Large language models can store extensive world knowledge, often extractable
through question-answering (e.g., "What is Abraham Lincoln's birthday?").
However, it's unclear whether the model answers questions based on exposure to
exact/similar questions during training, or if it genuinely extracts knowledge
from the source (e.g., Wikipedia biographies).
In this paper, we conduct an in-depth study of this problem using a
controlled set of semi-synthetic biography data. We uncover a relationship
between the model's knowledge extraction ability and different diversity
measures of the training data. We conduct (nearly) linear probing, revealing a
strong correlation between this relationship and whether the model (nearly)
linearly encodes the knowledge attributes at the hidden embedding of the entity
names, or across the embeddings of other tokens in the training text.

通过对半合成的传记数据进行深入研究，我们发现大语言模型的知识提取能力与训练数据的多样性之间存在着关联，同时通过 (几乎) 线性推测发现模型是否线性地编码实体名称的隐藏嵌入中的知识属性，或者在训练文本中其他标记的嵌入之间存在着强相关性。

语言模型的物理学：3.1 部分，知识存储与提取

Physics of Language Models: Part 3.1, Knowledge Storage and Extraction

Abstractive summarization systems based on pretrained language models often
generate coherent but factually inconsistent sentences. In this paper, we
present a counterfactual data augmentation approach where we augment data with
perturbed summaries that increase the training data diversity. Specifically, we
present three augmentation approaches based on replacing (i) entities from
other and the same category and (ii) nouns with their corresponding WordNet
hypernyms. We show that augmenting the training data with our approach improves
the factual correctness of summaries without significantly affecting the ROUGE
score. We show that in two commonly used summarization datasets (CNN/Dailymail
and XSum), we improve the factual correctness by about 2.5 points on average

该文章介绍了一种基于对抗数据增广的方法，用于提高 Abstractive summarization system 的 factual correctness, 相关方法包括将同一或不同类别的实体或名词的 WordNet hypernyms 进行替换，实验证明，使用我们的方法对 CNN/Dailymail 和 XSum 等数据集进行训练，可将 factual correctness 平均提高约 2.5 个点。

反事实数据增强改进抽象摘要的真实性

Counterfactual Data Augmentation improves Factuality of Abstractive Summarization

Ultrasound b-mode imaging is a qualitative approach and diagnostic quality
strongly depends on operators' training and experience. Quantitative approaches
can provide information about tissue properties; therefore, can be used for
identifying various tissue types, e.g., speed-of-sound in the tissue can be
used as a biomarker for tissue malignancy, especially in breast imaging. Recent
studies showed the possibility of speed-of-sound reconstruction using deep
neural networks that are fully trained on simulated data. However, because of
the ever present domain shift between simulated and measured data, the
stability and performance of these models in real setups are still under
debate. In this study, we investigated the impacts of training data diversity
on the robustness of these networks by using multiple kinds of geometrical and
natural simulated phantom structures. On the simulated data, we investigated
the performance of the networks on out-of-domain echogenicity, geometries, and
in the presence of noise. We further inspected the stability of employing such
tissue modeling in a real data acquisition setup. We demonstrated that training
the network with a joint set of datasets including both geometrical and natural
tissue models improves the stability of the predicted speed-of-sound values
both on simulated and measured data.

本文研究了使用多种几何和自然的模拟幻影结构对深度神经网络训练数据多样性的影响，并证明在使用同时包含几何和自然组织模型的数据集来训练网络时，可以提高模型在模拟和实测数据中预测声速的稳定性。