Summarizing book-length documents (>100K tokens) that exceed the context window size of large language models (LLMs) requires first breaking the input document into smaller chunks and then prompting an LLM to merge, update, and compress chunk-level summaries. Despite the complexity and importance of this task, it has yet to be meaningfully studied due to the challenges of evaluation: existing book-length summarization datasets (e.g., BookSum) are in the pretraining data of most public LLMs, and existing evaluation methods struggle to capture errors made by modern LLM summarizers. In this paper, we present the first study of the coherence of LLM-based book-length summarizers implemented via two prompting workflows: (1) hierarchically merging chunk-level summaries, and (2) incrementally updating a running summary. We obtain 1193 fine-grained human annotations on GPT-4 generated summaries of 100 recently-published books and identify eight common types of coherence errors made by LLMs. Because human evaluation is expensive and time-consuming, we develop an automatic metric, BooookScore, that measures the proportion of sentences in a summary that do not contain any of the identified error types. BooookScore has high agreement with human annotations and allows us to systematically evaluate the impact of many other critical parameters (e.g., chunk size, base LLM) while saving $15K USD and 500 hours in human evaluation costs. We find that closed-source LLMs such as GPT-4 and Claude 2 produce summaries with higher BooookScore than those generated by open-source models. While LLaMA 2 falls behind other models, Mixtral achieves performance on par with GPT-3.5-Turbo. Incremental updating yields lower BooookScore but higher level of detail than hierarchical merging, a trade-off sometimes preferred by annotators.
相關內容
大語言模型是基于海量文本數據訓練的深度學習模型。它不僅能夠生成自然語言文本，還能夠深入理解文本含義，處理各種自然語言任務，如文本摘要、問答、翻譯等。2023年，大語言模型及其在人工智能領域的應用已成為全球科技研究的熱點，其在規模上的增長尤為引人注目，參數量已從最初的十幾億躍升到如今的一萬億。參數量的提升使得模型能夠更加精細地捕捉人類語言微妙之處，更加深入地理解人類語言的復雜性。在過去的一年里，大語言模型在吸納新知識、分解復雜任務以及圖文對齊等多方面都有顯著提升。隨著技術的不斷成熟，它將不斷拓展其應用范圍，為人類提供更加智能化和個性化的服務，進一步改善人們的生活和生產方式。
Large language models (LLMs) increasingly serve as the backbone for classifying text associated with distinct domains and simultaneously several labels (classes). When encountering domain shifts, e.g., classifier of movie reviews from IMDb to Rotten Tomatoes, adapting such an LLM-based multi-label classifier is challenging due to incomplete label sets at the target domain and daunting training overhead. The existing domain adaptation methods address either image multi-label classifiers or text binary classifiers. In this paper, we design DALLMi, Domain Adaptation Large Language Model interpolator, a first-of-its-kind semi-supervised domain adaptation method for text data models based on LLMs, specifically BERT. The core of DALLMi is the novel variation loss and MixUp regularization, which jointly leverage the limited positively labeled and large quantity of unlabeled text and, importantly, their interpolation from the BERT word embeddings. DALLMi also introduces a label-balanced sampling strategy to overcome the imbalance between labeled and unlabeled data. We evaluate DALLMi against the partial-supervised and unsupervised approach on three datasets under different scenarios of label availability for the target domain. Our results show that DALLMi achieves higher mAP than unsupervised and partially-supervised approaches by 19.9% and 52.2%, respectively.
Likelihood-free inference methods based on neural conditional density estimation were shown to drastically reduce the simulation burden in comparison to classical methods such as ABC. When applied in the context of any latent variable model, such as a Hidden Markov model (HMM), these methods are designed to only estimate the parameters, rather than the joint distribution of the parameters and the hidden states. Naive application of these methods to a HMM, ignoring the inference of this joint posterior distribution, will thus produce an inaccurate estimate of the posterior predictive distribution, in turn hampering the assessment of goodness-of-fit. To rectify this problem, we propose a novel, sample-efficient likelihood-free method for estimating the high-dimensional hidden states of an implicit HMM. Our approach relies on learning directly the intractable posterior distribution of the hidden states, using an autoregressive-flow, by exploiting the Markov property. Upon evaluating our approach on some implicit HMMs, we found that the quality of the estimates retrieved using our method is comparable to what can be achieved using a much more computationally expensive SMC algorithm.
We generalize McDiarmid's inequality for functions with bounded differences on a high probability set, using an extension argument. Those functions concentrate around their conditional expectations. We further extend the results to concentration in general metric spaces.
Watermarking of language model outputs enables statistical detection of model-generated text, which can mitigate harms and misuses of language models. Existing watermarking strategies operate by altering the decoder of an existing language model. In this paper, we ask whether language models can directly learn to generate watermarked text, which would have significant implications for the real-world deployment of watermarks. First, learned watermarks could be used to build open models that naturally generate watermarked text, enabling watermarking for open models, where users can control the decoding procedure. Second, if watermarking is used to determine the provenance of generated text, an adversary can hurt the reputation of a victim model by spoofing its watermark and generating damaging watermarked text. To investigate the learnability of watermarks, we propose watermark distillation, which trains a student model to behave like a teacher model that uses decoding-based watermarking. We test our approach on three decoding-based watermarking strategies and various hyperparameter settings, finding that models can learn to generate watermarked text with high detectability. We also find limitations to learnability, including the loss of watermarking capabilities under fine-tuning on normal text and high sample complexity when learning low-distortion watermarks.
As the capabilities of large language models (LLMs) have expanded dramatically, aligning these models with human values presents a significant challenge, posing potential risks during deployment. Traditional alignment strategies rely heavily on human intervention, such as Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), or on the self-alignment capacities of LLMs, which usually require a strong LLM's emergent ability to improve its original bad answer. To address these challenges, we propose a novel self-alignment method that utilizes a Chain of Thought (CoT) approach, termed AlignCoT. This method encompasses stages of Question Analysis, Answer Guidance, and Safe Answer production. It is designed to enable LLMs to generate high-quality, safe responses throughout various stages of their development. Furthermore, we introduce the Mixture of insighTful Experts (MoTE) architecture, which applies the mixture of experts to enhance each component of the AlignCoT process, markedly increasing alignment efficiency. The MoTE approach not only outperforms existing methods in aligning LLMs with human values but also highlights the benefits of using self-generated data, revealing the dual benefits of improved alignment and training efficiency.
We develop an inferential toolkit for analyzing object-valued responses, which correspond to data situated in general metric spaces, paired with Euclidean predictors within the conformal framework. To this end we introduce conditional profile average transport costs, where we compare distance profiles that correspond to one-dimensional distributions of probability mass falling into balls of increasing radius through the optimal transport cost when moving from one distance profile to another. The average transport cost to transport a given distance profile to all others is crucial for statistical inference in metric spaces and underpins the proposed conditional profile scores. A key feature of the proposed approach is to utilize the distribution of conditional profile average transport costs as conformity score for general metric space-valued responses, which facilitates the construction of prediction sets by the split conformal algorithm. We derive the uniform convergence rate of the proposed conformity score estimators and establish asymptotic conditional validity for the prediction sets. The finite sample performance for synthetic data in various metric spaces demonstrates that the proposed conditional profile score outperforms existing methods in terms of both coverage level and size of the resulting prediction sets, even in the special case of scalar and thus Euclidean responses. We also demonstrate the practical utility of conditional profile scores for network data from New York taxi trips and for compositional data reflecting energy sourcing of U.S. states.
The Hierarchy Of Time-Surfaces (HOTS) algorithm, a neuromorphic approach for feature extraction from event data, presents promising capabilities but faces challenges in accuracy and compatibility with neuromorphic hardware. In this paper, we introduce Sup3r, a Semi-Supervised algorithm aimed at addressing these challenges. Sup3r enhances sparsity, stability, and separability in the HOTS networks. It enables end-to-end online training of HOTS networks replacing external classifiers, by leveraging semi-supervised learning. Sup3r learns class-informative patterns, mitigates confounding features, and reduces the number of processed events. Moreover, Sup3r facilitates continual and incremental learning, allowing adaptation to data distribution shifts and learning new tasks without forgetting. Preliminary results on N-MNIST demonstrate that Sup3r achieves comparable accuracy to similarly sized Artificial Neural Networks trained with back-propagation. This work showcases the potential of Sup3r to advance the capabilities of HOTS networks, offering a promising avenue for neuromorphic algorithms in real-world applications.
Purpose: Radiologists are tasked with visually scrutinizing large amounts of data produced by 3D volumetric imaging modalities. Small signals can go unnoticed during the 3d search because they are hard to detect in the visual periphery. Recent advances in machine learning and computer vision have led to effective computer-aided detection (CADe) support systems with the potential to mitigate perceptual errors. Approach: Sixteen non-expert observers searched through digital breast tomosynthesis (DBT) phantoms and single cross-sectional slices of the DBT phantoms. The 3D/2D searches occurred with and without a convolutional neural network (CNN)-based CADe support system. The model provided observers with bounding boxes superimposed on the image stimuli while they looked for a small microcalcification signal and a large mass signal. Eye gaze positions were recorded and correlated with changes in the area under the ROC curve (AUC). Results: The CNN-CADe improved the 3D search for the small microcalcification signal (delta AUC = 0.098, p = 0.0002) and the 2D search for the large mass signal (delta AUC = 0.076, p = 0.002). The CNN-CADe benefit in 3D for the small signal was markedly greater than in 2D (delta delta AUC = 0.066, p = 0.035). Analysis of individual differences suggests that those who explored the least with eye movements benefited the most from the CNN-CADe (r = -0.528, p = 0.036). However, for the large signal, the 2D benefit was not significantly greater than the 3D benefit (delta delta AUC = 0.033, p = 0.133). Conclusion: The CNN-CADe brings unique performance benefits to the 3D (vs. 2D) search of small signals by reducing errors caused by the under-exploration of the volumetric data.
State transition algorithm (STA) is a metaheuristic method for global optimization. Recently, a modified STA named parameter optimal state transition algorithm (POSTA) is proposed. In POSTA, the performance of expansion operator, rotation operator and axesion operator is optimized through a parameter selection mechanism. But due to the insufficient utilization of historical information, POSTA still suffers from slow convergence speed and low solution accuracy on specific problems. To make better use of the historical information, Nelder-Mead (NM) simplex search and quadratic interpolation (QI) are integrated into POSTA. The enhanced POSTA is tested against 14 benchmark functions with 20-D, 30-D and 50-D space. An experimental comparison with several competitive metaheuristic methods demonstrates the effectiveness of the proposed method.
Transformer-based pretrained language models (T-PTLMs) have achieved great success in almost every NLP task. The evolution of these models started with GPT and BERT. These models are built on the top of transformers, self-supervised learning and transfer learning. Transformed-based PTLMs learn universal language representations from large volumes of text data using self-supervised learning and transfer this knowledge to downstream tasks. These models provide good background knowledge to downstream tasks which avoids training of downstream models from scratch. In this comprehensive survey paper, we initially give a brief overview of self-supervised learning. Next, we explain various core concepts like pretraining, pretraining methods, pretraining tasks, embeddings and downstream adaptation methods. Next, we present a new taxonomy of T-PTLMs and then give brief overview of various benchmarks including both intrinsic and extrinsic. We present a summary of various useful libraries to work with T-PTLMs. Finally, we highlight some of the future research directions which will further improve these models. We strongly believe that this comprehensive survey paper will serve as a good reference to learn the core concepts as well as to stay updated with the recent happenings in T-PTLMs.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191