Large language models (LLMs) have recently become remarkably good at improving developer productivity for high-resource programming languages. These models use two kinds of data: large amounts of unlabeled code samples for pretraining and relatively smaller amounts of labeled code samples for fine-tuning or in-context learning. Unfortunately, many programming languages are low-resource, lacking labeled samples for most tasks and often even lacking unlabeled samples. Therefore, users of low-resource languages (e.g., legacy or new languages) miss out on the benefits of LLMs. Cross-lingual transfer learning uses data from a source language to improve model performance on a target language. It has been well-studied for natural languages, but has received little attention for programming languages. This paper reports extensive experiments on four tasks using a transformer-based LLM and 11 to 41 programming languages to explore the following questions. First, how well cross-lingual transfer works for a given task across different language pairs. Second, given a task and target language, how to best choose a source language. Third, the characteristics of a language pair that are predictive of transfer performance, and fourth, how that depends on the given task.

The growing capabilities of transformer models pave the way for solving increasingly complex NLP tasks. A key to supporting application-specific requirements is the ability to fine-tune. However, compiling a fine-tuning dataset tailored to complex tasks is tedious and results in large datasets, limiting the ability to control transformer output. We present an approach in which complex tasks are divided into simpler subtasks. Multiple transformer models are fine-tuned to one subtask each, and lined up to accomplish the complex task. This simplifies the compilation of fine-tuning datasets and increases overall controllability. Using the example of reducing gender bias as a complex task, we demonstrate our approach and show that it performs better than using a single model.

The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by GLM, we propose a General Point Model (GPM) which seamlessly integrates autoencoding and autoregressive tasks in point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks.

We propose LLM-FP4 for quantizing both weights and activations in large language models (LLMs) down to 4-bit floating-point values, in a post-training manner. Existing post-training quantization (PTQ) solutions are primarily integer-based and struggle with bit widths below 8 bits. Compared to integer quantization, floating-point (FP) quantization is more flexible and can better handle long-tail or bell-shaped distributions, and it has emerged as a default choice in many hardware platforms. One characteristic of FP quantization is that its performance largely depends on the choice of exponent bits and clipping range. In this regard, we construct a strong FP-PTQ baseline by searching for the optimal quantization parameters. Furthermore, we observe a high inter-channel variance and low intra-channel variance pattern in activation distributions, which adds activation quantization difficulty. We recognize this pattern to be consistent across a spectrum of transformer models designed for diverse tasks, such as LLMs, BERT, and Vision Transformer models. To tackle this, we propose per-channel activation quantization and show that these additional scaling factors can be reparameterized as exponential biases of weights, incurring a negligible cost. Our method, for the first time, can quantize both weights and activations in the LLaMA-13B to only 4-bit and achieves an average score of 63.1 on the common sense zero-shot reasoning tasks, which is only 5.8 lower than the full-precision model, significantly outperforming the previous state-of-the-art by 12.7 points. Code is available at: https://github.com/nbasyl/LLM-FP4.

This study explores the capabilities of prompt-driven Large Language Models (LLMs) like ChatGPT and GPT-4 in adhering to human guidelines for dialogue summarization. Experiments employed DialogSum (English social conversations) and DECODA (French call center interactions), testing various prompts: including prompts from existing literature and those from human summarization guidelines, as well as a two-step prompt approach. Our findings indicate that GPT models often produce lengthy summaries and deviate from human summarization guidelines. However, using human guidelines as an intermediate step shows promise, outperforming direct word-length constraint prompts in some cases. The results reveal that GPT models exhibit unique stylistic tendencies in their summaries. While BERTScores did not dramatically decrease for GPT outputs suggesting semantic similarity to human references and specialised pre-trained models, ROUGE scores reveal grammatical and lexical disparities between GPT-generated and human-written summaries. These findings shed light on the capabilities and limitations of GPT models in following human instructions for dialogue summarization.

Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing, bringing faster and more accurate models, at the cost of massive parameter counts. For example, the SwitchTransformer-c2048 model has 1.6 trillion parameters, requiring 3.2TB of accelerator memory to run efficiently, which makes practical deployment challenging and expensive. In this paper, we present a solution to this memory problem, in form of a new compression and execution framework called QMoE. Specifically, QMoE consists of a scalable algorithm which accurately compresses trillion-parameter MoEs to less than 1 bit per parameter, in a custom format co-designed with bespoke GPU decoding kernels to facilitate efficient end-to-end compressed inference, with minor runtime overheads relative to uncompressed execution. Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB (20x compression, 0.8 bits per parameter) at only minor accuracy loss, in less than a day on a single GPU. This enables, for the first time, the execution of a trillion-parameter model on affordable commodity hardware, like a single server with 4x NVIDIA A6000 or 8x NVIDIA 3090 GPUs, at less than 5% runtime overhead relative to ideal uncompressed inference. The source code and compressed models are available at github.com/IST-DASLab/qmoe.

Theory of Mind (ToM) is the ability to reason about one's own and others' mental states. ToM plays a critical role in the development of intelligence, language understanding, and cognitive processes. While previous work has primarily focused on first and second-order ToM, we explore higher-order ToM, which involves recursive reasoning on others' beliefs. We introduce HI-TOM, a Higher Order Theory of Mind benchmark. Our experimental evaluation using various Large Language Models (LLMs) indicates a decline in performance on higher-order ToM tasks, demonstrating the limitations of current LLMs. We conduct a thorough analysis of different failure cases of LLMs, and share our thoughts on the implications of our findings on the future of NLP.

Authorship Analysis, also known as stylometry, has been an essential aspect of Natural Language Processing (NLP) for a long time. Likewise, the recent advancement of Large Language Models (LLMs) has made authorship analysis increasingly crucial for distinguishing between human-written and AI-generated texts. However, these authorship analysis tasks have primarily been focused on written texts, not considering spoken texts. Thus, we introduce the largest benchmark for spoken texts - HANSEN (Human ANd ai Spoken tExt beNchmark). HANSEN encompasses meticulous curation of existing speech datasets accompanied by transcripts, alongside the creation of novel AI-generated spoken text datasets. Together, it comprises 17 human datasets, and AI-generated spoken texts created using 3 prominent LLMs: ChatGPT, PaLM2, and Vicuna13B. To evaluate and demonstrate the utility of HANSEN, we perform Authorship Attribution (AA) & Author Verification (AV) on human-spoken datasets and conducted Human vs. AI spoken text detection using state-of-the-art (SOTA) models. While SOTA methods, such as, character ngram or Transformer-based model, exhibit similar AA & AV performance in human-spoken datasets compared to written ones, there is much room for improvement in AI-generated spoken text detection. The HANSEN benchmark is available at: https://huggingface.co/datasets/HANSEN-REPO/HANSEN.

Large language models (LLMs) have become an integral component in solving a wide range of NLP tasks. In this work, we explore a novel use case of using LLMs to build performance predictors (PP): models that, given a specific deep neural network architecture, predict its performance on a downstream task. We design PP prompts for LLMs consisting of: (i) role: description of the role assigned to the LLM, (ii) instructions: set of instructions to be followed by the LLM to carry out performance prediction, (iii) hyperparameters: a definition of each architecture-specific hyperparameter and (iv) demonstrations: sample architectures along with their efficiency metrics and 'training from scratch' performance. For machine translation (MT) tasks, we discover that GPT-4 with our PP prompts (LLM-PP) can predict the performance of architecture with a mean absolute error matching the SOTA and a marginal degradation in rank correlation coefficient compared to SOTA performance predictors. Further, we show that the predictions from LLM-PP can be distilled to a small regression model (LLM-Distill-PP). LLM-Distill-PP models surprisingly retain the performance of LLM-PP largely and can be a cost-effective alternative for heavy use cases of performance estimation. Specifically, for neural architecture search (NAS), we propose a Hybrid-Search algorithm for NAS (HS-NAS), which uses LLM-Distill-PP for the initial part of search, resorting to the baseline predictor for rest of the search. We show that HS-NAS performs very similar to SOTA NAS across benchmarks, reduces search hours by 50% roughly, and in some cases, improves latency, GFLOPs, and model size.

Automating code documentation through explanatory text can prove highly beneficial in code understanding. Large Language Models (LLMs) have made remarkable strides in Natural Language Processing, especially within software engineering tasks such as code generation and code summarization. This study specifically delves into the task of generating natural-language summaries for code snippets, using various LLMs. The findings indicate that Code LLMs outperform their generic counterparts, and zero-shot methods yield superior results when dealing with datasets with dissimilar distributions between training and testing sets.