The emergence of large pretrained models has enabled language models to achieve superior performance in common NLP tasks, including language modeling and question answering, compared to previous static word representation methods. Augmenting these models with a retriever to retrieve the related text and documents as supporting information has shown promise in effectively solving NLP problems in a more interpretable way given that the additional knowledge is injected explicitly rather than being captured in the models' parameters. In spite of the recent progress, our analysis on retriever-augmented language models shows that this class of language models still lack reasoning over the retrieved documents. In this paper, we study the strengths and weaknesses of different retriever-augmented language models such as REALM, kNN-LM, FiD, ATLAS, and Flan-T5 in reasoning over the selected documents in different tasks. In particular, we analyze the reasoning failures of each of these models and study how the models' failures in reasoning are rooted in the retriever module as well as the language model.

Syntax is a latent hierarchical structure which underpins the robust and compositional nature of human language. An active line of inquiry is whether large pretrained language models (LLMs) are able to acquire syntax by training on text alone; understanding a model's syntactic capabilities is essential to understanding how it processes and makes use of language. In this paper, we propose a new method, SSUD, which allows for the induction of syntactic structures without supervision from gold-standard parses. Instead, we seek to define formalism-agnostic, model-intrinsic syntactic parses by using a property of syntactic relations: syntactic substitutability. We demonstrate both quantitative and qualitative gains on dependency parsing tasks using SSUD, and induce syntactic structures which we hope provide clarity into LLMs and linguistic representations, alike.

语言模型既展示了定量的改进，又展示了新的定性功能，随着规模的增加。尽管它们具有潜在的变革性影响，但这些新能力的特征却很差。为了为未来的研究提供信息，为破坏性的新模型能力做准备，并改善社会有害的效果，至关重要的是，我们必须了解目前和近乎未来的能力和语言模型的局限性。为了应对这一挑战，我们介绍了超越模仿游戏基准（Big Bench）。 Big Bench目前由204个任务组成，由132家机构的442位作者贡献。任务主题是多样的，从语言学，儿童发展，数学，常识性推理，生物学，物理学，社会偏见，软件开发等等。 Big-Bench专注于被认为超出当前语言模型的功能的任务。我们评估了OpenAI的GPT型号，Google内部密集变压器体系结构和大型基础上的开关稀疏变压器的行为，跨越了数百万到数十亿个参数。此外，一个人类专家评估者团队执行了所有任务，以提供强大的基准。研究结果包括：模型性能和校准都随规模改善，但绝对的术语（以及与评估者的性能相比）；在模型类中的性能非常相似，尽管带有稀疏性。逐渐和预测的任务通常涉及大量知识或记忆成分，而在临界规模上表现出“突破性”行为的任务通常涉及多个步骤或组成部分或脆性指标；社交偏见通常会随着含糊不清的环境而随着规模而增加，但这可以通过提示来改善。

在智能辅导系统中生成提示的现有工作（ITS）主要集中在手动和非个人反馈上。在这项工作中，我们探索了ITS中的个性化反馈作为个性化反馈。我们的个性化反馈可以在学生答案中查明正确，错误或缺失的短语，并通过提出自然语言问题来指导他们正确答案。我们的方法结合了因果分析，以使用基于文本相似性的NLP变压器模型来分解学生答案，以识别正确和不正确或缺失的零件。我们培训了一些弹药的神经问题生成和问题重新排序模型，以显示解决学生答案中缺少的组件的问题，这些组件使学生朝着正确的答案迈进。在基于真实对话的ITS测试时，我们的模型在学生学习的增长方面大大优于简单和强大的基线。最后，我们表明我们个性化的纠正反馈系统有可能改善生成的问答系统。

可靠的评估基准是为了可复制性和全面性而设计的，在机器学习方面取得了进步。但是，由于缺乏多语言基准，视觉和语言研究主要集中在英语任务上。为了填补这一空白，我们介绍了图像的语言理解评估基准。 Iglue通过汇总已有的数据集并创建新的数据来汇集 - 视觉问题回答，跨模式检索，扎根的推理以及跨20种不同语言的扎根成本。我们的基准测试能够评估多语言多模型用于转移学习的模型，不仅在零弹位设置中，而且还以新定义的少数图学习设置。根据对可用最新模型的评估，我们发现翻译测试转移优于零弹性转移，并且对于许多任务而言，很难利用射击的学习。此外，下游性能部分用可用的未标记文本数据进行预处理来解释，并且仅通过目标源语言的类型学距离而微弱。我们希望通过向社区释放基准来鼓励该领域的未来研究工作。

预先训练的语言模型（LMS）通常逻辑地扭转或以组成方式概括。最近的工作表明，结合外部实体知识可以提高LMS的能力和推广。然而，明确提供实体抽象的效果仍然不清楚，特别是在最近的研究表明，预先训练的LMS已经在其参数中编码了一些知识。我们研究将实体型抽象的实用程序融入预先训练的变压器，并在需要不同形式的逻辑推理的四个NLP任务上测试这些方法：（1）与基于文本的关系推理（CLUTRR）的组成语言理解，（2）绑架推理（校对者），（3）多跳问题应答（HotpotQA），和（4）会话问题应答（COQA）。我们提出并经验探索了三种方法来添加此类抽象：（i）作为附加输入嵌入式，（ii）作为编码的单独序列，（iii）作为模型的辅助预测任务。总体而言，我们的分析表明，具有抽象实体知识的模型比没有它更好。然而，我们的实验还表明，强烈的益处取决于所使用的技术和手头的任务。与基线模型相比，最佳抽象意识模型分别达到了88.8％和91.8％的总精度，分别在CLUTRR和校对者上实现了62.3％和89.8％。此外，抽象感知模型在插值和外推设置中显示出改善的组成概括。然而，对于热杆菌和CoQA，我们发现F1分数平均仅提高0.5％。我们的结果表明，明确抽象的好处在正式定义的逻辑推理设置中需要许多推理跳跃，但指向它对具有较少正式逻辑结构的NLP任务不利的概念。

我们提出了一种用于在生成答案时将信息与多个检索文件中的信息组合的可检索增强的开放式开放式开放式开放域问题训练方法。我们将检索决策模拟作为相关文件集的潜在变量。由于通过对所检索的文件集的边缘化，因此使用期望最大化算法估计这一点。我们迭代地估计我们的潜在变量的价值（给定问题的这些相关文档集），然后使用此估计来更新检索器和读取器参数。我们假设这种端到端的训练允许训练信号流到读者，然后比上演明智的训练更好地流到猎犬。这导致检索器能够为问题和读者选择更多相关文档，这些文件在更准确的文档中培训以生成答案。三个基准数据集的实验表明，我们所提出的方法优于所有现有的相当大小的方法2-3％绝对精确匹配点，实现了新的最先进的结果。我们的结果还展示了学习检索以改善答复的可行性，而无明确监督检索决策。

Humans gather information through conversations involving a series of interconnected questions and answers. For machines to assist in information gathering, it is therefore essential to enable them to answer conversational questions. We introduce CoQA, a novel dataset for building Conversational Question Answering systems. 1 Our dataset contains 127k questions with answers, obtained from 8k conversations about text passages from seven diverse domains. The questions are conversational, and the answers are free-form text with their corresponding evidence highlighted in the passage. We analyze CoQA in depth and show that conversational questions have challenging phenomena not present in existing reading comprehension datasets, e.g., coreference and pragmatic reasoning. We evaluate strong dialogue and reading comprehension models on CoQA. The best system obtains an F1 score of 65.4%, which is 23.4 points behind human performance (88.8%), indicating there is ample room for improvement. We present CoQA as a challenge to the community at https://stanfordnlp. github.io/coqa.

The decarbonization of buildings presents new challenges for the reliability of the electrical grid as a result of the intermittency of renewable energy sources and increase in grid load brought about by end-use electrification. To restore reliability, grid-interactive efficient buildings can provide flexibility services to the grid through demand response. Residential demand response programs are hindered by the need for manual intervention by customers. To maximize the energy flexibility potential of residential buildings, an advanced control architecture is needed. Reinforcement learning is well-suited for the control of flexible resources as it is able to adapt to unique building characteristics compared to expert systems. Yet, factors hindering the adoption of RL in real-world applications include its large data requirements for training, control security and generalizability. Here we address these challenges by proposing the MERLIN framework and using a digital twin of a real-world 17-building grid-interactive residential community in CityLearn. We show that 1) independent RL-controllers for batteries improve building and district level KPIs compared to a reference RBC by tailoring their policies to individual buildings, 2) despite unique occupant behaviours, transferring the RL policy of any one of the buildings to other buildings provides comparable performance while reducing the cost of training, 3) training RL-controllers on limited temporal data that does not capture full seasonality in occupant behaviour has little effect on performance. Although, the zero-net-energy (ZNE) condition of the buildings could be maintained or worsened as a result of controlled batteries, KPIs that are typically improved by ZNE condition (electricity price and carbon emissions) are further improved when the batteries are managed by an advanced controller.

The current reinforcement learning algorithm uses forward-generated trajectories to train the agent. The forward-generated trajectories give the agent little guidance, so the agent can explore as much as possible. While the appreciation of reinforcement learning comes from enough exploration, this gives the trade-off of losing sample efficiency. The sampling efficiency is an important factor that decides the performance of the algorithm. Past tasks use reward shaping techniques and changing the structure of the network to increase sample efficiency, however these methods require many steps to implement. In this work, we propose novel reverse curriculum reinforcement learning. Reverse curriculum learning starts training the agent using the backward trajectory of the episode rather than the original forward trajectory. This gives the agent a strong reward signal, so the agent can learn in a more sample-efficient manner. Moreover, our method only requires a minor change in algorithm, which is reversing the order of trajectory before training the agent. Therefore, it can be simply applied to any state-of-art algorithms.