Retrieval augmented generation (RAG) combines the generative abilities of large language models (LLMs) with external knowledge sources to provide more accurate and up-to-date responses. Recent RAG advancements focus on improving retrieval outcomes through iterative LLM refinement or self-critique capabilities acquired through additional instruction tuning of LLMs. In this work, we introduce Speculative RAG - a framework that leverages a larger generalist LM to efficiently verify multiple RAG drafts produced in parallel by a smaller, distilled specialist LM. Each draft is generated from a distinct subset of retrieved documents, offering diverse perspectives on the evidence while reducing input token counts per draft. This approach enhances comprehension of each subset and mitigates potential position bias over long context. Our method accelerates RAG by delegating drafting to the smaller specialist LM, with the larger generalist LM performing a single verification pass over the drafts. Extensive experiments demonstrate that Speculative RAG achieves state-of-the-art performance with reduced latency on TriviaQA, MuSiQue, PubHealth, and ARC-Challenge benchmarks. It notably enhances accuracy by up to 12.97% while reducing latency by 51% compared to conventional RAG systems on PubHealth.

With the rapid development of the large model domain, research related to fine-tuning has concurrently seen significant advancement, given that fine-tuning is a constituent part of the training process for large-scale models. Data engineering plays a fundamental role in the training process of models, which includes data infrastructure, data processing, etc. Data during fine-tuning likewise forms the base for large models. In order to embrace the power and explore new possibilities of fine-tuning datasets, this paper reviews current public fine-tuning datasets from the perspective of data construction. An overview of public fine-tuning datasets from two sides: evolution and taxonomy, is provided in this review, aiming to chart the development trajectory. Construction techniques and methods for public fine-tuning datasets of Large Language Models (LLMs), including data generation and data augmentation among others, are detailed. This elaboration follows the aforementioned taxonomy, specifically across demonstration, comparison, and generalist categories. Additionally, a category tree of data generation techniques has been abstracted in our review to assist researchers in gaining a deeper understanding of fine-tuning datasets from the construction dimension. Our review also summarizes the construction features in different data preparation phases of current practices in this field, aiming to provide a comprehensive overview and inform future research. Fine-tuning dataset practices, encompassing various data modalities, are also discussed from a construction perspective in our review. Towards the end of the article, we offer insights and considerations regarding the future construction and developments of fine-tuning datasets.

As an important part of Music Information Retrieval (MIR), Symbolic Music Understanding (SMU) has gained substantial attention, as it can assist musicians and amateurs in learning and creating music. Recently, pre-trained language models have been widely adopted in SMU because the symbolic music shares a huge similarity with natural language, and the pre-trained manner also helps make full use of limited music data. However, the issue of bias, such as sexism, ageism, and racism, has been observed in pre-trained language models, which is attributed to the imbalanced distribution of training data. It also has a significant influence on the performance of downstream tasks, which also happens in SMU. To address this challenge, we propose Adversarial-MidiBERT, a symbolic music understanding model based on Bidirectional Encoder Representations from Transformers (BERT). We introduce an unbiased pre-training method based on adversarial learning to minimize the participation of tokens that lead to biases during training. Furthermore, we propose a mask fine-tuning method to narrow the data gap between pre-training and fine-tuning, which can help the model converge faster and perform better. We evaluate our method on four music understanding tasks, and our approach demonstrates excellent performance in all of them. The code for our model is publicly available at https://github.com/RS2002/Adversarial-MidiBERT.

As Artificial Intelligence (AI) tools are increasingly employed in diverse real-world applications, there has been significant interest in regulating these tools. To this end, several regulatory frameworks have been introduced by different countries worldwide. For example, the European Union recently passed the AI Act, the White House issued an Executive Order on safe, secure, and trustworthy AI, and the White House Office of Science and Technology Policy issued the Blueprint for an AI Bill of Rights (AI BoR). Many of these frameworks emphasize the need for auditing and improving the trustworthiness of AI tools, underscoring the importance of safety, privacy, explainability, fairness, and human fallback options. Although these regulatory frameworks highlight the necessity of enforcement, practitioners often lack detailed guidance on implementing them. Furthermore, the extensive research on operationalizing each of these aspects is frequently buried in technical papers that are difficult for practitioners to parse. In this write-up, we address this shortcoming by providing an accessible overview of existing literature related to operationalizing regulatory principles. We provide easy-to-understand summaries of state-of-the-art literature and highlight various gaps that exist between regulatory guidelines and existing AI research, including the trade-offs that emerge during operationalization. We hope that this work not only serves as a starting point for practitioners interested in learning more about operationalizing the regulatory guidelines outlined in the Blueprint for an AI BoR but also provides researchers with a list of critical open problems and gaps between regulations and state-of-the-art AI research. Finally, we note that this is a working paper and we invite feedback in line with the purpose of this document as described in the introduction.

In this paper we argue that key, often sensational and misleading, claims regarding linguistic capabilities of Large Language Models (LLMs) are based on at least two unfounded assumptions; the assumption of language completeness and the assumption of data completeness. Language completeness assumes that a distinct and complete thing such as `a natural language' exists, the essential characteristics of which can be effectively and comprehensively modelled by an LLM. The assumption of data completeness relies on the belief that a language can be quantified and wholly captured by data. Work within the enactive approach to cognitive science makes clear that, rather than a distinct and complete thing, language is a means or way of acting. Languaging is not the kind of thing that can admit of a complete or comprehensive modelling. From an enactive perspective we identify three key characteristics of enacted language; embodiment, participation, and precariousness, that are absent in LLMs, and likely incompatible in principle with current architectures. We argue that these absences imply that LLMs are not now and cannot in their present form be linguistic agents the way humans are. We illustrate the point in particular through the phenomenon of `algospeak', a recently described pattern of high stakes human language activity in heavily controlled online environments. On the basis of these points, we conclude that sensational and misleading claims about LLM agency and capabilities emerge from a deep misconception of both what human language is and what LLMs are.

Cryptic crossword clues are challenging cognitive tasks, for which new test sets are released on a daily basis by multiple international newspapers. Each cryptic clue contains both the definition of the answer to be placed in the crossword grid (in common with regular crosswords), and `wordplay' that proves that the answer is correct (i.e. a human solver can be confident that an answer is correct without needing crossing words to confirm it). Using an existing cryptic wordplay proving framework (operating on Python proofs created by an LLM), we show that it is possible to distinguish between correct answers and almost-correct ones based upon whether the wordplay `works'.

Recently, various pre-trained language models (PLMs) have been proposed to prove their impressive performances on a wide range of few-shot tasks. However, limited by the unstructured prior knowledge in PLMs, it is difficult to maintain consistent performance on complex structured scenarios, such as hierarchical text classification (HTC), especially when the downstream data is extremely scarce. The main challenge is how to transfer the unstructured semantic space in PLMs to the downstream domain hierarchy. Unlike previous work on HTC which directly performs multi-label classification or uses graph neural network (GNN) to inject label hierarchy, in this work, we study the HTC problem under a few-shot setting to adapt knowledge in PLMs from an unstructured manner to the downstream hierarchy. Technically, we design a simple yet effective method named Hierarchical Iterative Conditional Random Field (HierICRF) to search the most domain-challenging directions and exquisitely crafts domain-hierarchy adaptation as a hierarchical iterative language modeling problem, and then it encourages the model to make hierarchical consistency self-correction during the inference, thereby achieving knowledge transfer with hierarchical consistency preservation. We perform HierICRF on various architectures, and extensive experiments on two popular HTC datasets demonstrate that prompt with HierICRF significantly boosts the few-shot HTC performance with an average Micro-F1 by 28.80% to 1.50% and Macro-F1 by 36.29% to 1.5% over the previous state-of-the-art (SOTA) baselines under few-shot settings, while remaining SOTA hierarchical consistency performance.

Large Language Models (LLMs) show promise for natural language generation in healthcare, but risk hallucinating factually incorrect information. Deploying LLMs for medical question answering necessitates reliable uncertainty estimation (UE) methods to detect hallucinations. In this work, we benchmark popular UE methods with different model sizes on medical question-answering datasets. Our results show that current approaches generally perform poorly in this domain, highlighting the challenge of UE for medical applications. We also observe that larger models tend to yield better results, suggesting a correlation between model size and the reliability of UE. To address these challenges, we propose Two-phase Verification, a probability-free Uncertainty Estimation approach. First, an LLM generates a step-by-step explanation alongside its initial answer, followed by formulating verification questions to check the factual claims in the explanation. The model then answers these questions twice: first independently, and then referencing the explanation. Inconsistencies between the two sets of answers measure the uncertainty in the original response. We evaluate our approach on three biomedical question-answering datasets using Llama 2 Chat models and compare it against the benchmarked baseline methods. The results show that our Two-phase Verification method achieves the best overall accuracy and stability across various datasets and model sizes, and its performance scales as the model size increases.

Current research on tool learning primarily focuses on selecting the most effective tool from a wide array of options, often overlooking cost-effectiveness, a crucial factor in human problem-solving. In this paper, we address the selection of homogeneous tools by predicting both their performance and the associated cost required to accomplish a given task. We then assign queries to the optimal tools in a cost-effective manner. Our experimental results demonstrate that our method achieves higher performance at a lower cost compared to strong baseline approaches.

Training large language models (LLMs) and multimodal LLMs necessitates significant computing resources, and existing publicly available LLMs are typically pre-trained on diverse, privately curated datasets spanning various tasks. For instance, LLaMA, Vicuna, and LLaVA are three LLM variants trained with LLaMA base models using very different training recipes, tasks, and data modalities. The training cost and complexity for such LLM variants grow rapidly. In this study, we propose to use a soup strategy to assemble these LLM variants into a single well-generalized multimodal LLM (SoupLM) in a cost-efficient manner. Assembling these LLM variants efficiently brings knowledge and specialities trained from different domains and data modalities into an integrated one (e.g., chatbot speciality from user-shared conversations for Vicuna, and visual capacity from vision-language data for LLaVA), therefore, to avoid computing costs of repetitive training on several different domains. We propose series of soup strategies to systematically benchmark performance gains across various configurations, and probe the soup behavior across base models in the interpolation space.