ABSTRACT We demonstrate the potential of the state-of-the-art OpenAI GPT-4 large language model to engage in meaningful interactions with Astronomy papers using in-context prompting. To optimize for efficiency, we employ a distillation technique that effectively reduces the size of the original input paper by 50%, while maintaining the paragraph structure and overall semantic integrity. We then explore the model's responses using a multi-document context (ten distilled documents). Our findings indicate that GPT-4 excels in the multi-document domain, providing detailed answers contextualized within the framework of related research findings. INTRODUCTION Large language models (LLMs) have significantly advanced natural language processing, allowing machines to process and generate intricate text with remarkable quality (e.g., Devlin et al. 2018; Brown et al. 2020; Chowdhery et al. 2022; Bubeck et al. 2023).

Studying data memorization in neural language models helps us understand the risks (e.g., to privacy or copyright) associated with models regurgitating training data and aids in the development of countermeasures. Many prior works -- and some recently deployed defenses -- focus on "verbatim memorization", defined as a model generation that exactly matches a substring from the training set. We argue that verbatim memorization definitions are too restrictive and fail to capture more subtle forms of memorization. Specifically, we design and implement an efficient defense that perfectly prevents all verbatim memorization. And yet, we demonstrate that this "perfect" filter does not prevent the leakage of training data. Indeed, it is easily circumvented by plausible and minimally modified "style-transfer" prompts -- and in some cases even the non-modified original prompts -- to extract memorized information. We conclude by discussing potential alternative definitions and why defining memorization is a difficult yet crucial open question for neural language models.

Although demonstrating superb performance on various NLP tasks, large language models (LLMs) still suffer from the hallucination problem, which threatens the reliability of LLMs. To measure the level of hallucination of LLMs, previous works first categorize the hallucination according to the phenomenon similarity, then quantify the proportion that model outputs contain hallucinatory contents. However, such hallucination rates could easily be distorted by confounders. Moreover, such hallucination rates could not reflect the reasons for the hallucination, as similar hallucinatory phenomena may originate from different sources. To address these issues, we propose to combine the hallucination level quantification and hallucination reason investigation through an association analysis, which builds the relationship between the hallucination rate of LLMs with a set of risk factors. In this way, we are able to observe the hallucination level under each value of each risk factor, examining the contribution and statistical significance of each risk factor, meanwhile excluding the confounding effect of other factors. Additionally, by recognizing the risk factors according to a taxonomy of model capability, we reveal a set of potential deficiencies in commonsense memorization, relational reasoning, and instruction following, which may further provide guidance for the pretraining and supervised fine-tuning process of LLMs to mitigate the hallucination.

Large language models (LLMs) have led to a surge in collaborative writing with model assistance. As different users incorporate suggestions from the same model, there is a risk of decreased diversity in the produced content, potentially limiting diverse perspectives in public discourse. In this work, we measure the impact of co-writing on diversity via a controlled experiment, where users write argumentative essays in three setups -- using a base LLM (GPT3), a feedback-tuned LLM (InstructGPT), and writing without model help. We develop a set of diversity metrics and find that writing with InstructGPT (but not the GPT3) results in a statistically significant reduction in diversity. Specifically, it increases the similarity between the writings of different authors and reduces the overall lexical and content diversity. We additionally find that this effect is mainly attributable to InstructGPT contributing less diverse text to co-written essays. In contrast, the user-contributed text remains unaffected by model collaboration. This suggests that the recent improvement in generation quality from adapting models to human feedback might come at the cost of more homogeneous and less diverse content.

We use large language models to aid learners enhance proficiency in a foreign language. This is accomplished by identifying content on topics that the user is interested in, and that closely align with the learner's proficiency level in that foreign language. Our work centers on French content, but our approach is readily transferable to other languages. Our solution offers several distinctive characteristics that differentiate it from existing language-learning solutions, such as, a) the discovery of content across topics that the learner cares about, thus increasing motivation, b) a more precise estimation of the linguistic difficulty of the content than traditional readability measures, and c) the availability of both textual and video-based content. The linguistic complexity of video content is derived from the video captions. It is our aspiration that such technology will enable learners to remain engaged in the language-learning process by continuously adapting the topics and the difficulty of the content to align with the learners' evolving interests and learning objectives. A video showcasing our solution can be found at: https://youtu.be/O6krGN-LTGI

The development of believable, natural, and interactive digital artificial agents is a field of growing interest. Theoretical uncertainties and technical barriers present considerable challenges to the field, particularly with regards to developing agents that effectively simulate human emotions. Large language models (LLMs) might address these issues by tapping common patterns in situational appraisal. In three empirical experiments, this study tests the capabilities of LLMs to solve emotional intelligence tasks and to simulate emotions. It presents and evaluates a new chain-of-emotion architecture for emotion simulation within video games, based on psychological appraisal research. Results show that it outperforms standard LLM architectures on a range of user experience and content analysis metrics. This study therefore provides early evidence of how to construct and test affective agents based on cognitive processes represented in language models.

Over decades, neuroscience has accumulated a wealth of research results in the text modality that can be used to explore cognitive processes. Meta-analysis is a typical method that successfully establishes a link from text queries to brain activation maps using these research results, but it still relies on an ideal query environment. In practical applications, text queries used for meta-analyses may encounter issues such as semantic redundancy and ambiguity, resulting in an inaccurate mapping to brain images. On the other hand, large language models (LLMs) like ChatGPT have shown great potential in tasks such as context understanding and reasoning, displaying a high degree of consistency with human natural language. Hence, LLMs could improve the connection between text modality and neuroscience, resolving existing challenges of meta-analyses. In this study, we propose a method called Chat2Brain that combines LLMs to basic text-2-image model, known as Text2Brain, to map open-ended semantic queries to brain activation maps in data-scarce and complex query environments. By utilizing the understanding and reasoning capabilities of LLMs, the performance of the mapping model is optimized by transferring text queries to semantic queries. We demonstrate that Chat2Brain can synthesize anatomically plausible neural activation patterns for more complex tasks of text queries.

Prompt-based classifiers are an attractive approach for zero-shot classification. However, the precise choice of the prompt template and label words can largely influence performance, with semantically equivalent settings often showing notable performance difference. This discrepancy can be partly attributed to word biases, where the classifier may be biased towards classes. To address this problem, it is possible to optimise classification thresholds on a labelled data set, however, this mitigates some of the advantages of prompt-based classifiers. This paper instead approaches this problem by examining the expected marginal probabilities of the classes. Here, probabilities are reweighted to have a uniform prior over classes, in an unsupervised fashion. Further, we draw a theoretical connection between the class priors and the language models' word prior, and offer the ability to set a threshold in a zero-resource fashion. We show that matching class priors correlates strongly with the oracle upper bound performance and demonstrate large consistent performance gains for prompt settings over a range of NLP tasks.

With the emergence of Large Language Models (LLMs), there has been a significant improvement in the programming capabilities of models, attracting growing attention from researchers. We propose CodeApex, a bilingual benchmark dataset focusing on the programming comprehension and code generation abilities of LLMs. CodeApex comprises three types of multiple-choice questions: conceptual understanding, commonsense reasoning, and multi-hop reasoning, designed to evaluate LLMs on programming comprehension tasks. Additionally, CodeApex utilizes algorithmic questions and corresponding test cases to assess the code quality generated by LLMs. We evaluate 14 state-of-the-art LLMs, including both general-purpose and specialized models. GPT exhibits the best programming capabilities, achieving approximate accuracies of 50% and 56% on the two tasks, respectively. There is still significant room for improvement in programming tasks. We hope that CodeApex can serve as a reference for evaluating the coding capabilities of LLMs, further promoting their development and growth. Datasets are released at https://github.com/APEXLAB/CodeApex.git. CodeApex submission website is https://apex.sjtu.edu.cn/codeapex/.

Knowledge graphs play a vital role in numerous artificial intelligence tasks, yet they frequently face the issue of incompleteness. In this study, we explore utilizing Large Language Models (LLM) for knowledge graph completion. We consider triples in knowledge graphs as text sequences and introduce an innovative framework called Knowledge Graph LLM (KG-LLM) to model these triples. Our technique employs entity and relation descriptions of a triple as prompts and utilizes the response for predictions. Experiments on various benchmark knowledge graphs demonstrate that our method attains state-of-the-art performance in tasks such as triple classification and relation prediction. We also find that fine-tuning relatively smaller models (e.g., LLaMA-7B, ChatGLM-6B) outperforms recent ChatGPT and GPT-4.