分享

Neurosymbolic approaches blend the effectiveness of symbolic reasoning with the flexibility of neural networks. In this work, we propose a neurosymbolic architecture for generating SQL queries that builds and explores a solution tree using Best-First Search, with the possibility of backtracking. For this purpose, it integrates a Language Model (LM) with symbolic modules that help catch and correct errors made by the LM on SQL queries, as well as guiding the exploration of the solution tree. We focus on improving the performance of smaller open-source LMs, and we find that our tool, Xander, increases accuracy by an average of 10.9% and reduces runtime by an average of 28% compared to the LM without Xander, enabling a smaller LM (with Xander) to outperform its four-times larger counterpart (without Xander).

This research seeks to obviate the need for creating QA datasets and grading (chatbot) LLM responses when comparing LLMs' knowledge in specific topic domains. This is done in an entirely end-user centric way without need for access to any inner workings of the LLM, so long as it can be prompted and given a random seed to create different generations to the same prompt. The paper does this by, for a given topic domain, defining the "response dispersion" of an LLM by repeatedly asking an LLM the same opinion question about that topic domain. Namely, the response dispersion is the count of singular values needed to explain 95% of the variance in the embedding matrix of the LLM's responses. It is found that the response dispersion is inversely correlated with accuracy on relevant QA evaluations (average spearman rank correlation stronger than -.59). A use-case analysis shows that when comparing two different LLMs on the same topic domain, comparing their response dispersion is a suitable replacement for comparing their QA accuracy between 74% and 89% of the time, the range depending on certain reasonable accuracy-difference tolerances that may be acceptable to an end-user in exchange for the labor being saved using response dispersion instead of QA accuracy for comparison. Two response embeddings are studied for creating the embedding matrix in this study, one is from OpenAI's APIs and one is a novel embedding, here named reference sentence similarity embeddings, that can be computed locally and performs very nearly as well in calculating response dispersion. Also in this research, a pre-existing dataset called the IRC-Wiki Trivia dataset, originally developed for trivia games, has been re-purposed, curated, and the curation, called IRC-WikiTriviaQA, is made available for the purpose of this research.

A keyword spotting (KWS) engine that is continuously running on device is exposed to various speech signals that are usually unseen before. It is a challenging problem to build a small-footprint and high-performing KWS model with robustness under different acoustic environments. In this paper, we explore how to effectively apply adversarial examples to improve KWS robustness. We propose datasource-aware disentangled learning with adversarial examples to reduce the mismatch between the original and adversarial data as well as the mismatch across original training datasources. The KWS model architecture is based on depth-wise separable convolution and a simple attention module. Experimental results demonstrate that the proposed learning strategy improves false reject rate by $40.31%$ at $1%$ false accept rate on the internal dataset, compared to the strongest baseline without using adversarial examples. Our best-performing system achieves $98.06%$ accuracy on the Google Speech Commands V1 dataset.

Advances in automatic speaker verification (ASV) promote research into the formulation of spoofing detection systems for real-world applications. The performance of ASV systems can be degraded severely by multiple types of spoofing attacks, namely, synthetic speech (SS), voice conversion (VC), replay, twins and impersonation, especially in the case of unseen synthetic spoofing attacks. A reliable and robust spoofing detection system can act as a security gate to filter out spoofing attacks instead of having them reach the ASV system. A weighted additive angular margin loss is proposed to address the data imbalance issue, and different margins has been assigned to improve generalization to unseen spoofing attacks in this study. Meanwhile, we incorporate a meta-learning loss function to optimize differences between the embeddings of support versus query set in order to learn a spoofing-category-independent embedding space for utterances. Furthermore, we craft adversarial examples by adding imperceptible perturbations to spoofing speech as a data augmentation strategy, then we use an auxiliary batch normalization (BN) to guarantee that corresponding normalization statistics are performed exclusively on the adversarial examples. Additionally, A simple attention module is integrated into the residual block to refine the feature extraction process. Evaluation results on the Logical Access (LA) track of the ASVspoof 2019 corpus provides confirmation of our proposed approaches' effectiveness in terms of a pooled EER of 0.87%, and a min t-DCF of 0.0277. These advancements offer effective options to reduce the impact of spoofing attacks on voice recognition/authentication systems.

Hand gesture-based sign language recognition (SLR) is one of the most advanced applications of machine learning, and computer vision uses hand gestures. Although, in the past few years, many researchers have widely explored and studied how to address BSL problems, specific unaddressed issues remain, such as skeleton and transformer-based BSL recognition. In addition, the lack of evaluation of the BSL model in various concealed environmental conditions can prove the generalized property of the existing model by facing daily life signs. As a consequence, existing BSL recognition systems provide a limited perspective of their generalisation ability as they are tested on datasets containing few BSL alphabets that have a wide disparity in gestures and are easy to differentiate. To overcome these limitations, we propose a spatial-temporal attention-based BSL recognition model considering hand joint skeletons extracted from the sequence of images. The main aim of utilising hand skeleton-based BSL data is to ensure the privacy and low-resolution sequence of images, which need minimum computational cost and low hardware configurations. Our model captures discriminative structural displacements and short-range dependency based on unified joint features projected onto high-dimensional feature space. Specifically, the use of Separable TCN combined with a powerful multi-head spatial-temporal attention architecture generated high-performance accuracy. The extensive experiments with a proposed dataset and two benchmark BSL datasets with a wide range of evaluations, such as intra- and inter-dataset evaluation settings, demonstrated that our proposed models achieve competitive performance with extremely low computational complexity and run faster than existing models.

The Lunar landing has drawn great interest in lunar exploration in recent years, and autonomous lunar landing navigation is fundamental to this task. AI is expected to play a critical role in autonomous and intelligent space missions, yet human experts question the reliability of AI solutions. Thus, the \gls{xai} for vision-based lunar landing is studied in this paper, aiming at providing transparent and understandable predictions for intelligent lunar landing. Attention-based Darknet53 is proposed as the feature extraction structure. For crater detection and navigation tasks, attention-based YOLOv3 and attention-Darknet53-LSTM are presented respectively. The experimental results show that the offered networks provide competitive performance on relative crater detection and pose estimation during the lunar landing. The explainability of the provided networks is achieved by introducing an attention mechanism into the network during model building. Moreover, the PCC is utilised to quantitively evaluate the explainability of the proposed networks, with the findings showing the functions of various convolutional layers in the network.

Our algorithm GNN: Graph Neural Network and Large Language Model for Data Discovery inherit the benefits of \cite{hoang2024plod} (PLOD: Predictive Learning Optimal Data Discovery), \cite{Hoang2024BODBO} (BOD: Blindly Optimal Data Discovery) in terms of overcoming the challenges of having to predefine utility function and the human input for attribute ranking, which helps prevent the time-consuming loop process. In addition to these previous works, our algorithm GNN leverages the advantages of graph neural networks and large language models to understand text type values that cannot be understood by PLOD and MOD, thus making the task of predicting outcomes more reliable. GNN could be seen as an extension of PLOD in terms of understanding the text type value and the user's preferences, not only numerical values but also text values, making the promise of data science and analytics purposes.

Hallucinations in large language models (LLMs) pose significant challenges in tasks requiring complex multi-step reasoning, such as mathematical problem-solving. Existing approaches primarily detect the presence of hallucinations but lack a nuanced understanding of their types and manifestations. In this paper, we first introduce a comprehensive taxonomy that categorizes the common hallucinations in mathematical reasoning task into six types: fabrication, factual inconsistency, context inconsistency, instruction inconsistency, logical inconsistency, and logical error. We then propose FG-PRM (Fine-Grained Process Reward Model), an augmented model designed to detect and mitigate hallucinations in a fine-grained, step-level manner. To address the limitations of manually labeling training data, we propose an automated method for generating fine-grained hallucination data using LLMs. By injecting hallucinations into reasoning steps of correct solutions, we create a diverse and balanced synthetic dataset for training FG-PRM, which consists of six specialized Process Reward Models (PRMs), each tailored to detect a specific hallucination type. Our FG-PRM demonstrates superior performance across two key tasks: 1) Fine-grained hallucination detection: classifying hallucination types for each reasoning step; and 2) Verification: ranking multiple LLM-generated outputs to select the most accurate solution, mitigating reasoning hallucinations. Our experiments show that FG-PRM outperforms ChatGPT-3.5 and Claude-3 on fine-grained hallucination detection and substantially boosts the performance of LLMs on GSM8K and MATH benchmarks.

The performance of Large Language Models (LLMs) is substantially influenced by the pretraining corpus, which consists of vast quantities of unsupervised data processed by the models. Despite its critical role in model performance, ensuring the quality of this data is challenging due to its sheer volume and the absence of sample-level quality annotations and enhancements. In this paper, we introduce DecorateLM, a data engineering method designed to refine the pretraining corpus through data rating, tagging and editing. Specifically, DecorateLM rates texts against quality criteria, tags texts with hierarchical labels, and edits texts into a more formalized format. Due to the massive size of the pretraining corpus, adopting an LLM for decorating the entire corpus is less efficient. Therefore, to balance performance with efficiency, we curate a meticulously annotated training corpus for DecorateLM using a large language model and distill data engineering expertise into a compact 1.2 billion parameter small language model (SLM). We then apply DecorateLM to enhance 100 billion tokens of the training corpus, selecting 45 billion tokens that exemplify high quality and diversity for the further training of another 1.2 billion parameter LLM. Our results demonstrate that employing such high-quality data can significantly boost model performance, showcasing a powerful approach to enhance the quality of the pretraining corpus.

In this technical report, we present Falcon Mamba 7B, a new base large language model based on the novel Mamba architecture. Falcon Mamba 7B is trained on 5.8 trillion tokens with carefully selected data mixtures. As a pure Mamba-based model, Falcon Mamba 7B surpasses leading open-weight models based on Transformers, such as Mistral 7B, Llama3.1 8B, and Falcon2 11B. It is on par with Gemma 7B and outperforms models with different architecture designs, such as RecurrentGemma 9B and RWKV-v6 Finch 7B/14B. Currently, Falcon Mamba 7B is the best-performing Mamba model in the literature at this scale, surpassing both existing Mamba and hybrid Mamba-Transformer models, according to the Open LLM Leaderboard. Due to its architecture, Falcon Mamba 7B is significantly faster at inference and requires substantially less memory for long sequence generation. Despite recent studies suggesting that hybrid Mamba-Transformer models outperform pure architecture designs, we demonstrate that even the pure Mamba design can achieve similar, or even superior results compared to the Transformer and hybrid designs. We make the weights of our implementation of Falcon Mamba 7B publicly available on https://huggingface.co/tiiuae/falcon-mamba-7b, under a permissive license.

One of the most challenging problems facing NLP today is evaluation. Some of the most pressing issues pertain to benchmark saturation, data contamination, and diversity in the quality of test examples. To address these concerns, we propose Selection Methodology for Accurate, Reduced, and Targeted (SMART) filtering, a novel approach to select a high-quality subset of examples from existing benchmark datasets by systematically removing less informative and less challenging examples. Our approach applies three filtering criteria, removing (i) easy examples, (ii) data-contaminated examples, and (iii) examples that are similar to each other based on distance in an embedding space. We demonstrate the effectiveness of SMART on three multiple choice QA datasets, where our methodology increases efficiency by reducing dataset size by 48\% on average, while increasing Pearson correlation with rankings from ChatBot Arena, a more open-ended human evaluation setting. Our method enables us to be more efficient, whether using SMART to make new benchmarks more challenging or to revitalize older datasets, while still preserving the relative model rankings.

Generating high-quality, in-depth textual documents, such as academic papers, news articles, Wikipedia entries, and books, remains a significant challenge for Large Language Models (LLMs). In this paper, we propose to use planning to generate long form content. To achieve our goal, we generate intermediate steps via an auxiliary task that teaches the LLM to plan, reason and structure before generating the final text. Our main novelty lies in a single auxiliary task that does not require multiple rounds of prompting or planning. To overcome the scarcity of training data for these intermediate steps, we leverage LLMs to generate synthetic intermediate writing data such as outlines, key information and summaries from existing full articles. Our experiments demonstrate on two datasets from different domains, namely the scientific news dataset SciNews and Wikipedia datasets in KILT-Wiki and FreshWiki, that LLMs fine-tuned with the auxiliary task generate higher quality documents. We observed +2.5% improvement in ROUGE-Lsum, and a strong 3.60 overall win/loss ratio via human SxS evaluation, with clear wins in organization, relevance, and verifiability.