Existing LLM training and inference frameworks struggle in boosting efficiency with sparsity while maintaining the integrity of context and model architecture. Inspired by the sharding concept in database and the fact that attention parallelizes over heads on accelerators, we propose Sparsely-Sharded (S2) Attention, an attention algorithm that allocates heterogeneous context partitions for different attention heads to divide and conquer. S2-Attention enforces each attention head to only attend to a partition of contexts following a strided sparsity pattern, while the full context is preserved as the union of all the shards. As attention heads are processed in separate thread blocks, the context reduction for each head can thus produce end-to-end speed-up and memory reduction. At inference, LLMs trained with S2-Attention can then take the KV cache reduction as free meals with guaranteed model quality preserve. In experiments, we show S2-Attentioncan provide as much as (1) 25.3X wall-clock attention speed-up over FlashAttention-2, resulting in 6X reduction in end-to-end training time and 10X inference latency, (2) on-par model training quality compared to default attention, (3)perfect needle retrieval accuracy over 32K context window. On top of the algorithm, we build DKernel, an LLM training and inference kernel library that allows users to customize sparsity patterns for their own models. We open-sourced DKerneland make it compatible with Megatron, Pytorch, and vLLM.

Large Language Models (LLMs) are implicit troublemakers. While they provide valuable insights and assist in problem-solving, they can also potentially serve as a resource for malicious activities. Implementing safety alignment could mitigate the risk of LLMs generating harmful responses. We argue that: even when an LLM appears to successfully block harmful queries, there may still be hidden vulnerabilities that could act as ticking time bombs. To identify these underlying weaknesses, we propose to use a cost value model as both a detector and an attacker. Trained on external or self-generated harmful datasets, the cost value model could successfully influence the original safe LLM to output toxic content in decoding process. For instance, LLaMA-2-chat 7B outputs 39.18% concrete toxic content, along with only 22.16% refusals without any harmful suffixes. These potential weaknesses can then be exploited via prompt optimization such as soft prompts on images. We name this decoding strategy: Jailbreak Value Decoding (JVD), emphasizing that seemingly secure LLMs may not be as safe as we initially believe. They could be used to gather harmful data or launch covert attacks.

Power grid operators depend on accurate and reliable energy forecasts, aiming to minimize cases of extreme errors, as these outliers are particularly challenging to manage during operation. Incorporating planning information - such as known data about users' future behavior or scheduled events - has the potential to significantly enhance the accuracy and specificity of forecasts. Although there have been attempts to integrate such expected future behavior, these efforts consistently rely on conventional regression models to process this information. These models often lack the flexibility and capability to effectively incorporate both dynamic, forward-looking contextual inputs and historical data. To address this challenge, we conceptualize this combined forecasting and regression challenge as a sequence-to-sequence modeling problem and demonstrate, with three distinct models, that our contextually enhanced transformer models excel in this task. By leveraging schedule-based contextual information from the Swiss railway traction network, our proposed method significantly improved the average forecasting accuracy of nationwide railway energy consumption. Specifically, enhancing the transformer models with contextual information resulted in an average reduction of mean absolute error by 40.6%, whereas other state-of-the-art methods did not demonstrate any significant improvement. Despite extensive research efforts to forecast energy usage in electrical grids, operators still encounter significant outliers when faced with unexpected scenarios, with relative errors occasionally exceeding 50%, posing considerable operational challenges. Our research reveals that a critical limitation of current forecasting approaches is their over-reliance on trends and periodic patterns from past observations. We challenge this conventional focus on historical data as the primary source for energy forecasts and advocate for integrating contextual information about the expected future, such as anticipated user behavior and scheduled events. By incorporating this expected future information, we significantly improve the accuracy and specificity of load forecasts. This approach proves crucial for improving forecasting capabilities, and this methodology can be broadly applied to other domains where similar planning information is available. Electrical energy distinguishes itself from other traded commodities because its transmission follows the power flow equations Kundur (2012); Pagnier & Chertkov (2021). These equations require that the amount of energy consumed (demand) must always match the amount of energy produced (supply) to maintain the stability of the power grid Ullah et al. (2021). In compliance with these laws, grid operators collaborate closely with energy traders to ensure frequency synchronization and voltage stabilization, preventing damage to infrastructure and grid-connected assets Klyuev et al. (2022).

Machine learning for early syndrome diagnosis aims to solve the intricate task of predicting a ground truth label that most often is the outcome (effect) of a medical consensus definition applied to observed clinical measurements (causes), given clinical measurements observed several hours before. Instead of focusing on the prediction of the future effect, we propose to directly predict the causes via time series forecasting (TSF) of clinical variables and determine the effect by applying the gold standard consensus definition to the forecasted values. This method has the invaluable advantage of being straightforwardly interpretable to clinical practitioners, and because model training does not rely on a particular label anymore, the forecasted data can be used to predict any consensus-based label. We exemplify our method by means of long-term TSF with Transformer models, with a focus on accurate prediction of sparse clinical variables involved in the SOFA-based Sepsis-3 definition and the new Simplified Acute Physiology Score (SAPS-II) definition. Our experiments are conducted on two datasets and show that contrary to recent proposals which advocate set function encoders for time series and direct multi-step decoders, best results are achieved by a combination of standard dense encoders with iterative multi-step decoders. The key for success of iterative multi-step decoding can be attributed to its ability to capture cross-variate dependencies and to a student forcing training strategy that teaches the model to rely on its own previous time step predictions for the next time step prediction.

The ability to detect and adapt to changes in data distributions is crucial to maintain the accuracy and reliability of machine learning models. Detection is generally approached by observing the drift of model performance from a global point of view. However, drifts occurring in (fine-grained) data subgroups may go unnoticed when monitoring global drift. We take a different perspective, and introduce methods for observing drift at the finer granularity of subgroups. Relevant data subgroups are identified during training and monitored efficiently throughout the model's life. Performance drifts in any subgroup are detected, quantified and characterized so as to provide an interpretable summary of the model behavior over time. Experimental results confirm that our subgroup-level drift analysis identifies drifts that do not show at the (coarser) global dataset level. The proposed approach provides a valuable tool for monitoring model performance in dynamic real-world applications, offering insights into the evolving nature of data and ultimately contributing to more robust and adaptive models.

Large Language Models (LLMs) are prone to inheriting and amplifying societal biases embedded within their training data, potentially reinforcing harmful stereotypes related to gender, occupation, and other sensitive categories. This issue becomes particularly problematic as biased LLMs can have far-reaching consequences, leading to unfair practices and exacerbating social inequalities across various domains, such as recruitment, online content moderation, or even the criminal justice system. Although prior research has focused on detecting bias in LLMs using specialized datasets designed to highlight intrinsic biases, there has been a notable lack of investigation into how these findings correlate with authoritative datasets, such as those from the U.S. National Bureau of Labor Statistics (NBLS). To address this gap, we conduct empirical research that evaluates LLMs in a ``bias-out-of-the-box" setting, analyzing how the generated outputs compare with the distributions found in NBLS data. Furthermore, we propose a straightforward yet effective debiasing mechanism that directly incorporates NBLS instances to mitigate bias within LLMs. Our study spans seven different LLMs, including instructable, base, and mixture-of-expert models, and reveals significant levels of bias that are often overlooked by existing bias detection techniques. Importantly, our debiasing method, which does not rely on external datasets, demonstrates a substantial reduction in bias scores, highlighting the efficacy of our approach in creating fairer and more reliable LLMs.

Writing clinical notes and documenting medical exams is a critical task for healthcare professionals, serving as a vital component of patient care documentation. However, manually writing these notes is time-consuming and can impact the amount of time clinicians can spend on direct patient interaction and other tasks. Consequently, the development of automated clinical note generation systems has emerged as a clinically meaningful area of research within AI for health. In this paper, we present three key contributions to the field of clinical note generation using large language models (LLMs). First, we introduce CliniKnote, a comprehensive dataset consisting of 1,200 complex doctor-patient conversations paired with their full clinical notes. This dataset, created and curated by medical experts with the help of modern neural networks, provides a valuable resource for training and evaluating models in clinical note generation tasks. Second, we propose the K-SOAP (Keyword, Subjective, Objective, Assessment, and Plan) note format, which enhances traditional SOAP~\cite{podder2023soap} (Subjective, Objective, Assessment, and Plan) notes by adding a keyword section at the top, allowing for quick identification of essential information. Third, we develop an automatic pipeline to generate K-SOAP notes from doctor-patient conversations and benchmark various modern LLMs using various metrics. Our results demonstrate significant improvements in efficiency and performance compared to standard LLM finetuning methods.

The worldwide impact of the recent COVID-19 pandemic has been substantial, necessitating the development of accurate forecasting models to predict the spread and course of a pandemic. Previous methods for outbreak forecasting have faced limitations by not utilizing multiple sources of input and yielding suboptimal performance due to the limited availability of data. In this study, we propose a novel approach to address the challenges of infectious disease forecasting. We introduce a Multilateral Attention-enhanced GRU model that leverages information from multiple sources, thus enabling a comprehensive analysis of factors influencing the spread of a pandemic. By incorporating attention mechanisms within a GRU framework, our model can effectively capture complex relationships and temporal dependencies in the data, leading to improved forecasting performance. Further, we have curated a well-structured multi-source dataset for the recent COVID-19 pandemic that the research community can utilize as a great resource to conduct experiments and analysis on time-series forecasting. We evaluated the proposed model on our COVID-19 dataset and reported the output in terms of RMSE and MAE. The experimental results provide evidence that our proposed model surpasses existing techniques in terms of performance. We also performed performance gain and qualitative analysis on our dataset to evaluate the impact of the attention mechanism and show that the proposed model closely follows the trajectory of the pandemic.

The Channel Quality Indicator (CQI) plays a pivotal role in 5G networks, optimizing infrastructure dynamically to ensure high Quality of Service (QoS). Recent research has focused on improving CQI estimation in 5G networks using machine learning. In this field, the selection of the proper loss function is critical for training an accurate model. Two commonly used loss functions are Mean Squared Error (MSE) and Mean Absolute Error (MAE). Roughly speaking, MSE put more weight on outliers, MAE on the majority. Here, we argue that the Huber loss function is more suitable for CQI prediction, since it combines the benefits of both MSE and MAE. To achieve this, the Huber loss transitions smoothly between MSE and MAE, controlled by a user-defined hyperparameter called delta. However, finding the right balance between sensitivity to small errors (MAE) and robustness to outliers (MSE) by manually choosing the optimal delta is challenging. To address this issue, we propose a novel loss function, named Residual-based Adaptive Huber Loss (RAHL). In RAHL, a learnable residual is added to the delta, enabling the model to adapt based on the distribution of errors in the data. Our approach effectively balances model robustness against outliers while preserving inlier data precision. The widely recognized Long Short-Term Memory (LSTM) model is employed in conjunction with RAHL, showcasing significantly improved results compared to the aforementioned loss functions. The obtained results affirm the superiority of RAHL, offering a promising avenue for enhanced CQI prediction in 5G networks.

With the development of artificial intelligence, particularly the success of Large Language Models (LLMs), the quantity and quality of automatically generated stories have significantly increased. This has led to the need for automatic story evaluation to assess the generative capabilities of computing systems and analyze the quality of both automatic-generated and human-written stories. Evaluating a story can be more challenging than other generation evaluation tasks. While tasks like machine translation primarily focus on assessing the aspects of fluency and accuracy, story evaluation demands complex additional measures such as overall coherence, character development, interestingness, etc. This requires a thorough review of relevant research. In this survey, we first summarize existing storytelling tasks, including text-to-text, visual-to-text, and text-to-visual. We highlight their evaluation challenges, identify various human criteria to measure stories, and present existing benchmark datasets. Then, we propose a taxonomy to organize evaluation metrics that have been developed or can be adopted for story evaluation. We also provide descriptions of these metrics, along with the discussion of their merits and limitations. Later, we discuss the human-AI collaboration for story evaluation and generation. Finally, we suggest potential future research directions, extending from story evaluation to general evaluations.