Search

Autoregressive language models are the currently dominant paradigm for text generation, but they have some fundamental limitations that cannot be remedied by scale-for example inherently sequential and unidirectional generation. While alternate classes of models have been explored, we have limited mathematical understanding of their fundamental power and limitations. In this paper we focus on Generative Masked Language Models (GMLMs), a non-autoregressive paradigm in which we train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model, These models empirically strike a promising speed-quality trade-off as each step can be typically parallelized by decoding the entire sequence in parallel. We develop a mathematical framework for analyzing and improving such models which sheds light on questions of sample complexity and inference speed and quality. Empirically, we adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality compared with autoregressive models. We run careful ablation experiments to give recommendations on key design choices, and make fine-grained observations on the common error modes in connection with our theory. Our mathematical analyses and empirical observations characterize both potentials and limitations of this approach, and can be applied to future works on improving understanding and performance of GMLMs. Our codes are released at https://github.com/google-research/google-research/tree/master/padir, Comment: ICML 2024

Pretrained language models (PLMs) display impressive performances and have captured the attention of the NLP community. Establishing best practices in pretraining has, therefore, become a major focus of NLP research, especially since insights gained from monolingual English models may not necessarily apply to more complex multilingual models. One significant caveat of the current state of the art is that different works are rarely comparable: they often discuss different parameter counts, training data, and evaluation methodology. This paper proposes a comparison of multilingual pretraining objectives in a controlled methodological environment. We ensure that training data and model architectures are comparable, and discuss the downstream performances across 6 languages that we observe in probing and fine-tuning scenarios. We make two key observations: (1) the architecture dictates which pretraining objective is optimal; (2) multilingual translation is a very effective pretraining objective under the right conditions. We make our code, data, and model weights available at \texttt{\url{https://github.com/Helsinki-NLP/lm-vs-mt}}., Comment: Proceedings of EMNLP 2024

Spoken dialogue plays a crucial role in human-AI interactions, necessitating dialogue-oriented spoken language models (SLMs). To develop versatile SLMs, large-scale and diverse speech datasets are essential. Additionally, to ensure hiqh-quality speech generation, the data must be spontaneous like in-wild data and must be acoustically clean with noise removed. Despite the critical need, no open-source corpus meeting all these criteria has been available. This study addresses this gap by constructing and releasing a large-scale spoken dialogue corpus, named Japanese Corpus for Human-AI Talks (J-CHAT), which is publicly accessible. Furthermore, this paper presents a language-independent method for corpus construction and describes experiments on dialogue generation using SLMs trained on J-CHAT. Experimental results indicate that the collected data from multiple domains by our method improve the naturalness and meaningfulness of dialogue generation., Comment: 8 pages, 6 figures

The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named $\text{Memory}^3$, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.

As the context length that large language models can handle continues to increase, these models demonstrate an enhanced ability to utilize distant information for tasks such as language modeling. This capability contrasts with human reading and writing habits, where it is uncommon to remember and use particularly distant information, except in cases of foreshadowing. In this paper, we aim to explore which kinds of words benefit more from long contexts in language models. By analyzing the changes in token probabilities with increasing context length, we find that content words (e.g., nouns, adjectives) and the initial tokens of words benefit the most. Frequent patterns in the context (N-grams) also significantly impact predictions. Additionally, the model's prior knowledge plays a crucial role in influencing predictions, especially for rare tokens. We also observe that language models become more confident with longer contexts, resulting in sharper probability distributions. This overconfidence may contribute to the increasing probabilities of tokens with distant contextual information. We hope that our analysis will help the community better understand long-text language modeling and contribute to the design of more reliable long-context models.

Towards energy-efficient artificial intelligence similar to the human brain, the bio-inspired spiking neural networks (SNNs) have advantages of biological plausibility, event-driven sparsity, and binary activation. Recently, large-scale language models exhibit promising generalization capability, making it a valuable issue to explore more general spike-driven models. However, the binary spikes in existing SNNs fail to encode adequate semantic information, placing technological challenges for generalization. This work proposes the first fully spiking mechanism for general language tasks, including both discriminative and generative ones. Different from previous spikes with {0,1} levels, we propose a more general spike formulation with bi-directional, elastic amplitude, and elastic frequency encoding, while still maintaining the addition nature of SNNs. In a single time step, the spike is enhanced by direction and amplitude information; in spike frequency, a strategy to control spike firing rate is well designed. We plug this elastic bi-spiking mechanism in language modeling, named SpikeLM. It is the first time to handle general language tasks with fully spike-driven models, which achieve much higher accuracy than previously possible. SpikeLM also greatly bridges the performance gap between SNNs and ANNs in language modeling. Our code is available at https://github.com/Xingrun-Xing/SpikeLM.

Recent language model-based text-to-speech (TTS) frameworks demonstrate scalability and in-context learning capabilities. However, they suffer from robustness issues due to the accumulation of errors in speech unit predictions during autoregressive language modeling. In this paper, we propose a phonetic enhanced language modeling method to improve the performance of TTS models. We leverage self-supervised representations that are phonetically rich as the training target for the autoregressive language model. Subsequently, a non-autoregressive model is employed to predict discrete acoustic codecs that contain fine-grained acoustic details. The TTS model focuses solely on linguistic modeling during autoregressive training, thereby reducing the error propagation that occurs in non-autoregressive training. Both objective and subjective evaluations validate the effectiveness of our proposed method., Comment: Accepted by Interspeech 2024

Tabular data -- structured, heterogeneous, spreadsheet-style data with rows and columns -- is widely used in practice across many domains. However, while recent foundation models have reduced the need for developing task-specific datasets and predictors in domains such as language modeling and computer vision, this transfer learning paradigm has not had similar impact in the tabular domain. In this work, we seek to narrow this gap and present TabuLa-8B, a language model for tabular prediction. We define a process for extracting a large, high-quality training dataset from the TabLib corpus, proposing methods for tabular data filtering and quality control. Using the resulting dataset, which comprises over 2.1B rows from over 4M unique tables, we fine-tune a Llama 3-8B large language model (LLM) for tabular data prediction (classification and binned regression) using a novel packing and attention scheme for tabular prediction. Through evaluation across a test suite of 329 datasets, we find that TabuLa-8B has zero-shot accuracy on unseen tables that is over 15 percentage points (pp) higher than random guessing, a feat that is not possible with existing state-of-the-art tabular prediction models (e.g. XGBoost, TabPFN). In the few-shot setting (1-32 shots), without any fine-tuning on the target datasets, TabuLa-8B is 5-15 pp more accurate than XGBoost and TabPFN models that are explicitly trained on equal, or even up to 16x more data. We release our model, code, and data along with the publication of this paper., Comment: NeurIPS 2024 camera-ready updates

Large language models (LLMs) have shown promising efficacy across various tasks, becoming powerful tools in numerous aspects of human life. However, Transformer-based LLMs suffer a performance degradation when modeling long-term contexts due to they discard some information to reduce computational overhead. In this work, we propose a simple yet effective method to enable LLMs to take a deep breath, encouraging them to summarize information contained within discrete text chunks. Specifically, we segment the text into multiple chunks and insert special token at the end of each chunk. We then modify the attention mask to integrate the chunk's information into the corresponding token. This facilitates LLMs to interpret information not only from historical individual tokens but also from the token, aggregating the chunk's semantic information. Experiments on language modeling and out-of-domain downstream tasks validate the superiority of our approach.

The visual quality of an image is confounded by a number of intertwined factors including its semantic content, distortion characteristics and appearance properties such as brightness, contrast, sharpness, and colourfulness. Distilling high level knowledge about all these quality bearing attributes is crucial for developing objective Image Quality Assessment (IQA).While existing solutions have modeled some of these aspects, a comprehensive solution that involves all these important quality related attributes has not yet been developed. In this paper, we present a new blind IQA (BIQA) model termed Self-supervision and Vision-Language supervision Image QUality Evaluator (SLIQUE) that features a joint vision-language and visual contrastive representation learning framework for acquiring high level knowledge about the images semantic contents, distortion characteristics and appearance properties for IQA. For training SLIQUE, we have developed a systematic approach to constructing a first of its kind large image database annotated with all three categories of quality relevant texts. The Text Annotated Distortion, Appearance and Content (TADAC) database has over 1.6 million images annotated with textual descriptions of their semantic contents, distortion characteristics and appearance properties. The method for constructing TADAC and the database itself will be particularly useful for exploiting vision-language modeling for advanced IQA applications. Extensive experimental results show that SLIQUE has superior performances over state of the art, demonstrating the soundness of its design principle and the effectiveness of its implementation., Comment: This paper has been accepted for publication in Journal of Selected Topics in Signal Processing, \c{opyright} 2024 IEEE. Personal use is permitted. For other uses, permission must be obtained from IEEE

Dysarthric speech reconstruction (DSR) aims to transform dysarthric speech into normal speech. It still suffers from low speaker similarity and poor prosody naturalness. In this paper, we propose a multi-modal DSR model by leveraging neural codec language modeling to improve the reconstruction results, especially for the speaker similarity and prosody naturalness. Our proposed model consists of: (i) a multi-modal content encoder to extract robust phoneme embeddings from dysarthric speech with auxiliary visual inputs; (ii) a speaker codec encoder to extract and normalize the speaker-aware codecs from the dysarthric speech, in order to provide original timbre and normal prosody; (iii) a codec language model based speech decoder to reconstruct the speech based on the extracted phoneme embeddings and normalized codecs. Evaluations on the commonly used UASpeech corpus show that our proposed model can achieve significant improvements in terms of speaker similarity and prosody naturalness., Comment: Accepted by Interspeech 2024

In this paper, our goal is to generate synthetic data for heterogeneous (mixed-type) tabular datasets with high machine learning utility (MLu). Since the MLu performance depends on accurately approximating the conditional distributions, we focus on devising a synthetic data generation method based on conditional distribution estimation. We introduce MaCoDE by redefining the consecutive multi-class classification task of Masked Language Modeling (MLM) as histogram-based non-parametric conditional density estimation. Our approach enables the estimation of conditional densities across arbitrary combinations of target and conditional variables. We bridge the theoretical gap between distributional learning and MLM by demonstrating that minimizing the orderless multi-class classification loss leads to minimizing the total variation distance between conditional distributions. To validate our proposed model, we evaluate its performance in synthetic data generation across 10 real-world datasets, demonstrating its ability to adjust data privacy levels easily without re-training. Additionally, since masked input tokens in MLM are analogous to missing data, we further assess its effectiveness in handling training datasets with missing values, including multiple imputations of the missing entries.

We propose a novel tensor network language model based on the simplest tensor network (i.e., tensor trains), called `Tensor Train Language Model' (TTLM). TTLM represents sentences in an exponential space constructed by the tensor product of words, but computing the probabilities of sentences in a low-dimensional fashion. We demonstrate that the architectures of Second-order RNNs, Recurrent Arithmetic Circuits (RACs), and Multiplicative Integration RNNs are, essentially, special cases of TTLM. Experimental evaluations on real language modeling tasks show that the proposed variants of TTLM (i.e., TTLM-Large and TTLM-Tiny) outperform the vanilla Recurrent Neural Networks (RNNs) with low-scale of hidden units. (The code is available at https://github.com/shuishen112/tensortrainlm.)

Retrieval-Augmented Language Modeling (RALM) by integrating large language models (LLM) with relevant documents from an external corpus is a proven method for enabling the LLM to generate information beyond the scope of its pre-training corpus. Previous work utilizing retrieved content by simply prepending it to the input poses a high runtime issue, which degrades the inference efficiency of the LLMs because they fail to use the Key-Value (KV) cache efficiently. In this paper, we propose FlashBack, a modular RALM designed to improve the inference efficiency of RALM with appending context pattern while maintaining decent performance after fine-tuning by Low-Rank Adaption. FlashBack appends retrieved documents at the end of the context for efficiently utilizing the KV cache instead of prepending them. And we introduce Marking Token as two special prompt tokens for marking the boundary of the appending context during fine-tuning. Our experiments on testing generation quality show that FlashBack can remain decent generation quality in perplexity. And the inference speed of FlashBack is up to $4\times$ faster than the prepending counterpart on a 7B LLM (Llama 2) in the runtime test. Via bypassing unnecessary re-computation, it demonstrates an advancement by achieving significantly faster inference speed, and this heightened efficiency will substantially reduce inferential cost., Comment: 14 pages

Tokenization is widely used in large language models because it significantly improves performance. However, tokenization imposes several disadvantages, such as performance biases, increased adversarial vulnerability, decreased character-level modeling performance, and increased modeling complexity. To address these disadvantages without sacrificing performance, we propose SpaceByte, a novel byte-level decoder architecture that closes the performance gap between byte-level and subword autoregressive language modeling. SpaceByte consists of a byte-level Transformer model, but with extra larger transformer blocks inserted in the middle of the layers. We find that performance is significantly improved by applying these larger blocks only after certain bytes, such as space characters, which typically denote word boundaries. Our experiments show that for a fixed training and inference compute budget, SpaceByte outperforms other byte-level architectures and roughly matches the performance of tokenized Transformer architectures., Comment: 9+10 pages, 3+1 figures, 2+4 tables

Predicting the future is of great interest across many aspects of human activity. Businesses are interested in future trends, traders are interested in future stock prices, and companies are highly interested in future technological breakthroughs. While there are many automated systems for predicting future numerical data, such as weather, stock prices, and demand for products, there is relatively little work in automatically predicting textual data. Humans are interested in textual data predictions because it is a natural format for our consumption, and experts routinely make predictions in a textual format (Christensen et al., 2004; Tetlock & Gardner, 2015; Frick, 2015). However, there has been relatively little formalization of this general problem in the machine learning or natural language processing communities. To address this gap, we introduce the task of future language modeling: probabilistic modeling of texts in the future based on a temporal history of texts. To our knowledge, our work is the first work to formalize the task of predicting the future in this way. We show that it is indeed possible to build future language models that improve upon strong non-temporal language model baselines, opening the door to working on this important, and widely applicable problem., Comment: Accepted by ICLR 2024

In recent years, the use of open-source models has gained immense popularity in various fields, including legal language modelling and analysis. These models have proven to be highly effective in tasks such as summarizing legal documents, extracting key information, and even predicting case outcomes. This has revolutionized the legal industry, enabling lawyers, researchers, and policymakers to quickly access and analyse vast amounts of legal text, saving time and resources. This paper presents a novel approach to legal language modeling (LLM) and analysis using open-source models from Hugging Face. We leverage Hugging Face embeddings via LangChain and Sentence Transformers to develop an LLM tailored for legal texts. We then demonstrate the application of this model by extracting insights from the official Constitution of India. Our methodology involves preprocessing the data, splitting it into chunks, using ChromaDB and LangChainVectorStores, and employing the Google/Flan-T5-XXL model for analysis. The trained model is tested on the Indian Constitution, which is available in PDF format. Our findings suggest that our approach holds promise for efficient legal language processing and analysis., Comment: 10 Pages , 3 figures

Methods for learning latent user representations from historical behavior logs have gained traction for recommendation tasks in e-commerce, content streaming, and other settings. However, this area still remains relatively underexplored in video and mobile gaming contexts. In this work, we present a novel method for overcoming this limitation by extending a long-range Transformer model from the natural language processing domain to player behavior data. We discuss specifics of behavior tracking in games and propose preprocessing and tokenization approaches by viewing in-game events in an analogous way to words in sentences, thus enabling learning player representations in a self-supervised manner in the absence of ground-truth annotations. We experimentally demonstrate the efficacy of the proposed approach in fitting the distribution of behavior events by evaluating intrinsic language modeling metrics. Furthermore, we qualitatively analyze the emerging structure of the learned embedding space and show its value for generating insights into behavior patterns to inform downstream applications.

The emergence of Large Language Models (LLMs) has achieved tremendous success in the field of Natural Language Processing owing to diverse training paradigms that empower LLMs to effectively capture intricate linguistic patterns and semantic representations. In particular, the recent "pre-train, prompt and predict" training paradigm has attracted significant attention as an approach for learning generalizable models with limited labeled data. In line with this advancement, these training paradigms have recently been adapted to the recommendation domain and are seen as a promising direction in both academia and industry. This half-day tutorial aims to provide a thorough understanding of extracting and transferring knowledge from pre-trained models learned through different training paradigms to improve recommender systems from various perspectives, such as generality, sparsity, effectiveness and trustworthiness. In this tutorial, we first introduce the basic concepts and a generic architecture of the language modeling paradigm for recommendation purposes. Then, we focus on recent advancements in adapting LLM-related training strategies and optimization objectives for different recommendation tasks. After that, we will systematically introduce ethical issues in LLM-based recommender systems and discuss possible approaches to assessing and mitigating them. We will also summarize the relevant datasets, evaluation metrics, and an empirical study on the recommendation performance of training paradigms. Finally, we will conclude the tutorial with a discussion of open challenges and future directions., Comment: Tutorial held at the 27th European Conference on Artificial Intelligence (ECAI) in Santiago de Compostela, Spain, on October 19-24, 2024

We present RALL-E, a robust language modeling method for text-to-speech (TTS) synthesis. While previous work based on large language models (LLMs) shows impressive performance on zero-shot TTS, such methods often suffer from poor robustness, such as unstable prosody (weird pitch and rhythm/duration) and a high word error rate (WER), due to the autoregressive prediction style of language models. The core idea behind RALL-E is chain-of-thought (CoT) prompting, which decomposes the task into simpler steps to enhance the robustness of LLM-based TTS. To accomplish this idea, RALL-E first predicts prosody features (pitch and duration) of the input text and uses them as intermediate conditions to predict speech tokens in a CoT style. Second, RALL-E utilizes the predicted duration prompt to guide the computing of self-attention weights in Transformer to enforce the model to focus on the corresponding phonemes and prosody features when predicting speech tokens. Results of comprehensive objective and subjective evaluations demonstrate that, compared to a powerful baseline method VALL-E, RALL-E significantly improves the WER of zero-shot TTS from $5.6\%$ (without reranking) and $1.7\%$ (with reranking) to $2.5\%$ and $1.0\%$, respectively. Furthermore, we demonstrate that RALL-E correctly synthesizes sentences that are hard for VALL-E and reduces the error rate from $68\%$ to $4\%$.

The modern autoregressive Large Language Models (LLMs) have achieved outstanding performance on NLP benchmarks, and they are deployed in the real world. However, they still suffer from limitations of the autoregressive training paradigm. For example, autoregressive token generation is notably slow and can be prone to \textit{exposure bias}. The diffusion-based language models were proposed as an alternative to autoregressive generation to address some of these limitations. We evaluate the recently proposed Score Entropy Discrete Diffusion (SEDD) approach and show it is a promising alternative to autoregressive generation but it has some short-comings too. We empirically demonstrate the advantages and challenges of SEDD, and observe that SEDD generally matches autoregressive models in perplexity and on benchmarks such as HellaSwag, Arc or WinoGrande. Additionally, we show that in terms of inference latency, SEDD can be up to 4.5$\times$ more efficient than GPT-2. While SEDD allows conditioning on tokens at abitrary positions, SEDD appears slightly weaker than GPT-2 for conditional generation given short prompts. Finally, we reproduced the main results from the original SEDD paper.

When the world changes, so does the text that humans write about it. How do we build language models that can be easily updated to reflect these changes? One popular approach is retrieval-augmented generation, in which new documents are inserted into a knowledge base and retrieved during prediction for downstream tasks. Most prior work on these systems have focused on improving behavior during prediction through better retrieval or reasoning. This paper introduces ERASE, which instead improves model behavior when new documents are acquired, by incrementally deleting or rewriting other entries in the knowledge base each time a document is added. In two new benchmark datasets evaluating models' ability to answer questions about a stream of news articles or conversations, ERASE improves accuracy relative to conventional retrieval-augmented generation by 7-13% (Mixtral-8x7B) and 6-10% (Llama-3-8B) absolute. Code and data are available at https://github.com/belindal/ERASE

The identification of Figurative Language (FL) features in text is crucial for various Natural Language Processing (NLP) tasks, where understanding of the author's intended meaning and its nuances is key for successful communication. At the same time, the use of a specific blend of various FL forms most accurately reflects a writer's style, rather than the use of any single construct, such as just metaphors or irony. Thus, we postulate that FL features could play an important role in Authorship Attribution (AA) tasks. We believe that our is the first computational study of AA based on FL use. Accordingly, we propose a Multi-task Figurative Language Model (MFLM) that learns to detect multiple FL features in text at once. We demonstrate, through detailed evaluation across multiple test sets, that the our model tends to perform equally or outperform specialized binary models in FL detection. Subsequently, we evaluate the predictive capability of joint FL features towards the AA task on three datasets, observing improved AA performance through the integration of MFLM embeddings.

Retrieval-augmented language models (RALMs) have recently shown great potential in mitigating the limitations of implicit knowledge in LLMs, such as untimely updating of the latest expertise and unreliable retention of long-tail knowledge. However, since the external knowledge base, as well as the retriever, can not guarantee reliability, potentially leading to the knowledge retrieved not being helpful or even misleading for LLM generation. In this paper, we introduce Supportiveness-based Knowledge Rewriting (SKR), a robust and pluggable knowledge rewriter inherently optimized for LLM generation. Specifically, we introduce the novel concept of "supportiveness"--which represents how effectively a knowledge piece facilitates downstream tasks--by considering the perplexity impact of augmented knowledge on the response text of a white-box LLM. Based on knowledge supportiveness, we first design a training data curation strategy for our rewriter model, effectively identifying and filtering out poor or irrelevant rewrites (e.g., with low supportiveness scores) to improve data efficacy. We then introduce the direct preference optimization (DPO) algorithm to align the generated rewrites to optimal supportiveness, guiding the rewriter model to summarize augmented content that better improves the final response. Comprehensive evaluations across six popular knowledge-intensive tasks and four LLMs have demonstrated the effectiveness and superiority of SKR. With only 7B parameters, SKR has shown better knowledge rewriting capability over GPT-4, the current state-of-the-art general-purpose LLM.

Efficiently modeling sequences with infinite context length has long been a challenging problem. Previous approaches have either suffered from quadratic computational complexity or limited extrapolation ability in length generalization. In this work, we present Samba, a simple hybrid architecture that layer-wise combines Mamba, a selective State Space Model (SSM), with Sliding Window Attention (SWA). Samba selectively compresses a given sequence into recurrent hidden states while still maintaining the ability to precisely recall recent memories with the attention mechanism. We scale Samba up to 3.8B parameters with 3.2T training tokens and demonstrate that it significantly outperforms state-of-the-art models across a variety of benchmarks. Pretrained on sequences of 4K length, Samba shows improved perplexity in context lengths of up to 1M in zero-shot. When finetuned on 4K-length sequences, Samba efficiently extrapolates to a 256K context length with perfect memory recall on the Passkey Retrieval task, and exhibits superior retrieval extrapolation on the challenging Phonebook task compared to full-attention models. As a linear-time sequence model, Samba achieves a 3.73x higher throughput compared to Transformers with grouped-query attention for user prompts of 128K length, and a 3.64x speedup when generating 64K tokens with unlimited streaming. Our code for training on open source data is publicly available at https://github.com/microsoft/Samba.

Generating novel active molecules for a given protein is an extremely challenging task for generative models that requires an understanding of the complex physical interactions between the molecule and its environment. In this paper, we present a novel generative model, BindGPT which uses a conceptually simple but powerful approach to create 3D molecules within the protein's binding site. Our model produces molecular graphs and conformations jointly, eliminating the need for an extra graph reconstruction step. We pretrain BindGPT on a large-scale dataset and fine-tune it with reinforcement learning using scores from external simulation software. We demonstrate how a single pretrained language model can serve at the same time as a 3D molecular generative model, conformer generator conditioned on the molecular graph, and a pocket-conditioned 3D molecule generator. Notably, the model does not make any representational equivariance assumptions about the domain of generation. We show how such simple conceptual approach combined with pretraining and scaling can perform on par or better than the current best specialized diffusion models, language models, and graph neural networks while being two orders of magnitude cheaper to sample.

The size of large language models (LLMs) has scaled dramatically in recent years and their computational and data requirements have surged correspondingly. State-of-the-art language models, even at relatively smaller sizes, typically require training on at least a trillion tokens. This rapid advancement has eclipsed the growth of open-source datasets available for large-scale LLM pretraining. In this paper, we introduce Zyda (Zyphra Dataset), a dataset under a permissive license comprising 1.3 trillion tokens, assembled by integrating several major respected open-source datasets into a single, high-quality corpus. We apply rigorous filtering and deduplication processes, both within and across datasets, to maintain and enhance the quality derived from the original datasets. Our evaluations show that Zyda not only competes favorably with other open datasets like Dolma, FineWeb, and RefinedWeb, but also substantially improves the performance of comparable models from the Pythia suite. Our rigorous data processing methods significantly enhance Zyda's effectiveness, outperforming even the best of its constituent datasets when used independently.

We introduce the Block Transformer which adopts hierarchical global-to-local modeling to autoregressive transformers to mitigate the inference bottlenecks associated with self-attention. Self-attention requires the key-value (KV) cache of all previous sequences to be retrieved from memory at every decoding step to retrieve context information, leading to two primary bottlenecks during batch inference. First, there is a significant delay in obtaining the first token, as the information of the entire prompt must first be processed to prefill the KV cache. Second, computation of subsequent tokens is bottlenecked by the high memory I/O demand of fetching the entire KV cache, which grows linearly with sequence length, incurring quadratic memory reads overall. We design the Block Transformer to strategically mitigate these costs, by incorporating coarsity and locality into an integrated global-to-local architecture. At the lower layers, we aggregate tokens into fixed size blocks to apply attention across the entire sequence at coarse-grained detail, to capture the global context while minimizing KV cache overhead. At upper layers, we apply attention within each block to decode individual tokens, to model fine-grained details with a lightweight local KV cache. We pretrain vanilla and Block Transformers from scratch and demonstrate that Block Transformers reach 10--20x inference throughput compared to vanilla transformers with equivalent perplexity and zero-shot task performance. Code is available at https://github.com/itsnamgyu/block-transformer., Comment: 37 pages, 24 figures, 7 tables

Matrix multiplication (MatMul) typically dominates the overall computational cost of large language models (LLMs). This cost only grows as LLMs scale to larger embedding dimensions and context lengths. In this work, we show that MatMul operations can be completely eliminated from LLMs while maintaining strong performance at billion-parameter scales. Our experiments show that our proposed MatMul-free models achieve performance on-par with state-of-the-art Transformers that require far more memory during inference at a scale up to at least 2.7B parameters. We investigate the scaling laws and find that the performance gap between our MatMul-free models and full precision Transformers narrows as the model size increases. We also provide a GPU-efficient implementation of this model which reduces memory usage by up to 61% over an unoptimized baseline during training. By utilizing an optimized kernel during inference, our model's memory consumption can be reduced by more than 10x compared to unoptimized models. To properly quantify the efficiency of our architecture, we build a custom hardware solution on an FPGA which exploits lightweight operations beyond what GPUs are capable of. We processed billion-parameter scale models at 13W beyond human readable throughput, moving LLMs closer to brain-like efficiency. This work not only shows how far LLMs can be stripped back while still performing effectively, but also points at the types of operations future accelerators should be optimized for in processing the next generation of lightweight LLMs. Our code implementation is available at https://github.com/ridgerchu/matmulfreellm.

We propose a new transformer model for the Traveling Salesman Problem (TSP) called CycleFormer. We identified distinctive characteristics that need to be considered when applying a conventional transformer model to TSP and aimed to fully incorporate these elements into the TSP-specific transformer. Unlike the token sets in typical language models, which are limited and static, the token (node) set in TSP is unlimited and dynamic. To exploit this fact to the fullest, we equated the encoder output with the decoder linear layer and directly connected the context vector of the encoder to the decoder encoding. Additionally, we added a positional encoding to the encoder tokens that reflects the two-dimensional nature of TSP, and devised a circular positional encoding for the decoder tokens that considers the cyclic properties of a tour. By incorporating these ideas, CycleFormer outperforms state-of-the-art (SOTA) transformer models for TSP from TSP-50 to TSP-500. Notably, on TSP-500, the optimality gap was reduced by approximately 2.8 times, from 3.09% to 1.10%, compared to the existing SOTA. The code will be made available at https://github.com/Giventicket/CycleFormer., Comment: The paper's content (experiments) is insufficient

We present Lightning Attention, the first linear attention implementation that maintains a constant training speed for various sequence lengths under fixed memory consumption. Due to the issue with cumulative summation operations (cumsum), previous linear attention implementations cannot achieve their theoretical advantage in a casual setting. However, this issue can be effectively solved by utilizing different attention calculation strategies to compute the different parts of attention. Specifically, we split the attention calculation into intra-blocks and inter-blocks and use conventional attention computation for intra-blocks and linear attention kernel tricks for inter-blocks. This eliminates the need for cumsum in the linear attention calculation. Furthermore, a tiling technique is adopted through both forward and backward procedures to take full advantage of the GPU hardware. To enhance accuracy while preserving efficacy, we introduce TransNormerLLM (TNL), a new architecture that is tailored to our lightning attention. We conduct rigorous testing on standard and self-collected datasets with varying model sizes and sequence lengths. TNL is notably more efficient than other language models. In addition, benchmark results indicate that TNL performs on par with state-of-the-art LLMs utilizing conventional transformer structures. The source code is released at github.com/OpenNLPLab/TransnormerLLM., Comment: Accepted by ICML 2024. Yiran Zhong is the corresponding author. Code is released at github.com/OpenNLPLab/TransnormerLLM

Following the recent popularity of Large Language Models (LLMs), several attempts have been made to extend them to the visual domain. From having a visual assistant that could guide us through unfamiliar environments to generative models that produce images using only a high-level text description, the vision-language model (VLM) applications will significantly impact our relationship with technology. However, there are many challenges that need to be addressed to improve the reliability of those models. While language is discrete, vision evolves in a much higher dimensional space in which concepts cannot always be easily discretized. To better understand the mechanics behind mapping vision to language, we present this introduction to VLMs which we hope will help anyone who would like to enter the field. First, we introduce what VLMs are, how they work, and how to train them. Then, we present and discuss approaches to evaluate VLMs. Although this work primarily focuses on mapping images to language, we also discuss extending VLMs to videos.

This paper will present textual corpora for Serbian (and Serbo-Croatian), usable for the training of large language models and publicly available at one of the several notable online repositories. Each corpus will be classified using multiple methods and its characteristics will be detailed. Additionally, the paper will introduce three new corpora: a new umbrella web corpus of Serbo-Croatian, a new high-quality corpus based on the doctoral dissertations stored within National Repository of Doctoral Dissertations from all Universities in Serbia, and a parallel corpus of abstract translation from the same source. The uniqueness of both old and new corpora will be accessed via frequency-based stylometric methods, and the results will be briefly discussed.

How much is 56 times 37? Language models often make mistakes in these types of difficult calculations. This is usually explained by their inability to perform complex reasoning. Since language models rely on large training sets and great memorization capability, naturally they are not equipped to run complex calculations. However, one can argue that humans also cannot perform this calculation immediately and require a considerable amount of time to construct the solution. In order to enhance the generalization capability of language models, and as a parallel to human behavior, we propose to use special 'thinking tokens' which allow the model to perform much more calculations whenever a complex problem is encountered., Comment: AITP 2023 (May 10, 2023)

SMS, or short messaging service, is a widely used and cost-effective communication medium that has sadly turned into a haven for unwanted messages, commonly known as SMS spam. With the rapid adoption of smartphones and Internet connectivity, SMS spam has emerged as a prevalent threat. Spammers have taken notice of the significance of SMS for mobile phone users. Consequently, with the emergence of new cybersecurity threats, the number of SMS spam has expanded significantly in recent years. The unstructured format of SMS data creates significant challenges for SMS spam detection, making it more difficult to successfully fight spam attacks in the cybersecurity domain. In this work, we employ optimized and fine-tuned transformer-based Large Language Models (LLMs) to solve the problem of spam message detection. We use a benchmark SMS spam dataset for this spam detection and utilize several preprocessing techniques to get clean and noise-free data and solve the class imbalance problem using the text augmentation technique. The overall experiment showed that our optimized fine-tuned BERT (Bidirectional Encoder Representations from Transformers) variant model RoBERTa obtained high accuracy with 99.84\%. We also work with Explainable Artificial Intelligence (XAI) techniques to calculate the positive and negative coefficient scores which explore and explain the fine-tuned model transparency in this text-based spam SMS detection task. In addition, traditional Machine Learning (ML) models were also examined to compare their performance with the transformer-based models. This analysis describes how LLMs can make a good impact on complex textual-based spam data in the cybersecurity field.

The intrinsic capability of the Human Vision System (HVS) to perceive depth of field and failure of Instrument Landing Systems (ILS) stimulates a pilot to perform a vision-based manual landing over an autoland approach. However, harsh weather creates challenges, and a pilot must have a clear view of runway elements before the minimum decision altitude. To aid in manual landing, a vision-based system trained to clear weather-induced visual degradations requires a robust landing dataset under various climatic conditions. Nevertheless, to acquire a dataset, flying an aircraft in dangerous weather impacts safety. Also, this system fails to generate reliable warnings, as localization of runway elements suffers from projective distortion while landing at crosswind. To combat, we propose to synthesize harsh weather landing images by training a prompt-based climatic diffusion network. Also, we optimize a weather distillation model using a novel diffusion-distillation loss to learn to clear these visual degradations. Precisely, the distillation model learns an inverse relationship with the diffusion network. Inference time, pre-trained distillation network directly clears weather-impacted onboard camera images, which can be further projected to display devices for improved visibility.Then, to tackle crosswind landing, a novel Regularized Spatial Transformer Networks (RuSTaN) module accurately warps landing images. It minimizes the localization error of runway object detector and helps generate reliable internal software warnings. Finally, we curated an aircraft landing dataset (AIRLAD) by simulating a landing scenario under various weather degradations and experimentally validated our contributions., Comment: Accepted in Indian Conference on Vision Graphics and Image Processing - ICVGIP 2024

Chain-of-Thought (CoT) reasoning holds great promise for explaining the outputs of language models, but recent studies have highlighted significant challenges in its practical application for interpretability. We propose to address this issue via two key components: a technique to factor next-token prediction through intermediate CoT text, ensuring the CoT is causally load-bearing, and a reinforcement learning approach to train CoT to predict future tokens independently of other context. This results in "Markovian" language models, where CoT serves as a fixed-size state for future token prediction. Our approach optimizes for "informativeness" -- the improvement in next-token predictions using a trained CoT compared to a baseline. We demonstrate our method's effectiveness using Proximal Policy Optimization (PPO) on arithmetic problems and achieve an 11% performance boost on the GSM8K benchmark using Mistral 7B Inst V2. The increased sensitivity of model performance to CoT perturbations provides strong evidence of CoT reliance. This work advances the development of more transparent and interpretable language models, potentially enabling their extension to arbitrarily long contexts and enhancing AI reasoning capabilities across various domains., Comment: 16 pages, 5 figures

This study introduces a novel method for irony detection, applying Large Language Models (LLMs) with prompt-based learning to facilitate emotion-centric text augmentation. Traditional irony detection techniques typically fall short due to their reliance on static linguistic features and predefined knowledge bases, often overlooking the nuanced emotional dimensions integral to irony. In contrast, our methodology augments the detection process by integrating subtle emotional cues, augmented through LLMs, into three benchmark pre-trained NLP models - BERT, T5, and GPT-2 - which are widely recognized as foundational in irony detection. We assessed our method using the SemEval-2018 Task 3 dataset and observed substantial enhancements in irony detection capabilities., Comment: 11 pages, 3 tables, 2 figures. Accepted by the 25th Chinese Lexical Semantics Workshop

In large language model training, input documents are typically concatenated together and then split into sequences of equal length to avoid padding tokens. Despite its efficiency, the concatenation approach compromises data integrity -- it inevitably breaks many documents into incomplete pieces, leading to excessive truncations that hinder the model from learning to compose logically coherent and factually consistent content that is grounded on the complete context. To address the issue, we propose Best-fit Packing, a scalable and efficient method that packs documents into training sequences through length-aware combinatorial optimization. Our method completely eliminates unnecessary truncations while retaining the same training efficiency as concatenation. Empirical results from both text and code pre-training show that our method achieves superior performance (e.g., relatively +4.7% on reading comprehension; +16.8% in context following; and +9.2% on program synthesis), and reduces closed-domain hallucination effectively by up to 58.3%., Comment: ICML 2024

Off-the-shelf pre-trained language models have become the de facto standard in NLP pipelines for a multitude of downstream tasks. However, the inability of these models to properly encode numerals limits their performance on tasks requiring numeric comprehension. We introduce strategies to semantically prime numerals in any corpus by generating anchors governed by the distribution of numerals in said corpus, thereby enabling mathematically grounded representations of these numeral tokens. We establish the superiority of our proposed techniques through evaluation on a range of numeracy tasks for both in-domain (seen) and out-domain (unseen) numerals. Further, we expand our empirical evaluations to numerals ranging from 1 to 10 billion, a significantly broader range compared to previous studies of the same nature, and we demonstrate significant improvements in the mathematical grounding of our learned embeddings., Comment: Accepted to the findings of NAACL 2024

Ribosomally synthesized and post-translationally modified peptide (RiPP) biosynthetic enzymes often exhibit promiscuous substrate preferences that cannot be reduced to simple rules. Large language models are promising tools for predicting such peptide fitness landscapes. However, state-of-the-art protein language models are trained on relatively few peptide sequences. A previous study comprehensively profiled the peptide substrate preferences of LazBF (a two-component serine dehydratase) and LazDEF (a three-component azole synthetase) from the lactazole biosynthetic pathway. We demonstrated that masked language modeling of LazBF substrate preferences produced language model embeddings that improved downstream classification models of both LazBF and LazDEF substrates. Similarly, masked language modeling of LazDEF substrate preferences produced embeddings that improved the performance of classification models of both LazBF and LazDEF substrates. Our results suggest that the models learned functional forms that are transferable between distinct enzymatic transformations that act within the same biosynthetic pathway. Our transfer learning method improved performance and data efficiency in data-scarce scenarios. We then fine-tuned models on each data set and showed that the fine-tuned models provided interpretable insight that we anticipate will facilitate the design of substrate libraries that are compatible with desired RiPP biosynthetic pathways.

BERT (Bidirectional Encoder Representations from Transformers) has revolutionized the field of natural language processing through its exceptional performance on numerous tasks. Yet, the majority of researchers have mainly concentrated on enhancements related to the model structure, such as relative position embedding and more efficient attention mechanisms. Others have delved into pretraining tricks associated with Masked Language Modeling, including whole word masking. DeBERTa introduced an enhanced decoder adapted for BERT's encoder model for pretraining, proving to be highly effective. We argue that the design and research around enhanced masked language modeling decoders have been underappreciated. In this paper, we propose several designs of enhanced decoders and introduce BPDec (BERT Pretraining Decoder), a novel method for modeling training. Typically, a pretrained BERT model is fine-tuned for specific Natural Language Understanding (NLU) tasks. In our approach, we utilize the original BERT model as the encoder, making only changes to the decoder without altering the encoder. This approach does not necessitate extensive modifications to the encoder architecture and can be seamlessly integrated into existing fine-tuning pipelines and services, offering an efficient and effective enhancement strategy. Compared to other methods, while we also incur a moderate training cost for the decoder during the pretraining process, our approach does not introduce additional training costs during the fine-tuning phase. We test multiple enhanced decoder structures after pretraining and evaluate their performance on the GLUE tasks and SQuAD tasks. Our results demonstrate that BPDec, having only undergone subtle refinements to the model structure during pretraining, significantly enhances model performance without escalating the finetuning cost, inference time and serving budget.

By training to predict the next token in an unlabeled corpus, large language models learn to perform many tasks without any labeled data. However, their next-token-prediction objective arguably limits their performance in scenarios that require planning, such as writing a coherent article. In this paper, we train a module for planning the future writing process via a self-supervised learning objective. Given the textual context, this planning module learns to predict future abstract writing actions, which correspond to centroids in a clustered text embedding space. By conditioning on these actions, our model extends the successful language model formula to more abstract planning in an unsupervised way. Empirically, we demonstrate that our method improves language modeling performance in general, particularly with respect to the text structure. Because our framework uses a planner module that is unsupervised and external to the language model, new planner modules can be trained at large scale and easily be shared with the community., Comment: Published at COLM 2024

Reference resolution is an important problem, one that is essential to understand and successfully handle context of different kinds. This context includes both previous turns and context that pertains to non-conversational entities, such as entities on the user's screen or those running in the background. While LLMs have been shown to be extremely powerful for a variety of tasks, their use in reference resolution, particularly for non-conversational entities, remains underutilized. This paper demonstrates how LLMs can be used to create an extremely effective system to resolve references of various types, by showing how reference resolution can be converted into a language modeling problem, despite involving forms of entities like those on screen that are not traditionally conducive to being reduced to a text-only modality. We demonstrate large improvements over an existing system with similar functionality across different types of references, with our smallest model obtaining absolute gains of over 5% for on-screen references. We also benchmark against GPT-3.5 and GPT-4, with our smallest model achieving performance comparable to that of GPT-4, and our larger models substantially outperforming it., Comment: Accepted at SIGDIAL 2024 (Oral presentation)

Today's most accurate language models are trained on orders of magnitude more language data than human language learners receive - but with no supervision from other sensory modalities that play a crucial role in human learning. Can we make LMs' representations and predictions more accurate (and more human-like) with more ecologically plausible supervision? This paper describes LexiContrastive Grounding (LCG), a grounded language learning procedure that leverages visual supervision to improve textual representations. LexiContrastive Grounding combines a next token prediction strategy with a contrastive visual grounding objective, focusing on early-layer representations that encode lexical information. Across multiple word-learning and sentence-understanding benchmarks, LexiContrastive Grounding not only outperforms standard language-only models in learning efficiency, but also improves upon vision-and-language learning procedures including CLIP, GIT, Flamingo, and Vokenization. Moreover, LexiContrastive Grounding improves perplexity by around 5% on multiple language modeling tasks. This work underscores the potential of incorporating visual grounding into language models, aligning more closely with the multimodal nature of human language acquisition.

A major consideration in multilingual language modeling is how to best represent languages with diverse vocabularies and scripts. Although contemporary text encoding methods cover most of the world's writing systems, they exhibit bias towards the high-resource languages of the Global West. As a result, texts of underrepresented languages tend to be segmented into long sequences of linguistically meaningless units. To address the disparities, we introduce a new paradigm that encodes the same information with segments of consistent size across diverse languages. Our encoding convention (MYTE) is based on morphemes, as their inventories are more balanced across languages than characters, which are used in previous methods. We show that MYTE produces shorter encodings for all 99 analyzed languages, with the most notable improvements for non-European languages and non-Latin scripts. This, in turn, improves multilingual LM performance and diminishes the perplexity gap throughout diverse languages., Comment: Published at ACL 2024

While (large) language models have significantly improved over the last years, they still struggle to sensibly process long sequences found, e.g., in books, due to the quadratic scaling of the underlying attention mechanism. To address this, we propose NextLevelBERT, a Masked Language Model operating not on tokens, but on higher-level semantic representations in the form of text embeddings. We pretrain NextLevelBERT to predict the vector representation of entire masked text chunks and evaluate the effectiveness of the resulting document vectors on three types of tasks: 1) Semantic Textual Similarity via zero-shot document embeddings, 2) Long document classification, 3) Multiple-choice question answering. We find that next-level Masked Language Modeling is an effective technique to tackle long-document use cases and can outperfor much larger embedding models as long as the required level of detail of semantic information is not too fine. Our models and code are publicly available online., Comment: accepted at ACL 2024; camera-ready version; 9 pages

Extending large language models (LLMs) to process longer inputs is crucial for a wide range of applications. However, the substantial computational cost of transformers and limited generalization of positional encoding restrict the size of their context window. We introduce Context Expansion with Parallel Encoding (CEPE), a framework that can be applied to any existing decoder-only LLMs to extend their context window. CEPE employs a small encoder to process long inputs chunk by chunk, enabling the frozen decoder to utilize additional contexts via cross-attention. CEPE is efficient, generalizable, and versatile: trained with 8K-token documents, it extends the context window of LLAMA-2 to 128K tokens, offering 10x the throughput with only 1/6 of the memory. CEPE yields strong performance on language modeling and in-context learning. CEPE also excels in retrieval-augmented applications, while existing long-context models degenerate with retrieved contexts. We further introduce a CEPE variant that can extend the context window of instruction-tuned models using only unlabeled data, and showcase its effectiveness on LLAMA-2-CHAT, leading to a strong instruction-following model that can leverage very long contexts on downstream tasks., Comment: ACL 2024. Code, models, and data are available at https://github.com/princeton-nlp/CEPE. arXiv admin note: text overlap with arXiv:1912.01214 by other authors