Perception of toxicity evolves over time and often differs between geographies and cultural backgrounds. Similarly, black-box commercially available APIs for detecting toxicity, such as the Perspective API, are not static, but frequently retrained to address any unattended weaknesses and biases. We evaluate the implications of these changes on the reproducibility of findings that compare the relative merits of models and methods that aim to curb toxicity. Our findings suggest that research that relied on inherited automatic toxicity scores to compare models and techniques may have resulted in inaccurate findings. Rescoring all models from HELM, a widely respected living benchmark, for toxicity with the recent version of the API led to a different ranking of widely used foundation models. We suggest caution in applying apples-to-apples comparisons between studies and lay recommendations for a more structured approach to evaluating toxicity over time. Code and data are available at https://github.com/for-ai/black-box-api-challenges.

Large language models (LLMs) have received significant attention by achieving remarkable performance across various tasks. However, their fixed context length poses challenges when processing long documents or maintaining extended conversations. This paper proposes a method called \textit{Selective Context} that employs self-information to filter out less informative content, thereby enhancing the efficiency of the fixed context length. We demonstrate the effectiveness of our approach on tasks of summarisation and question answering across different data sources, including academic papers, news articles, and conversation transcripts.

Pre-trained large-scale language models (LLMs) have shown revolutionary performance in many downstream tasks (Guo et al., 2023; Wei et al., 2021). One crucial ability of LLMs is called instruction following. That is, models can complete the tasks described by instructions given as input. This ability is based on a specialized training stage called instruction tuning. Compared to unlabeled data used for pre-training, the data for instruction tuning is typically more goal-oriented, and it should explicitly demonstrate how a response follows its corresponding instruction with a given input. There are many instruction tuning datasets in English. For example, the FLAN collection (Longpre et al., 2023) contains 15M examples covering 1836 tasks, and OPT-IML (Iyer et al., 2022b) claims to have 18M examples for more than 2000 tasks (although it is still not publicly available). In contrast, existing data resources for Chinese instruction tuning are either small in scale or have questionable quality. For example, Ziang Leng and Li (2023) directly translate English instruction tuning data into Chinese, but do not consider mitigating translation errors or potential cultural gaps, e.g.

Large language models (LLMs) are capable to perform complex reasoning by in-context learning (ICL) when provided with a few input-output demonstrations (demos) and more powerful when intermediate reasoning steps ("chain of thoughts (CoT)") of the demos are given. Is it necessary to use multi-demo in ICL? In this paper, we study ICL using fewer demos for each test query on the tasks in~\cite{wei2022chain}. Surprisingly, we do not observe significant degradation when using only one randomly chosen demo. To study this phenomenon, for each test query, we categorize demos into "correct demos" leading to the correct answer, and "wrong demos" resulting in wrong answers. Our analysis reveals an inherent bias in those widely studied datasets: most demos are correct for a majority of test queries, which explains the good performance of using one random demo. Moreover, ICL (with and w/o CoT) using only one correct demo significantly outperforms all-demo ICL adopted by most previous works, indicating the weakness of LLMs in finding correct demo(s) for input queries, which is difficult to evaluate on the biased datasets. Furthermore, we observe a counterintuitive behavior of ICL using multi-demo, i.e., its accuracy degrades(improves) when given more correct(wrong) demos. This implies that ICL can be easily misguided by interference among demos and their spurious correlations. Our analyses highlight several fundamental challenges that need to be addressed in LLMs training, ICL, and benchmark design.

The rise of large language models (LLMs) is revolutionizing information retrieval, question answering, summarization, and code generation tasks. However, in addition to confidently presenting factually inaccurate information at times (known as "hallucinations"), LLMs are also inherently limited by the number of input and output tokens that can be processed at once, making them potentially less effective on tasks that require processing a large set or continuous stream of information. A common approach to reducing the size of data is through lossless or lossy compression. Yet, in some cases it may not be strictly necessary to perfectly recover every detail from the original data, as long as a requisite level of semantic precision or intent is conveyed. This paper presents three contributions to research on LLMs. First, we present the results from experiments exploring the viability of approximate compression using LLMs, focusing specifically on GPT-3.5 and GPT-4 via ChatGPT interfaces. Second, we investigate and quantify the capability of LLMs to compress text and code, as well as to recall and manipulate compressed representations of prompts. Third, we present two novel metrics -- Exact Reconstructive Effectiveness (ERE) and Semantic Reconstruction Effectiveness (SRE) -- that quantify the level of preserved intent between text compressed and decompressed by the LLMs we studied. Our initial results indicate that GPT-4 can effectively compress and reconstruct text while preserving the semantic essence of the original text, providing a path to leverage $\sim$5$\times$ more tokens than present limits allow.

Human robot collaboration is becoming increasingly important as robots become more involved in various aspects of human life in the era of Artificial Intelligence. However, the issue of human operators trust in robots remains a significant concern, primarily due to the lack of adequate semantic understanding and communication between humans and robots. The emergence of Large Language Models (LLMs), such as ChatGPT, provides an opportunity to develop an interactive, communicative, and robust human-robot collaboration approach. This paper explores the impact of ChatGPT on trust in a human-robot collaboration assembly task. This study designs a robot control system called RoboGPT using ChatGPT to control a 7-degree-of-freedom robot arm to help human operators fetch, and place tools, while human operators can communicate with and control the robot arm using natural language. A human-subject experiment showed that incorporating ChatGPT in robots significantly increased trust in human-robot collaboration, which can be attributed to the robot's ability to communicate more effectively with humans. Furthermore, ChatGPT ability to understand the nuances of human language and respond appropriately helps to build a more natural and intuitive human-robot interaction. The findings of this study have significant implications for the development of human-robot collaboration systems.

Human label variation (Plank 2022), or annotation disagreement, exists in many natural language processing (NLP) tasks. To be robust and trusted, NLP models need to identify such variation and be able to explain it. To this end, we created the first ecologically valid explanation dataset with diverse reasoning, LiveNLI. LiveNLI contains annotators' highlights and free-text explanations for the label(s) of their choice for 122 English Natural Language Inference items, each with at least 10 annotations. We used its explanations for chain-of-thought prompting, and found there is still room for improvement in GPT-3's ability to predict label distribution with in-context learning.

The powerful ability of ChatGPT has caused widespread concern in the academic community. Malicious users could synthesize dummy academic content through ChatGPT, which is extremely harmful to academic rigor and originality. The need to develop ChatGPT-written content detection algorithms call for large-scale datasets. In this paper, we initially investigate the possible negative impact of ChatGPT on academia,and present a large-scale CHatGPT-writtEn AbsTract dataset (CHEAT) to support the development of detection algorithms. In particular, the ChatGPT-written abstract dataset contains 35,304 synthetic abstracts, with Generation, Polish, and Mix as prominent representatives. Based on these data, we perform a thorough analysis of the existing text synthesis detection algorithms. We show that ChatGPT-written abstracts are detectable, while the detection difficulty increases with human involvement.

We introduce compositional soft prompting (CSP), a parameter-efficient learning technique to improve the zero-shot compositionality of large-scale pretrained vision-language models (VLMs) like CLIP. We develop CSP for compositional zero-shot learning, the task of predicting unseen attribute-object compositions (e.g., old cat and young tiger). VLMs have a flexible text encoder that can represent arbitrary classes as natural language prompts but they often underperform task-specific architectures on the compositional zero-shot benchmark datasets. CSP treats the attributes and objects that define classes as learnable tokens of vocabulary. During training, the vocabulary is tuned to recognize classes that compose tokens in multiple ways (e.g., old cat and white cat). At test time, we recompose the learned attribute-object vocabulary in new combinations to recognize novel classes. We show that CSP outperforms the CLIP on benchmark datasets by an average of 10.9 percentage points on AUC. CSP also outperforms CoOp, a soft prompting method that fine-tunes the prefix context tokens, by an average of 5.8 percentage points on AUC. We perform additional experiments to show that CSP improves generalization to higher-order attribute-attribute-object compositions (e.g., old white cat) and combinations of pretrained attributes and fine-tuned objects. The code is available at https://github.com/BatsResearch/csp.

This paper describes the submission of UZH_CLyp for the SemEval 2023 Task 9 "Multilingual Tweet Intimacy Analysis". We achieved second-best results in all 10 languages according to the official Pearson's correlation regression evaluation measure. Our cross-lingual transfer learning approach explores the benefits of using a Head-First Fine-Tuning method (HeFiT) that first updates only the regression head parameters and then also updates the pre-trained transformer encoder parameters at a reduced learning rate. Additionally, we study the impact of using a small set of automatically generated examples (in our case, from ChatGPT) for low-resource settings where no human-labeled data is available. Our study shows that HeFiT stabilizes training and consistently improves results for pre-trained models that lack domain adaptation to tweets. Our study also shows a noticeable performance increase in cross-lingual learning when synthetic data is used, confirming the usefulness of current text generation systems to improve zero-shot baseline results. Finally, we examine how possible inconsistencies in the annotated data contribute to cross-lingual interference issues.