Annotators are not fungible. Their demographics, life experiences, and backgrounds all contribute to how they label data. However, NLP has only recently considered how annotator identity might influence their decisions. Here, we present POPQUORN (the POtato-Prolific dataset for QUestion-Answering, Offensiveness, text Rewriting, and politeness rating with demographic Nuance). POPQUORN contains 45,000 annotations from 1,484 annotators, drawn from a representative sample regarding sex, age, and race as the US population. Through a series of analyses, we show that annotators' background plays a significant role in their judgments. Further, our work shows that backgrounds not previously considered in NLP (e.g., education), are meaningful and should be considered. Our study suggests that understanding the background of annotators and collecting labels from a demographically balanced pool of crowd workers is important to reduce the bias of datasets. The dataset, annotator background, and annotation interface are available at https://github.com/Jiaxin-Pei/potato-prolific-dataset .

AI powered code-recommendation systems, such as Copilot and CodeWhisperer, provide code suggestions inside a programmer's environment (e.g., an IDE) with the aim to improve their productivity. Since, in these scenarios, programmers accept and reject suggestions, ideally, such a system should use this feedback in furtherance of this goal. In this work, we leverage prior data of programmers interacting with GitHub Copilot, a system used by millions of programmers, to develop interventions that can save programmer time. We propose a utility theory framework, which models this interaction with programmers and decides which suggestions to display. Our framework Conditional suggestion Display from Human Feedback (CDHF), relies on a cascade of models that predict suggestion acceptance to selectively hide suggestions reducing both latency and programmer verification time. Using data from 535 programmers, we perform a retrospective evaluation of CDHF and show that we can avoid displaying a significant fraction of suggestions that would have been rejected doing so without total knowledge of the suggestions themselves. We further demonstrate the importance of incorporating the programmer's latent unobserved state in deciding when to display suggestions through ablations on user study data. Finally, we showcase that using suggestion acceptance as a reward signal to know which suggestions to display leads to reduced quality suggestions indicating an unexpected pitfall.

This progress comes at a cost, however, of increased resource consumption and latency during both training and inference, presenting challenges not only in real-world deployment but also in terms of environmental impact and energy usage (Sharir et al., 2020; Patterson et al., 2021; Bommasani et al., 2022). For instance, LLM-based chatbots typically consist of large transformer-based networks with parameter counts ranging from one to several hundred billion (Zhou et al., 2023). Moreover, the auto-regressive nature of LLMs exacerbates the issue of latency and resource consumption because the model can only generate one token at a time. Thus, compared to traditional AI-powered services, language model inference costs are much higher and the latency is significantly longer, making it nearly impossible to process each query using LLMs in high-throughput query systems such as search engines. In this paper, we explore two simple yet effective strategies to mitigate this problem: (1) employing a caching system to store previous queries, and (2) developing a model multiplexer to choose the most appropriate model from a set of models for processing the queries. The general workflow of our proposed LLM-based inference system is shown in Figure 1: upon receiving a query or prompt, we initially check if it can be retrieved from the cache. If the query is not found in the cache, we employ the model multiplexer to determine which model should be used for processing it first, based on the estimated cost for both models. The choice of cost function and models can vary based on the goal. One measure of cost, for example, could be floating point operations (FLOPs).

Transformers have emerged as the cornerstone of state-of-the-art natural language processing models, showcasing exceptional performance across a wide range of AI applications. However, the memory demands posed by the self-attention mechanism and the large feedforward network in Transformers limit their ability to handle long sequences, thereby creating challenges for tasks involving multiple long sequences or long-term dependencies. We present a distinct approach, Blockwise Parallel Transformer (BPT), that leverages blockwise computation of self-attention and feedforward network fusion to minimize memory costs. By processing longer input sequences while maintaining memory efficiency, BPT enables training sequences 32 times longer than vanilla Transformers and up to 4 times longer than previous memory-efficient methods. Extensive experiments on language modeling and reinforcement learning tasks demonstrate the effectiveness of BPT in reducing memory requirements and improving performance.

Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose Scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that Scissorhands reduces the inference memory usage of the KV cache by up to 5X without compromising model quality. We further demonstrate that Scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20X compression.

Large Language Models (LLMs) have gathered significant attention due to their impressive performance on a variety of tasks. ChatGPT, developed by OpenAI, is a recent addition to the family of language models and is being called a disruptive technology by a few, owing to its human-like text-generation capabilities. Although, many anecdotal examples across the internet have evaluated ChatGPT's strength and weakness, only a few systematic research studies exist. To contribute to the body of literature of systematic research on ChatGPT, we evaluate the performance of ChatGPT on Abstractive Summarization by the means of automated metrics and blinded human reviewers. We also build automatic text classifiers to detect ChatGPT generated summaries. We found that while text classification algorithms can distinguish between real and generated summaries, humans are unable to distinguish between real summaries and those produced by ChatGPT.

The language ability of Large Language Models (LLMs) is often unbalanced towards English because of the imbalance in the distribution of the pre-training data. This disparity is demanded in further fine-tuning and affecting the cross-lingual abilities of LLMs. In this paper, we propose to empower Instructiontuned LLMs (It-LLMs) in languages other than English by building semantic alignment between them. Hence, we propose CrossAlpaca, an It-LLM with cross-lingual instruction-following and Translation-following demonstrations to improve semantic alignment between languages. We validate our approach on the multilingual Question Answering (QA) benchmarks XQUAD and MLQA and adapted versions of MMLU and BBH. Our models, tested over six different languages, outperform the It-LLMs tuned on monolingual data. The final results show that instruction tuning on non-English data is not enough and that semantic alignment can be further improved by Translation-following demonstrations.

Large Language Models (LLMs) have exploded a new heatwave of AI for their ability to engage end-users in human-level conversations with detailed and articulate answers across many knowledge domains. In response to their fast adoption in many industrial applications, this survey concerns their safety and trustworthiness. First, we review known vulnerabilities and limitations of the LLMs, categorising them into inherent issues, attacks, and unintended bugs. Then, we consider if and how the Verification and Validation (V&V) techniques, which have been widely developed for traditional software and deep learning models such as convolutional neural networks as independent processes to check the alignment of their implementations against the specifications, can be integrated and further extended throughout the lifecycle of the LLMs to provide rigorous analysis to the safety and trustworthiness of LLMs and their applications. Specifically, we consider four complementary techniques: falsification and evaluation, verification, runtime monitoring, and regulations and ethical use. In total, 370+ references are considered to support the quick understanding of the safety and trustworthiness issues from the perspective of V&V. While intensive research has been conducted to identify the safety and trustworthiness issues, rigorous yet practical methods are called for to ensure the alignment of LLMs with safety and trustworthiness requirements.

Recent advances in Language Models have enabled the protein modeling community with a powerful tool since protein sequences can be represented as text. Specifically, by taking advantage of Transformers, sequence-to-property prediction will be amenable without the need for explicit structural data. In this work, inspired by recent progress in Large Language Models (LLMs), we introduce PeptideBERT, a protein language model for predicting three key properties of peptides (hemolysis, solubility, and non-fouling). The PeptideBert utilizes the ProtBERT pretrained transformer model with 12 attention heads and 12 hidden layers. We then finetuned the pretrained model for the three downstream tasks. Our model has achieved state of the art (SOTA) for predicting Hemolysis, which is a task for determining peptide's potential to induce red blood cell lysis. Our PeptideBert non-fouling model also achieved remarkable accuracy in predicting peptide's capacity to resist non-specific interactions. This model, trained predominantly on shorter sequences, benefits from the dataset where negative examples are largely associated with insoluble peptides. Codes, models, and data used in this study are freely available at: https://github.com/ChakradharG/PeptideBERT

Natural language is among the most accessible tools for explaining decisions to humans, and large pretrained language models (PLMs) have demonstrated impressive abilities to generate coherent natural language explanations (NLE). The existing NLE research perspectives do not take the audience into account. An NLE can have high textual quality, but it might not accommodate audiences' needs and preference. To address this limitation, we propose an alternative perspective, situated NLE, including a situated generation framework and a situated evaluation framework. On the generation side, we propose simple prompt engineering methods that adapt the NLEs to situations. In human studies, the annotators preferred the situated NLEs. On the evaluation side, we set up automated evaluation scores in lexical, semantic, and pragmatic categories. The scores can be used to select the most suitable prompts to generate NLEs. Situated NLE provides a perspective to conduct further research on automatic NLE generations.