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.

关于语言模型（LMS）的先前工作表明，对大量不同任务进行培训可以改善新任务的少数学习（FSL）表现。我们将其带到极端，从Internet表中自动提取413,299个任务 - 比下一大公共数据集的数量级多。在生成的数据集上进行填充可改善自然语言处理（NLP）任务的FSL性能，但与数据集量表不成比例。实际上，我们发现数据集的狭窄子集有时比数据集更优于多样化的数据集。例如，从support.google.com上对软件文档进行填充，在52个下游任务上提高了FSL性能，平均 +7.5％，该任务击败了40个人类策划的NLP数据集（ +6.7％）的培训。在各种狭窄数据集上进行填充会导致在测试任务之间进行类似的广泛改进，这表明收益不是来自域的适应性，而是适应FSL。我们没有观察到导致FSL收益的数据集之间的清晰模式，而是关于某些数据为何有助于FSL的开放问题。

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。

从头开始解决复杂问题通常是有挑战性的，但如果我们可以访问其解决方案的其他类似问题，则更容易 - 一种称为基于案例的推理（CBR）的范式。我们提出了一种神经象征性的CBR方法（CBR-KBQA），用于在大知识库上应答。 CBR-KBQA由非参数内存组成，该内存存储案例（问题和逻辑表单）和参数模型，该参数模型可以通过检索与其相关的案例来为新问题生成逻辑表单。在包含复杂问题的几个KBQA数据集上，CBR-KBQA实现了竞争性能。例如，在ComplexWebQuestions数据集上，CBR-KBQA以11 \％的准确度优于当前最新状态。此外，我们表明CBR-KBQA能够使用新案例\ EMPH {没有}任何进一步的培训：通过在案例存储器中纳入一些人类标记的示例，CBR-KBQA能够成功地生成包含未经看线KB实体的逻辑表格以及关系。

We introduce a general-purpose conditioning method for neural networks called FiLM: Feature-wise Linear Modulation. FiLM layers influence neural network computation via a simple, feature-wise affine transformation based on conditioning information. We show that FiLM layers are highly effective for visual reasoning -answering image-related questions which require a multi-step, high-level process -a task which has proven difficult for standard deep learning methods that do not explicitly model reasoning. Specifically, we show on visual reasoning tasks that FiLM layers 1) halve state-of-theart error for the CLEVR benchmark, 2) modulate features in a coherent manner, 3) are robust to ablations and architectural modifications, and 4) generalize well to challenging, new data from few examples or even zero-shot.

Research has shown that climate change creates warmer temperatures and drier conditions, leading to longer wildfire seasons and increased wildfire risks in the United States. These factors have in turn led to increases in the frequency, extent, and severity of wildfires in recent years. Given the danger posed by wildland fires to people, property, wildlife, and the environment, there is an urgency to provide tools for effective wildfire management. Early detection of wildfires is essential to minimizing potentially catastrophic destruction. In this paper, we present our work on integrating multiple data sources in SmokeyNet, a deep learning model using spatio-temporal information to detect smoke from wildland fires. Camera image data is integrated with weather sensor measurements and processed by SmokeyNet to create a multimodal wildland fire smoke detection system. We present our results comparing performance in terms of both accuracy and time-to-detection for multimodal data vs. a single data source. With a time-to-detection of only a few minutes, SmokeyNet can serve as an automated early notification system, providing a useful tool in the fight against destructive wildfires.

The United States coastline spans 95,471 miles; a distance that cannot be effectively patrolled or secured by manual human effort alone. Unmanned Aerial Vehicles (UAVs) equipped with infrared cameras and deep-learning based algorithms represent a more efficient alternative for identifying and segmenting objects of interest - namely, ships. However, standard approaches to training these algorithms require large-scale datasets of densely labeled infrared maritime images. Such datasets are not publicly available and manually annotating every pixel in a large-scale dataset would have an extreme labor cost. In this work we demonstrate that, in the context of segmenting ships in infrared imagery, weakly-supervising an algorithm with sparsely labeled data can drastically reduce data labeling costs with minimal impact on system performance. We apply weakly-supervised learning to an unlabeled dataset of 7055 infrared images sourced from the Naval Air Warfare Center Aircraft Division (NAWCAD). We find that by sparsely labeling only 32 points per image, weakly-supervised segmentation models can still effectively detect and segment ships, with a Jaccard score of up to 0.756.

We present a human-in-the-loop evaluation framework for fact-checking novel misinformation claims and identifying social media messages that violate relevant policies. Our approach extracts structured representations of check-worthy claims, which are aggregated and ranked for review. Stance classifiers are then used to identify tweets supporting novel misinformation claims, which are further reviewed to determine whether they violate relevant policies. To demonstrate the feasibility of our approach, we develop a baseline system based on modern NLP methods for human-in-the-loop fact-checking in the domain of COVID-19 treatments. Using our baseline system, we show that human fact-checkers can identify 124 tweets per hour that violate Twitter's policies on COVID-19 misinformation. We will make our code, data, and detailed annotation guidelines available to support the evaluation of human-in-the-loop systems that identify novel misinformation directly from raw user-generated content.