我们呈现GSPMD，一种用于公共机器学习计算的自动，基于编译的并行化系统。它允许用户以与单个设备的方式相同的方式编写程序，然后通过关于如何分发Tensors的一些注释来提供提示，基于哪个GSPMD将并行化计算。其分区的表示简单尚不一般，允许它在各种模型上表达并行性的不同或混合范式。GSPMD基于有限的用户注释为每个运算符的分区Inventing，使得缩放现有的单设备程序方便。它解决了生产使用的几种技术挑战，允许GSPMD实现50％至62％的计算利用率，用于高达2048个云TPUv3核心，适用于高达1万亿参数的模型。

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.

Humans form mental images of 3D scenes to support counterfactual imagination, planning, and motor control. Our abilities to predict the appearance and affordance of the scene from previously unobserved viewpoints aid us in performing manipulation tasks (e.g., 6-DoF kitting) with a level of ease that is currently out of reach for existing robot learning frameworks. In this work, we aim to build artificial systems that can analogously plan actions on top of imagined images. To this end, we introduce Mental Imagery for Robotic Affordances (MIRA), an action reasoning framework that optimizes actions with novel-view synthesis and affordance prediction in the loop. Given a set of 2D RGB images, MIRA builds a consistent 3D scene representation, through which we synthesize novel orthographic views amenable to pixel-wise affordances prediction for action optimization. We illustrate how this optimization process enables us to generalize to unseen out-of-plane rotations for 6-DoF robotic manipulation tasks given a limited number of demonstrations, paving the way toward machines that autonomously learn to understand the world around them for planning actions.

Advancements in reinforcement learning (RL) have inspired new directions in intelligent automation of network defense. However, many of these advancements have either outpaced their application to network security or have not considered the challenges associated with implementing them in the real-world. To understand these problems, this work evaluates several RL approaches implemented in the second edition of the CAGE Challenge, a public competition to build an autonomous network defender agent in a high-fidelity network simulator. Our approaches all build on the Proximal Policy Optimization (PPO) family of algorithms, and include hierarchical RL, action masking, custom training, and ensemble RL. We find that the ensemble RL technique performs strongest, outperforming our other models and taking second place in the competition. To understand applicability to real environments we evaluate each method's ability to generalize to unseen networks and against an unknown attack strategy. In unseen environments, all of our approaches perform worse, with degradation varied based on the type of environmental change. Against an unknown attacker strategy, we found that our models had reduced overall performance even though the new strategy was less efficient than the ones our models trained on. Together, these results highlight promising research directions for autonomous network defense in the real world.

Measuring and monitoring soil organic carbon is critical for agricultural productivity and for addressing critical environmental problems. Soil organic carbon not only enriches nutrition in soil, but also has a gamut of co-benefits such as improving water storage and limiting physical erosion. Despite a litany of work in soil organic carbon estimation, current approaches do not generalize well across soil conditions and management practices. We empirically show that explicit modeling of cause-and-effect relationships among the soil processes improves the out-of-distribution generalizability of prediction models. We provide a comparative analysis of soil organic carbon estimation models where the skeleton is estimated using causal discovery methods. Our framework provide an average improvement of 81% in test mean squared error and 52% in test mean absolute error.

Recent developments in natural language generation (NLG) using neural language models have brought us closer than ever to the goal of building AI-powered creative writing tools. However, most prior work on human-AI collaboration in the creative writing domain has evaluated new systems with amateur writers, typically in contrived user studies of limited scope. In this work, we commissioned 13 professional, published writers from a diverse set of creative writing backgrounds to craft stories using Wordcraft, a text editor with built-in AI-powered writing assistance tools. Using interviews and participant journals, we discuss the potential of NLG to have significant impact in the creative writing domain--especially with respect to brainstorming, generation of story details, world-building, and research assistance. Experienced writers, more so than amateurs, typically have well-developed systems and methodologies for writing, as well as distinctive voices and target audiences. Our work highlights the challenges in building for these writers; NLG technologies struggle to preserve style and authorial voice, and they lack deep understanding of story contents. In order for AI-powered writing assistants to realize their full potential, it is essential that they take into account the diverse goals and expertise of human writers.

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。

我们提出语言学家，这是一种通过微调Alexatm 5B生成带注释数据的方法，用于生成意图分类和插槽标记（IC+ST），这是一种5亿参数的多语言序列到序列（SEQ2SEQ）模型，在灵活的指令上迅速的。在SNIP数据集的10次新颖意图设置中，语言学家超过了最新的方法（反向翻译和示例外推），可以通过宽阔的边距，显示出IC回忆中+1.9点的目标意图的绝对改善ST F1分数和+2.5分。在MATIS ++数据集的零击跨语言设置中，语言学家表现出强大的机器翻译基线，插槽对齐的基线是+4.14的+4.14点在6个语言上绝对在ST F1分数上，同时在IC上匹配IC的性能。最后，我们在用于对话代理IC+ST的内部大规模多语言数据集上验证了我们的结果，并显示了使用背面翻译，释义和插槽目录重新采样采样的基线的显着改进。据我们所知，我们是第一个展示大规模SEQ2SEQ模型的指导微调的人，以控制多语言意图和插槽标记的数据生成的输出。