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.

“感应头”是注意力头，它实现了一种简单的算法来完成令牌序列，例如[a] [b] ... [a] - > [b]。在这项工作中，我们提供了一个假设的初步和间接证据，即诱导头可能构成大型大型变压器模型中所有“文本学习”中大多数的机制（即减少在增加代币指数时损失的损失）。我们发现，诱导头在与秘密学习能力突然急剧上的急剧上升的位置完全相同，这是训练损失的颠簸。我们提出了六种互补的证据，认为诱导头可能是任何大小的变压器模型中一般性内部学习的机理来源。对于仅关注的小型模型，我们提供了有力的因果证据。对于具有MLP的较大模型，我们提供相关证据。

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。

大型变压器模型实现了自然语言理解任务的最新状态，并越来越成为建模源代码的基线模型体系结构。通常，变压器在大型无监督的语料库中进行预训练，学习令牌表示和与通常可用的文本相关的转换，然后对特定的下游感兴趣的任务进行微调。虽然微调是一种尝试将模型调整为新领域的久经考验的方法（例如，在给定主题上提出问题，概括仍然是一个持续的挑战。在本文中，我们探索并评估了变形金刚的模型以进行个性化。在为Java方法生成单元测试的背景下，我们评估学习以使用多种个性化技术为特定的软件项目个性化。我们考虑三种关键方法：（i）自定义微调，这允许调整所有模型参数； （ii）轻巧的微调，它冻结了大多数模型的参数，可以单独调整令牌嵌入和SoftMax层或单独的最终层； （iii）前缀调整，该调谐使模型参数冻结，但优化了小型项目特定的前缀矢量。这些技术中的每一个都提供了总计算成本和预测性能的权衡，我们通过代码和特定任务指标，培训时间和总计算操作进行评估。我们比较了这些微调策略以生成代码，并讨论了各种部署方案中每个策略的潜在概括和成本益处。

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。

鉴于大型语言模型的广泛能力，应该有可能朝着一般的文本的助手工作，这些助手与人类价值一致，这意味着它是有帮助，诚实的和无害的。在此方向上的初始遗传，我们研究简单的基线技术和评估，例如提示。我们发现，从模型规模增加适度的干预措施的好处，概括为各种对准评估，并不会损害大型模型的性能。接下来，我们调查与对齐，比较仿制，二进制歧视和排名偏好建模相关的几个培训目标的缩放趋势。我们发现排名优先级模型比模仿学习更好地表现得多，并且通常以模型大小更有利地缩放。相比之下，二进制歧视通常与模仿学习非常类似地执行和缩放。最后，我们研究了一种“偏好模型预训练阶段的培训阶段，其目的是在对人偏好的芬明时提高样本效率。

A popular approach to creating a zero-shot cross-language retrieval model is to substitute a monolingual pretrained language model in the retrieval model with a multilingual pretrained language model such as Multilingual BERT. This multilingual model is fined-tuned to the retrieval task with monolingual data such as English MS MARCO using the same training recipe as the monolingual retrieval model used. However, such transferred models suffer from mismatches in the languages of the input text during training and inference. In this work, we propose transferring monolingual retrieval models using adapters, a parameter-efficient component for a transformer network. By adding adapters pretrained on language tasks for a specific language with task-specific adapters, prior work has shown that the adapter-enhanced models perform better than fine-tuning the entire model when transferring across languages in various NLP tasks. By constructing dense retrieval models with adapters, we show that models trained with monolingual data are more effective than fine-tuning the entire model when transferring to a Cross Language Information Retrieval (CLIR) setting. However, we found that the prior suggestion of replacing the language adapters to match the target language at inference time is suboptimal for dense retrieval models. We provide an in-depth analysis of this discrepancy between other cross-language NLP tasks and CLIR.

Answering complex questions often requires multi-step reasoning in order to obtain the final answer. Most research into decompositions of complex questions involves open-domain systems, which have shown success in using these decompositions for improved retrieval. In the machine reading setting, however, work to understand when decompositions are helpful is understudied. We conduct experiments on decompositions in machine reading to unify recent work in this space, using a range of models and datasets. We find that decompositions can be helpful in the few-shot case, giving several points of improvement in exact match scores. However, we also show that when models are given access to datasets with around a few hundred or more examples, decompositions are not helpful (and can actually be detrimental). Thus, our analysis implies that models can learn decompositions implicitly even with limited data.