, in order to determine its numerical stability. In addition, this paper investigates the possibility to use reduced-precision floating point formats for DeepGOPlus inference to reduce memory consumption and latency. This is achieved with Monte Carlo Arithmetic, a technique that experimentally quantifies floating point operation errors and VPREC, a tool that emulates results with customizable floating point precision formats. Focus is placed on the inference stage as it is the main deliverable of the DeepGOPlus model that will be used across environments and therefore most likely be subjected to the most amount of noise. Furthermore, studies have shown that the inference stage is the part of the model which is most disposed to being scaled down in terms of reduced precision. All in all, it has been found that the numerical uncertainty of the DeepGOPlus CNN is very low at its current numerical precision format, but the model cannot currently be reduced to a lower precision that might render it more lightweight.

Deep learning classifiers provide the most accurate means of automatically diagnosing diabetic retinopathy (DR) based on optical coherence tomography (OCT) and its angiography (OCTA). The power of these models is attributable in part to the inclusion of hidden layers that provide the complexity required to achieve a desired task. However, hidden layers also render algorithm outputs difficult to interpret. Here we introduce a novel biomarker activation map (BAM) framework based on generative adversarial learning that allows clinicians to verify and understand classifiers decision-making. A data set including 456 macular scans were graded as non-referable or referable DR based on current clinical standards. A DR classifier that was used to evaluate our BAM was first trained based on this data set. The BAM generation framework was designed by combing two U-shaped generators to provide meaningful interpretability to this classifier. The main generator was trained to take referable scans as input and produce an output that would be classified by the classifier as non-referable. The BAM is then constructed as the difference image between the output and input of the main generator. To ensure that the BAM only highlights classifier-utilized biomarkers an assistant generator was trained to do the opposite, producing scans that would be classified as referable by the classifier from non-referable scans. The generated BAMs highlighted known pathologic features including nonperfusion area and retinal fluid. A fully interpretable classifier based on these highlights could help clinicians better utilize and verify automated DR diagnosis.

We identify the task of measuring data to quantitatively characterize the composition of machine learning data and datasets. Similar to an object's height, width, and volume, data measurements quantify different attributes of data along common dimensions that support comparison. Several lines of research have proposed what we refer to as measurements, with differing terminology; we bring some of this work together, particularly in fields of computer vision and language, and build from it to motivate measuring data as a critical component of responsible AI development. Measuring data aids in systematically building and analyzing machine learning (ML) data towards specific goals and gaining better control of what modern ML systems will learn. We conclude with a discussion of the many avenues of future work, the limitations of data measurements, and how to leverage these measurement approaches in research and practice.