迄今为止对文本生成的评估主要集中在依次创建的内容上，而不是对文本的改进。但是，写作自然是一个迭代和增量过程，需要在不同的模块化技能上进行专业知识，例如修复过时的信息或使样式更加一致。即便如此，对模型执行这些技能和编辑能力的模型能力的全面评估仍然很少。这项工作介绍了EditeVal：基于指导的，基准和评估套件，该套件利用现有的现有和新数据集自动评估编辑功能，例如使文本更具凝聚力和释义。我们评估了几种预训练的模型，这表明指令和同伴表现最好，但是大多数基准都落在监督的SOTA以下，尤其是在中和和更新信息时。我们的分析还表明，用于编辑任务的常用指标并不总是很好地关联，并且对具有最高性能的提示的优化并不一定带来对不同模型的最强鲁棒性。通过发布此基准和公开可用的排行榜挑战，我们希望在开发能够迭代和更可控制的编辑模型中解锁未来的研究。

文本内容通常是协作写作过程的输出：我们从初始草稿开始，提出建议并反复进行更改。不可知的是，当今的语言模型只能产生最终结果。结果，他们缺乏对协作写作至关重要的几种能力：他们无法更新现有文本，难以控制和无法进行口头计划或解释其行为。为了解决这些缺点，我们介绍了Peer，这是一种协作语言模型，经过训练以模仿整个写作过程本身：Peer可以编写草稿，添加建议，提出编辑并为其行为提供解释。至关重要的是，我们训练多个同伴能够填补写作过程的各个部分的实例，从而可以使用自训练技术来提高培训数据的质量，数量和多样性。这通过使其适用于没有编辑历史的域，并提高其遵循说明，编写有用的评论并解释其动作的能力，从而释放了Peer的全部潜力。我们表明，同行在各个领域和编辑任务上取得了强大的性能。

大型语言模型在各种任务上显示出令人印象深刻的几次结果。但是，当知识是此类结果的关键时，就像问题回答和事实检查之类的任务一样，似乎需要存储知识的大量参数计数。众所周知，检索增强模型可以在不需要多个参数的情况下在知识密集的任务上表现出色，但是目前尚不清楚它们是否在几个弹药设置中工作。在这项工作中，我们介绍了地图集，这是一个经过精心设计和预先训练的增强语言模型，能够通过很少的培训示例学习知识密集型任务。我们对包括MMLU，苏格兰短裙和归类等各种任务进行评估，并研究文档索引内容的影响，表明它可以很容易地进行更新。值得注意的是，在自然问题上仅使用64个示例在自然问题上达到超过42 \％的准确性，尽管参数少了50倍，但比540B参数模型的表现优于540b参数模型。

在本文中，我们考虑了从长时间的视频到几分钟的长视频进行分类的问题（例如，烹饪不同的食谱，烹饪不同的食谱，进行不同的家庭装修，创建各种形式的艺术和手工艺品）。准确地对这些活动进行分类，不仅需要识别构成任务的单个步骤，还需要捕获其时间依赖性。这个问题与传统的动作分类大不相同，在传统的动作分类中，模型通常在跨越几秒钟的视频上进行了优化，并且手动修剪以包含简单的原子动作。虽然步骤注释可以使模型的培训能够识别程序活动的各个步骤，但由于长时间视频中手动注释时间界的超级注释，因此该领域的现有大规模数据集不包括此类段标签。为了解决这个问题，我们建议通过利用文本知识库（Wikihow）的遥远监督来自动确定教学视频中的步骤，其中包括对执行各种复杂活动所需的步骤的详细描述。我们的方法使用语言模型来匹配视频中自动转录的语音，以在知识库中逐步描述。我们证明，经过训练的视频模型可以识别这些自动标记的步骤（无手动监督）产生了在四个下游任务上实现卓越的概括性能的表示：识别程序活动，步骤分类，步骤预测和以自我为中心的视频分类。

为了解决现实世界应用需求的日益增长，知识密集型NLP（KI-NLP）的研究应通过捕获真正开放域环境的挑战：网络规模知识，结构缺乏，质量不一致，和噪音。为此，我们提出了一种新的设置，用于评估现有的KI-NLP任务，其中我们将背景语料库概括为通用Web快照。我们重新保证Kilt，最初为维基百科最初开发的标准Ki-NLP基准测试，并要求系统使用CCNet的子集 - 球体语料库 - 作为知识源。与维基百科相比，球体是较大的数量级，更好地反映了互联网上的全部知识。我们发现，尽管潜在的覆盖范围，规模挑战，结构缺乏，质量较低，来自领域的检索可以实现最先进的检索系统，以匹配和甚至优于基于Wikipedia的模型在几个kilt上任务 - 即使我们积极过滤看起来像维基百科的内容。我们还观察到Wikipedia的单一密集通道指数可以胜过稀疏的BM25版本，而在球体上尚不实现。为了促进进一步研究该领域，并尽量减少社区对专有黑匣子搜索引擎的依赖，我们将分享我们的指数，评估指标和基础设施。

我们提出了Drboost，一个受升压启发的密集检索合奏。Drboost在阶段接受培训：通过仅关注当前合奏制作的检索错误来依次学习和专注于每个组件模型。最终的表示是所有组件模型的输出矢量的串联，使其成为测试时间标准密集检索器的替代品。与标准密集检索模型相比，Drboost享有几个优点。它产生的表示是4x更紧凑，同时提供可比的检索结果。它还在具有粗量化的近似搜索下进行令人惊讶的良好，从而减少另一个4x的延迟和带宽需求。在实践中，这可以在从内存中服务索引之间的服务指数之间的区别，为更便宜的部署铺平道路。

深度强化学习（RL）的进展是通过用于培训代理商的具有挑战性的基准的可用性来驱动。但是，社区广泛采用的基准未明确设计用于评估RL方法的特定功能。虽然存在用于评估RL的特定打开问题的环境（例如探索，转移学习，无监督环境设计，甚至语言辅助RL），但一旦研究超出证明，通常难以将这些更富有，更复杂的环境 - 概念结果。我们展示了一个强大的沙箱框架，用于易于设计新颖的RL环境。 Minihack是一个停止商店，用于RL实验，环境包括从小房间到复杂的，程序生成的世界。通过利用来自Nethack的全套实体和环境动态，MiniHack是最富有的基网上的视频游戏之一，允许设计快速方便的定制RL测试台。使用这种沙箱框架，可以轻松设计新颖的环境，可以使用人类可读的描述语言或简单的Python接口来设计。除了各种RL任务和基线外，Minihack还可以包装现有的RL基准，并提供无缝添加额外复杂性的方法。

Recent progress in pretraining language models on large textual corpora led to a surge of improvements for downstream NLP tasks. Whilst learning linguistic knowledge, these models may also be storing relational knowledge present in the training data, and may be able to answer queries structured as "fillin-the-blank" cloze statements. Language models have many advantages over structured knowledge bases: they require no schema engineering, allow practitioners to query about an open class of relations, are easy to extend to more data, and require no human supervision to train. We present an in-depth analysis of the relational knowledge already present (without fine-tuning) in a wide range of state-of-theart pretrained language models. We find that (i) without fine-tuning, BERT contains relational knowledge competitive with traditional NLP methods that have some access to oracle knowledge, (ii) BERT also does remarkably well on open-domain question answering against a supervised baseline, and (iii) certain types of factual knowledge are learned much more readily than others by standard language model pretraining approaches. The surprisingly strong ability of these models to recall factual knowledge without any fine-tuning demonstrates their potential as unsupervised open-domain QA systems. The code to reproduce our analysis is available at https: //github.com/facebookresearch/LAMA.

Recent years have seen a proliferation of research on adversarial machine learning. Numerous papers demonstrate powerful algorithmic attacks against a wide variety of machine learning (ML) models, and numerous other papers propose defenses that can withstand most attacks. However, abundant real-world evidence suggests that actual attackers use simple tactics to subvert ML-driven systems, and as a result security practitioners have not prioritized adversarial ML defenses. Motivated by the apparent gap between researchers and practitioners, this position paper aims to bridge the two domains. We first present three real-world case studies from which we can glean practical insights unknown or neglected in research. Next we analyze all adversarial ML papers recently published in top security conferences, highlighting positive trends and blind spots. Finally, we state positions on precise and cost-driven threat modeling, collaboration between industry and academia, and reproducible research. We believe that our positions, if adopted, will increase the real-world impact of future endeavours in adversarial ML, bringing both researchers and practitioners closer to their shared goal of improving the security of ML systems.

When simulating soft robots, both their morphology and their controllers play important roles in task performance. This paper introduces a new method to co-evolve these two components in the same process. We do that by using the hyperNEAT algorithm to generate two separate neural networks in one pass, one responsible for the design of the robot body structure and the other for the control of the robot. The key difference between our method and most existing approaches is that it does not treat the development of the morphology and the controller as separate processes. Similar to nature, our method derives both the "brain" and the "body" of an agent from a single genome and develops them together. While our approach is more realistic and doesn't require an arbitrary separation of processes during evolution, it also makes the problem more complex because the search space for this single genome becomes larger and any mutation to the genome affects "brain" and the "body" at the same time. Additionally, we present a new speciation function that takes into consideration both the genotypic distance, as is the standard for NEAT, and the similarity between robot bodies. By using this function, agents with very different bodies are more likely to be in different species, this allows robots with different morphologies to have more specialized controllers since they won't crossover with other robots that are too different from them. We evaluate the presented methods on four tasks and observe that even if the search space was larger, having a single genome makes the evolution process converge faster when compared to having separated genomes for body and control. The agents in our population also show morphologies with a high degree of regularity and controllers capable of coordinating the voxels to produce the necessary movements.