Retrieval-augmented generation (RAG) grounds large language model (LLM) output by leveraging external knowledge sources to reduce factual hallucinations. However, prior works lack a comprehensive evaluation of different language families, making it challenging to evaluate LLM robustness against errors in external retrieved knowledge. To overcome this, we establish NoMIRACL, a human-annotated dataset for evaluating LLM robustness in RAG across 18 typologically diverse languages. NoMIRACL includes both a non-relevant and a relevant subset. Queries in the non-relevant subset contain passages manually judged as non-relevant or noisy, whereas queries in the relevant subset include at least a single judged relevant passage. We measure LLM robustness using two metrics: (i) hallucination rate, measuring model tendency to hallucinate an answer, when the answer is not present in passages in the non-relevant subset, and (ii) error rate, measuring model inaccuracy to recognize relevant passages in the relevant subset. We build a GPT-4 baseline which achieves a 33.2% hallucination rate on the non-relevant and a 14.9% error rate on the relevant subset on average. Our evaluation reveals that GPT-4 hallucinates frequently in high-resource languages, such as French or English. This work highlights an important avenue for future research to improve LLM robustness to learn how to better reject non-relevant information in RAG.

Efficient attentions have greatly improved the computational efficiency of Transformers. However, most existing linear attention mechanisms suffer from an \emph{efficiency degradation} problem, leading to inefficiencies in causal language modeling and hindering their application in long-range language models. This problem is more pronounced under language modeling with unbounded contexts. In this paper, we propose \textbf{L}inear \textbf{A}ttention \textbf{V}ia \textbf{O}rthogonal memory~(\shortname) to address these limitations, achieving strong performance while maintaining linear complexity. \shortname employs orthogonal decomposition to compress a context into a fixed-size orthogonal memory while effectively minimizing redundancy within the context. Given that orthogonal memory compresses global information, we further dissect the context to amplify fine-grained local information. Additionally, we embed the relative position encoding into \shortname to improve the extrapolation ability. Experimental results show that \shortname greatly improves the efficiency of the causal language model with the best extrapolation performance and outperforms other efficient baselines. Further, we endeavor to employ \shortname for unbounded language modeling and successfully scale the context length to 128K.

Federated Learning (FL) has achieved significant achievements recently, enabling collaborative model training on distributed data over edge devices. Iterative gradient or model exchanges between devices and the centralized server in the standard FL paradigm suffer from severe efficiency bottlenecks on the server. While enabling collaborative training without a central server, existing decentralized FL approaches either focus on the synchronous mechanism that deteriorates FL convergence or ignore device staleness with an asynchronous mechanism, resulting in inferior FL accuracy. In this paper, we propose an Asynchronous Efficient Decentralized FL framework, i.e., AEDFL, in heterogeneous environments with three unique contributions. First, we propose an asynchronous FL system model with an efficient model aggregation method for improving the FL convergence. Second, we propose a dynamic staleness-aware model update approach to achieve superior accuracy. Third, we propose an adaptive sparse training method to reduce communication and computation costs without significant accuracy degradation. Extensive experimentation on four public datasets and four models demonstrates the strength of AEDFL in terms of accuracy (up to 16.3% higher), efficiency (up to 92.9% faster), and computation costs (up to 42.3% lower).

In streaming settings, speech recognition models have to map sub-sequences of speech to text before the full audio stream becomes available. However, since alignment information between speech and text is rarely available during training, models need to learn it in a completely self-supervised way. In practice, the exponential number of possible alignments makes this extremely challenging, with models often learning peaky or sub-optimal alignments. Prima facie, the exponential nature of the alignment space makes it difficult to even quantify the uncertainty of a model's alignment distribution. Fortunately, it has been known for decades that the entropy of a probabilistic finite state transducer can be computed in time linear to the size of the transducer via a dynamic programming reduction based on semirings. In this work, we revisit the entropy semiring for neural speech recognition models, and show how alignment entropy can be used to supervise models through regularization or distillation. We also contribute an open-source implementation of CTC and RNN-T in the semiring framework that includes numerically stable and highly parallel variants of the entropy semiring. Empirically, we observe that the addition of alignment distillation improves the accuracy and latency of an already well-optimized teacher-student distillation model, achieving state-of-the-art performance on the Librispeech dataset in the streaming scenario. Modern automatic speech recognition (ASR) systems deploy a single neural network trained in an end-to-end differentiable manner on a paired corpus of speech and text (Graves et al., 2006; Graves, 2012; Chan et al., 2015; Sak et al., 2017; He et al., 2019).

In large-scale datacenters, memory failure is a common cause of server crashes, with Uncorrectable Errors (UEs) being a major indicator of Dual Inline Memory Module (DIMM) defects. Existing approaches primarily focus on predicting UEs using Correctable Errors (CEs), without fully considering the information provided by error bits. However, error bit patterns have a strong correlation with the occurrence of UEs. In this paper, we present a comprehensive study on the correlation between CEs and UEs, specifically emphasizing the importance of spatio-temporal error bit information. Our analysis reveals a strong correlation between spatio-temporal error bits and UE occurrence. Through evaluations using real-world datasets, we demonstrate that our approach significantly improves prediction performance by 15% in F1-score compared to the state-of-the-art algorithms. Overall, our approach effectively reduces the number of virtual machine interruptions caused by UEs by approximately 59%.

Responding to the thousands of student questions on online QA platforms each semester has a considerable human cost, particularly in computing courses with rapidly growing enrollments. To address the challenges of scalable and intelligent question-answering (QA), we introduce an innovative solution that leverages open-source Large Language Models (LLMs) from the LLaMA-2 family to ensure data privacy. Our approach combines augmentation techniques such as retrieval augmented generation (RAG), supervised fine-tuning (SFT), and learning from human preferences data using Direct Preference Optimization (DPO). Through extensive experimentation on a Piazza dataset from an introductory CS course, comprising 10,000 QA pairs and 1,500 pairs of preference data, we demonstrate a significant 30% improvement in the quality of answers, with RAG being a particularly impactful addition. Our contributions include the development of a novel architecture for educational QA, extensive evaluations of LLM performance utilizing both human assessments and LLM-based metrics, and insights into the challenges and future directions of educational data processing. This work paves the way for the development of AI-TA, an intelligent QA assistant customizable for courses with an online QA platform

Representation learning stands as one of the critical machine learning techniques across various domains. Through the acquisition of high-quality features, pre-trained embeddings significantly reduce input space redundancy, benefiting downstream pattern recognition tasks such as classification, regression, or detection. Nonetheless, in the domain of tabular data, feature engineering and selection still heavily rely on manual intervention, leading to time-consuming processes and necessitating domain expertise. In response to this challenge, we introduce ReConTab, a deep automatic representation learning framework with regularized contrastive learning. Agnostic to any type of modeling task, ReConTab constructs an asymmetric autoencoder based on the same raw features from model inputs, producing low-dimensional representative embeddings. Specifically, regularization techniques are applied for raw feature selection. Meanwhile, ReConTab leverages contrastive learning to distill the most pertinent information for downstream tasks. Experiments conducted on extensive real-world datasets substantiate the framework's capacity to yield substantial and robust performance improvements. Furthermore, we empirically demonstrate that pre-trained embeddings can seamlessly integrate as easily adaptable features, enhancing the performance of various traditional methods such as XGBoost and Random Forest.

Learning from corrupted labels is very common in real-world machine-learning applications. Memorizing such noisy labels could affect the learning of the model, leading to sub-optimal performances. In this work, we propose a novel framework to learn robust machine-learning models from noisy labels. Through an empirical study, we find that different models make relatively similar predictions on clean examples, while the predictions on noisy examples vary much more across different models. Motivated by this observation, we propose \em denoising with cross-model agreement \em (DeCA) which aims to minimize the KL-divergence between the true label distributions parameterized by two machine learning models while maximizing the likelihood of data observation. We employ the proposed DeCA on both the binary label scenario and the multiple label scenario. For the binary label scenario, we select implicit feedback recommendation as the downstream task and conduct experiments with four state-of-the-art recommendation models on four datasets. For the multiple-label scenario, the downstream application is image classification on two benchmark datasets. Experimental results demonstrate that the proposed methods significantly improve the model performance compared with normal training and other denoising methods on both binary and multiple-label scenarios.

This paper studies Chinese Spelling Correction (CSC), which aims to detect and correct the potential spelling errors in a given sentence. Current state-of-the-art methods regard CSC as a sequence tagging task and fine-tune BERT-based models on sentence pairs. However, we note a critical flaw in the process of tagging one character to another, that the correction is excessively conditioned on the error. This is opposite from human mindset, where individuals rephrase the complete sentence based on its semantics, rather than solely on the error patterns memorized before. Such a counter-intuitive learning process results in the bottleneck of generalizability and transferability of machine spelling correction. To address this, we propose Rephrasing Language Model (ReLM), where the model is trained to rephrase the entire sentence by infilling additional slots, instead of character-to-character tagging. This novel training paradigm achieves the new state-of-the-art results across fine-tuned and zero-shot CSC benchmarks, outperforming previous counterparts by a large margin. Our method also learns transferable language representation when CSC is jointly trained with other tasks.

Large Language Models (LLMs), despite their recent impressive accomplishments, are notably cost-prohibitive to deploy, particularly for applications involving long-content generation, such as dialogue systems and story writing. Often, a large amount of transient state information, referred to as the KV cache, is stored in GPU memory in addition to model parameters, scaling linearly with the sequence length and batch size. In this paper, we introduce a novel approach for implementing the KV cache which significantly reduces its memory footprint. Our approach is based on the noteworthy observation that a small portion of tokens contributes most of the value when computing attention scores. We call these tokens Heavy Hitters (H$_2$). Through a comprehensive investigation, we find that (i) the emergence of H$_2$ is natural and strongly correlates with the frequent co-occurrence of tokens in the text, and (ii) removing them results in significant performance degradation. Based on these insights, we propose Heavy Hitter Oracle (H$_2$O), a KV cache eviction policy that dynamically retains a balance of recent and H$_2$ tokens. We formulate the KV cache eviction as a dynamic submodular problem and prove (under mild assumptions) a theoretical guarantee for our novel eviction algorithm which could help guide future work. We validate the accuracy of our algorithm with OPT, LLaMA, and GPT-NeoX across a wide range of tasks. Our implementation of H$_2$O with 20% heavy hitters improves the throughput over three leading inference systems DeepSpeed Zero-Inference, Hugging Face Accelerate, and FlexGen by up to 29$\times$, 29$\times$, and 3$\times$ on OPT-6.7B and OPT-30B. With the same batch size, H2O can reduce the latency by up to 1.9$\times$. The code is available at https://github.com/FMInference/H2O.