Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks but are constrained by their small context window sizes. Various efforts have been proposed to expand the context window to accommodate even up to 200K input tokens. Meanwhile, building high-quality benchmarks with much longer text lengths and more demanding tasks to provide comprehensive evaluations is of immense practical interest to facilitate long context understanding research of LLMs. However, prior benchmarks create datasets that ostensibly cater to long-text comprehension by expanding the input of traditional tasks, which falls short to exhibit the unique characteristics of long-text understanding, including long dependency tasks and longer text length compatible with modern LLMs' context window size. In this paper, we introduce a benchmark for extremely long context understanding with long-range dependencies, XL$^2$Bench, which includes three scenarios: Fiction Reading, Paper Reading, and Law Reading, and four tasks of increasing complexity: Memory Retrieval, Detailed Understanding, Overall Understanding, and Open-ended Generation, covering 27 subtasks in English and Chinese. It has an average length of 100K+ words (English) and 200K+ characters (Chinese). Evaluating six leading LLMs on XL$^2$Bench, we find that their performance significantly lags behind human levels. Moreover, the observed decline in performance across both the original and enhanced datasets underscores the efficacy of our approach to mitigating data contamination.

Reading comprehension continues to be a crucial research focus in the NLP community. Recent advances in Machine Reading Comprehension (MRC) have mostly centered on literal comprehension, referring to the surface-level understanding of content. In this work, we focus on the next level - interpretive comprehension, with a particular emphasis on inferring the themes of a narrative text. We introduce the first dataset specifically designed for interpretive comprehension of educational narratives, providing corresponding well-edited theme texts. The dataset spans a variety of genres and cultural origins and includes human-annotated theme keywords with varying levels of granularity. We further formulate NLP tasks under different abstractions of interpretive comprehension toward the main idea of a story. After conducting extensive experiments with state-of-the-art methods, we found the task to be both challenging and significant for NLP research.

The last two years have seen a rapid growth in concerns around the safety of large language models (LLMs). Researchers and practitioners have met these concerns by introducing an abundance of new datasets for evaluating and improving LLM safety. However, much of this work has happened in parallel, and with very different goals in mind, ranging from the mitigation of near-term risks around bias and toxic content generation to the assessment of longer-term catastrophic risk potential. This makes it difficult for researchers and practitioners to find the most relevant datasets for a given use case, and to identify gaps in dataset coverage that future work may fill. To remedy these issues, we conduct a first systematic review of open datasets for evaluating and improving LLM safety. We review 102 datasets, which we identified through an iterative and community-driven process over the course of several months. We highlight patterns and trends, such as a a trend towards fully synthetic datasets, as well as gaps in dataset coverage, such as a clear lack of non-English datasets. We also examine how LLM safety datasets are used in practice -- in LLM release publications and popular LLM benchmarks -- finding that current evaluation practices are highly idiosyncratic and make use of only a small fraction of available datasets. Our contributions are based on SafetyPrompts.com, a living catalogue of open datasets for LLM safety, which we commit to updating continuously as the field of LLM safety develops.

Scaling laws describe the relationship between the size of language models and their capabilities. Unlike prior studies that evaluate a model's capability via loss or benchmarks, we estimate the number of knowledge bits a model stores. We focus on factual knowledge represented as tuples, such as (USA, capital, Washington D.C.) from a Wikipedia page. Through multiple controlled datasets, we establish that language models can and only can store 2 bits of knowledge per parameter, even when quantized to int8, and such knowledge can be flexibly extracted for downstream applications. Consequently, a 7B model can store 14B bits of knowledge, surpassing the English Wikipedia and textbooks combined based on our estimation. More broadly, we present 12 results on how (1) training duration, (2) model architecture, (3) quantization, (4) sparsity constraints such as MoE, and (5) data signal-to-noise ratio affect a model's knowledge storage capacity. Notable insights include: * The GPT-2 architecture, with rotary embedding, matches or even surpasses LLaMA/Mistral architectures in knowledge storage, particularly over shorter training durations. This arises because LLaMA/Mistral uses GatedMLP, which is less stable and harder to train. * Prepending training data with domain names (e.g., wikipedia.org) significantly increases a model's knowledge capacity. Language models can autonomously identify and prioritize domains rich in knowledge, optimizing their storage capacity.

This paper studies retrieval-augmented approaches for personalizing large language models (LLMs), which potentially have a substantial impact on various applications and domains. We propose the first attempt to optimize the retrieval models that deliver a limited number of personal documents to large language models for the purpose of personalized generation. We develop two optimization algorithms that solicit feedback from the downstream personalized generation tasks for retrieval optimization -- one based on reinforcement learning whose reward function is defined using any arbitrary metric for personalized generation and another based on knowledge distillation from the downstream LLM to the retrieval model. This paper also introduces a pre- and post-generation retriever selection model that decides what retriever to choose for each LLM input. Extensive experiments on diverse tasks from the language model personalization (LaMP) benchmark reveal statistically significant improvements in six out of seven datasets.

Generating accurate step-by-step reasoning is essential for Large Language Models (LLMs) to address complex problems and enhance robustness and interpretability. Despite the flux of research on developing advanced reasoning approaches, systematically analyzing the diverse LLMs and reasoning strategies in generating reasoning chains remains a significant challenge. The difficulties stem from the lack of two key elements: (1) an automatic method for evaluating the generated reasoning chains on different tasks, and (2) a unified formalism and implementation of the diverse reasoning approaches for systematic comparison. This paper aims to close the gap: (1) We introduce AutoRace for fully automated reasoning chain evaluation. Existing metrics rely on expensive human annotations or pre-defined LLM prompts not adaptable to different tasks. In contrast, AutoRace automatically creates detailed evaluation criteria tailored for each task, and uses GPT-4 for accurate evaluation following the criteria. (2) We develop LLM Reasoners, a library for standardized modular implementation of existing and new reasoning algorithms, under a unified formulation of the search, reward, and world model components. With the new evaluation and library, (3) we conduct extensive study of different reasoning approaches (e.g., CoT, ToT, RAP). The analysis reveals interesting findings about different factors contributing to reasoning, including the reward-guidance, breadth-vs-depth in search, world model, and prompt formats, etc.

Transductive inference has been widely investigated in few-shot image classification, but completely overlooked in the recent, fast growing literature on adapting vision-langage models like CLIP. This paper addresses the transductive zero-shot and few-shot CLIP classification challenge, in which inference is performed jointly across a mini-batch of unlabeled query samples, rather than treating each instance independently. We initially construct informative vision-text probability features, leading to a classification problem on the unit simplex set. Inspired by Expectation-Maximization (EM), our optimization-based classification objective models the data probability distribution for each class using a Dirichlet law. The minimization problem is then tackled with a novel block Majorization-Minimization algorithm, which simultaneously estimates the distribution parameters and class assignments. Extensive numerical experiments on 11 datasets underscore the benefits and efficacy of our batch inference approach.On zero-shot tasks with test batches of 75 samples, our approach yields near 20% improvement in ImageNet accuracy over CLIP's zero-shot performance. Additionally, we outperform state-of-the-art methods in the few-shot setting. The code is available at: https://github.com/SegoleneMartin/transductive-CLIP.

As Large Language Models (LLMs) and generative AI become more widespread, the content safety risks associated with their use also increase. We find a notable deficiency in high-quality content safety datasets and benchmarks that comprehensively cover a wide range of critical safety areas. To address this, we define a broad content safety risk taxonomy, comprising 13 critical risk and 9 sparse risk categories. Additionally, we curate AEGISSAFETYDATASET, a new dataset of approximately 26, 000 human-LLM interaction instances, complete with human annotations adhering to the taxonomy. We plan to release this dataset to the community to further research and to help benchmark LLM models for safety. To demonstrate the effectiveness of the dataset, we instruction-tune multiple LLM-based safety models. We show that our models (named AEGISSAFETYEXPERTS), not only surpass or perform competitively with the state-of-the-art LLM-based safety models and general purpose LLMs, but also exhibit robustness across multiple jail-break attack categories. We also show how using AEGISSAFETYDATASET during the LLM alignment phase does not negatively impact the performance of the aligned models on MT Bench scores. Furthermore, we propose AEGIS, a novel application of a no-regret online adaptation framework with strong theoretical guarantees, to perform content moderation with an ensemble of LLM content safety experts in deployment

Instruction tuning has emerged as the key in aligning large language models (LLMs) with specific task instructions, thereby mitigating the discrepancy between the next-token prediction objective and users' actual goals. To reduce the labor and time cost to collect or annotate data by humans, researchers start to explore the use of LLMs to generate instruction-aligned synthetic data. Recent works focus on generating diverse instructions and applying LLM to increase instruction complexity, often neglecting downstream use cases. It remains unclear how to tailor high-quality data to elicit better instruction-following abilities in different target instruction distributions and LLMs. To this end, we introduce CodecLM, a general framework for adaptively generating high-quality synthetic data for LLM alignment with different downstream instruction distributions and LLMs. Drawing on the Encode-Decode principles, we use LLMs as codecs to guide the data generation process. We first encode seed instructions into metadata, which are concise keywords generated on-the-fly to capture the target instruction distribution, and then decode metadata to create tailored instructions. We also introduce Self-Rubrics and Contrastive Filtering during decoding to tailor data-efficient samples. Extensive experiments on four open-domain instruction following benchmarks validate the effectiveness of CodecLM over the current state-of-the-arts.

In this work, we explicitly show that modern LLMs tend to generate correct facts first, then "drift away" and generate incorrect facts later: this was occasionally observed but never properly measured. We develop a semantic drift score that measures the degree of separation between correct and incorrect facts in generated texts and confirm our hypothesis when generating Wikipedia-style biographies. This correct-then-incorrect generation pattern suggests that factual accuracy can be improved by knowing when to stop generation. Therefore, we explore the trade-off between information quantity and factual accuracy for several early stopping methods and manage to improve factuality by a large margin. We further show that reranking with semantic similarity can further improve these results, both compared to the baseline and when combined with early stopping. Finally, we try calling external API to bring the model back to the right generation path, but do not get positive results. Overall, our methods generalize and can be applied to any long-form text generation to produce more reliable information, by balancing trade-offs between factual accuracy, information quantity and computational cost.