While Retrieval-Augmented Generation (RAG) has emerged as a promising paradigm for boosting large language models (LLMs) in knowledge-intensive tasks, it often overlooks the crucial aspect of text chunking within its workflow. This paper proposes the Meta-Chunking framework, which specifically enhances chunking quality through a dual strategy that identifies optimal segmentation points and preserves global information. Initially, breaking limitations of similarity-based chunking, we design two adaptive chunking techniques based on uncertainty, namely Perplexity Chunking and Margin Sampling Chunking, by utilizing the logical perception capabilities of LLMs. Given the inherent complexity across different texts, we integrate meta-chunk with dynamic merging, striking a balance between fine-grained and coarse-grained text chunking. Furthermore, we establish the global information compensation mechanism, encompassing a two-stage hierarchical summary generation process and a three-stage text chunk rewriting procedure focused on missing reflection, refinement, and completion. These components collectively strengthen the semantic integrity and contextual coherence of chunks. Extensive experiments demonstrate that Meta-Chunking effectively addresses challenges of the chunking task within the RAG system, providing LLMs with more logically coherent text chunks. Additionally, our methodology validates the feasibility of implementing high-quality chunking tasks with smaller-scale models, thereby eliminating the reliance on robust instruction-following capabilities.
As large language models (LLMs) continue to advance, ensuring their alignment with human values becomes increasingly critical. Traditional alignment methods heavily rely on human feedback to fine-tune models. With the emergence of superhuman models whose outputs may surpass human understanding, evaluating and aligning these models using human judgments poses significant challenges. To address the challenges, recent works use weak supervisors to elicit knowledge from much stronger models. However, there are important disanalogies between the empirical setup in the existing works and the genuine goal of alignment. We remark that existing works investigate the phenomenon of weak-to-strong generation in analogous setup (i.e., binary classification), rather than practical alignment-relevant tasks (e.g., safety). In this paper, we bridge this gap by extending weak-to-strong generation to the context of practical alignment. We empirically demonstrate the widespread phenomenon of weak-to-strong generation in three complicated alignment tasks: safety, toxicity, and legal reasoning}. Furthermore, we explore efficient strategies for improving alignment performance to enhance the quality of model outcomes. Lastly, we summarize and analyze the challenges and potential solutions in regard to specific alignment tasks, which we hope to catalyze the research progress on the topic of weak-to-strong generalization. Our code is released at https://github.com/yeruimeng/WTS.git.
In this paper, we propose the $\textit{geometric invariance hypothesis (GIH)}$, which argues that the input space curvature of a neural network remains invariant under transformation in certain architecture-dependent directions during training. We investigate a simple, non-linear binary classification problem residing on a plane in a high dimensional space and observe that$\unicode{x2014}$unlike MLPs$\unicode{x2014}$ResNets fail to generalize depending on the orientation of the plane. Motivated by this example, we define a neural network's $\textbf{average geometry}$ and $\textbf{average geometry evolution}$ as compact $\textit{architecture-dependent}$ summaries of the model's input-output geometry and its evolution during training. By investigating the average geometry evolution at initialization, we discover that the geometry of a neural network evolves according to the data covariance projected onto its average geometry. This means that the geometry only changes in a subset of the input space when the average geometry is low-rank, such as in ResNets. This causes an architecture-dependent invariance property in the input space curvature, which we dub GIH. Finally, we present extensive experimental results to observe the consequences of GIH and how it relates to generalization in neural networks.
Citation networks are critical in modern science, and predicting which previous papers (candidates) will a new paper (query) cite is a critical problem. However, the roles of a paper's citations vary significantly, ranging from foundational knowledge basis to superficial contexts. Distinguishing these roles requires a deeper understanding of the logical relationships among papers, beyond simple edges in citation networks. The emergence of LLMs with textual reasoning capabilities offers new possibilities for discerning these relationships, but there are two major challenges. First, in practice, a new paper may select its citations from gigantic existing papers, where the texts exceed the context length of LLMs. Second, logical relationships between papers are implicit, and directly prompting an LLM to predict citations may result in surface-level textual similarities rather than the deeper logical reasoning. In this paper, we introduce the novel concept of core citation, which identifies the critical references that go beyond superficial mentions. Thereby, we elevate the citation prediction task from a simple binary classification to distinguishing core citations from both superficial citations and non-citations. To address this, we propose $\textbf{HLM-Cite}$, a $\textbf{H}$ybrid $\textbf{L}$anguage $\textbf{M}$odel workflow for citation prediction, which combines embedding and generative LMs. We design a curriculum finetune procedure to adapt a pretrained text embedding model to coarsely retrieve high-likelihood core citations from vast candidates and then design an LLM agentic workflow to rank the retrieved papers through one-shot reasoning, revealing the implicit relationships among papers. With the pipeline, we can scale the candidate sets to 100K papers. We evaluate HLM-Cite across 19 scientific fields, demonstrating a 17.6% performance improvement comparing SOTA methods.
Receiver Operating Characteristic (ROC) curves are useful for evaluation in binary classification and changepoint detection, but difficult to use for learning since the Area Under the Curve (AUC) is piecewise constant (gradient zero almost everywhere). Recently the Area Under Min (AUM) of false positive and false negative rates has been proposed as a differentiable surrogate for AUC. In this paper we study the piecewise linear/constant nature of the AUM/AUC, and propose new efficient path-following algorithms for choosing the learning rate which is optimal for each step of gradient descent (line search), when optimizing a linear model. Remarkably, our proposed line search algorithm has the same log-linear asymptotic time complexity as gradient descent with constant step size, but it computes a complete representation of the AUM/AUC as a function of step size. In our empirical study of binary classification problems, we verify that our proposed algorithm is fast and exact; in changepoint detection problems we show that the proposed algorithm is just as accurate as grid search, but faster.
Numerous methods have been implemented that pursue fairness with respect to sensitive features by mitigating biases in machine learning. Yet, the problem settings that each method tackles vary significantly, including the stage of intervention, the composition of sensitive features, the fairness notion, and the distribution of the output. Even in binary classification, these subtle differences make it highly complicated to benchmark fairness methods, as their performance can strongly depend on exactly how the bias mitigation problem was originally framed. Hence, we introduce ABCFair, a benchmark approach which allows adapting to the desiderata of the real-world problem setting, enabling proper comparability between methods for any use case. We apply ABCFair to a range of pre-, in-, and postprocessing methods on both large-scale, traditional datasets and on a dual label (biased and unbiased) dataset to sidestep the fairness-accuracy trade-off.
Logic Tensor Networks (LTN) is a Neuro-Symbolic framework that effectively incorporates deep learning and logical reasoning. In particular, LTN allows defining a logical knowledge base and using it as the objective of a neural model. This makes learning by logical reasoning possible as the parameters of the model are optimized by minimizing a loss function composed of a set of logical formulas expressing facts about the learning task. The framework learns via gradient-descent optimization. Fuzzy logic, a relaxation of classical logic permitting continuous truth values in the interval [0,1], makes this learning possible. Specifically, the training of an LTN consists of three steps. Firstly, (1) the training data is used to ground the formulas. Then, (2) the formulas are evaluated, and the loss function is computed. Lastly, (3) the gradients are back-propagated through the logical computational graph, and the weights of the neural model are changed so the knowledge base is maximally satisfied. LTNtorch is the fully documented and tested PyTorch implementation of Logic Tensor Networks. This paper presents the formalization of LTN and how LTNtorch implements it. Moreover, it provides a basic binary classification example.
Hallucination, the generation of factually incorrect content, is a growing challenge in Large Language Models (LLMs). Existing detection and mitigation methods are often isolated and insufficient for domain-specific needs, lacking a standardized pipeline. This paper introduces THaMES (Tool for Hallucination Mitigations and EvaluationS), an integrated framework and library addressing this gap. THaMES offers an end-to-end solution for evaluating and mitigating hallucinations in LLMs, featuring automated test set generation, multifaceted benchmarking, and adaptable mitigation strategies. It automates test set creation from any corpus, ensuring high data quality, diversity, and cost-efficiency through techniques like batch processing, weighted sampling, and counterfactual validation. THaMES assesses a model's ability to detect and reduce hallucinations across various tasks, including text generation and binary classification, applying optimal mitigation strategies like In-Context Learning (ICL), Retrieval Augmented Generation (RAG), and Parameter-Efficient Fine-tuning (PEFT). Evaluations of state-of-the-art LLMs using a knowledge base of academic papers, political news, and Wikipedia reveal that commercial models like GPT-4o benefit more from RAG than ICL, while open-weight models like Llama-3.1-8B-Instruct and Mistral-Nemo gain more from ICL. Additionally, PEFT significantly enhances the performance of Llama-3.1-8B-Instruct in both evaluation tasks.
Anomaly detection is an important problem in many application areas, such as network security. Many deep learning methods for unsupervised anomaly detection produce good empirical performance but lack theoretical guarantees. By casting anomaly detection into a binary classification problem, we establish non-asymptotic upper bounds and a convergence rate on the excess risk on rectified linear unit (ReLU) neural networks trained on synthetic anomalies. Our convergence rate on the excess risk matches the minimax optimal rate in the literature. Furthermore, we provide lower and upper bounds on the number of synthetic anomalies that can attain this optimality. For practical implementation, we relax some conditions to improve the search for the empirical risk minimizer, which leads to competitive performance to other classification-based methods for anomaly detection. Overall, our work provides the first theoretical guarantees of unsupervised neural network-based anomaly detectors and empirical insights on how to design them well.
Tabular data is the most common type of data in real-life scenarios. In this study, we propose the TabKANet model for tabular data modeling, which targets the bottlenecks in learning from numerical content. We constructed a Kolmogorov-Arnold Network (KAN) based Numerical Embedding Module and unified numerical and categorical features encoding within a Transformer architecture. TabKANet has demonstrated stable and significantly superior performance compared to Neural Networks (NNs) across multiple public datasets in binary classification, multi-class classification, and regression tasks. Its performance is comparable to or surpasses that of Gradient Boosted Decision Tree models (GBDTs). Our code is publicly available on GitHub: https://github.com/AI-thpremed/TabKANet.
In recent years there have been a growing interest in online auditing of information flow over social networks with the goal of monitoring undesirable effects, such as, misinformation and fake news. Most previous work on the subject, focus on the binary classification problem of classifying information as fake or genuine. Nonetheless, in many practical scenarios, the multi-class/label setting is of particular importance. For example, it could be the case that a social media platform may want to distinguish between ``true", ``partly-true", and ``false" information. Accordingly, in this paper, we consider the problem of online multiclass classification of information flow. To that end, driven by empirical studies on information flow over real-world social media networks, we propose a probabilistic information flow model over graphs. Then, the learning task is to detect the label of the information flow, with the goal of minimizing a combination of the classification error and the detection time. For this problem, we propose two detection algorithms; the first is based on the well-known multiple sequential probability ratio test, while the second is a novel graph neural network based sequential decision algorithm. For both algorithms, we prove several strong statistical guarantees. We also construct a data driven algorithm for learning the proposed probabilistic model. Finally, we test our algorithms over two real-world datasets, and show that they outperform other state-of-the-art misinformation detection algorithms, in terms of detection time and classification error.
In practice, we are often faced with small-sized tabular data. However, current tabular benchmarks are not geared towards data-scarce applications, making it very difficult to derive meaningful conclusions from empirical comparisons. We introduce PMLBmini, a tabular benchmark suite of 44 binary classification datasets with sample sizes $\leq$ 500. We use our suite to thoroughly evaluate current automated machine learning (AutoML) frameworks, off-the-shelf tabular deep neural networks, as well as classical linear models in the low-data regime. Our analysis reveals that state-of-the-art AutoML and deep learning approaches often fail to appreciably outperform even a simple logistic regression baseline, but we also identify scenarios where AutoML and deep learning methods are indeed reasonable to apply. Our benchmark suite, available on https://github.com/RicardoKnauer/TabMini , allows researchers and practitioners to analyze their own methods and challenge their data efficiency.
In recent years, the multimedia forensics and security community has seen remarkable progress in multitask learning for DeepFake (i.e., face forgery) detection. The prevailing approach has been to frame DeepFake detection as a binary classification problem augmented by manipulation-oriented auxiliary tasks. This scheme focuses on learning features specific to face manipulations with limited generalizability. In this paper, we delve deeper into semantics-oriented multitask learning for DeepFake detection, capturing the relationships among face semantics via joint embedding. We first propose an automated dataset expansion technique that broadens current face forgery datasets to support semantics-oriented DeepFake detection tasks at both the global face attribute and local face region levels. Furthermore, we resort to the joint embedding of face images and labels (depicted by text descriptions) for prediction. This approach eliminates the need for manually setting task-agnostic and task-specific parameters, which is typically required when predicting multiple labels directly from images. In addition, we employ bi-level optimization to dynamically balance the fidelity loss weightings of various tasks, making the training process fully automated. Extensive experiments on six DeepFake datasets show that our method improves the generalizability of DeepFake detection and renders some degree of model interpretation by providing human-understandable explanations.
Although advancements in deep learning have significantly enhanced the recommendation accuracy of deep recommendation models, these methods still suffer from low recommendation efficiency. Recently proposed tree-based deep recommendation models alleviate the problem by directly learning tree structure and representations under the guidance of recommendation objectives. To guarantee the effectiveness of beam search for recommendation accuracy, these models strive to ensure that the tree adheres to the max-heap assumption, where a parent node's preference should be the maximum among its children's preferences. However, they employ a one-versus-all strategy, framing the training task as a series of independent binary classification objectives for each node, which limits their ability to fully satisfy the max-heap assumption. To this end, we propose a Deep Tree-based Retriever (DTR for short) for efficient recommendation. DTR frames the training task as a softmax-based multi-class classification over tree nodes at the same level, enabling explicit horizontal competition and more discriminative top-k selection among them, which mimics the beam search behavior during training. To mitigate the suboptimality induced by the labeling of non-leaf nodes, we propose a rectification method for the loss function, which further aligns with the max-heap assumption in expectation. As the number of tree nodes grows exponentially with the levels, we employ sampled softmax to approximate optimization and thereby enhance efficiency. Furthermore, we propose a tree-based sampling method to reduce the bias inherent in sampled softmax. Theoretical results reveal DTR's generalization capability, and both the rectification method and tree-based sampling contribute to improved generalization. The experiments are conducted on four real-world datasets, validating the effectiveness of the proposed method.
This study explores the concept of cross-disease transferability (XDT) in medical imaging, focusing on the potential of binary classifiers trained on one disease to perform zero-shot classification on another disease affecting the same organ. Utilizing chest X-rays (CXR) as the primary modality, we investigate whether a model trained on one pulmonary disease can make predictions about another novel pulmonary disease, a scenario with significant implications for medical settings with limited data on emerging diseases. The XDT framework leverages the embedding space of a vision encoder, which, through kernel transformation, aids in distinguishing between diseased and non-diseased classes in the latent space. This capability is especially beneficial in resource-limited environments or in regions with low prevalence of certain diseases, where conventional diagnostic practices may fail. However, the XDT framework is currently limited to binary classification, determining only the presence or absence of a disease rather than differentiating among multiple diseases. This limitation underscores the supplementary role of XDT to traditional diagnostic tests in clinical settings. Furthermore, results show that XDT-CXR as a framework is able to make better predictions compared to other zero-shot learning (ZSL) baselines.
Support Vector Machines (SVMs) based on hinge loss have been extensively discussed and applied to various binary classification tasks. These SVMs achieve a balance between margin maximization and the minimization of slack due to outliers. Although many efforts have been dedicated to enhancing the performance of SVMs with hinge loss, studies on $p$SVMs, soft-margin SVMs with $p$-norm hinge loss, remain relatively scarce. In this paper, we explore the properties, performance, and training algorithms of $p$SVMs. We first derive the generalization bound of $p$SVMs, then formulate the dual optimization problem, comparing it with the traditional approach. Furthermore, we discuss a generalized version of the Sequential Minimal Optimization (SMO) algorithm, $p$SMO, to train our $p$SVM model. Comparative experiments on various datasets, including binary and multi-class classification tasks, demonstrate the effectiveness and advantages of our $p$SVM model and the $p$SMO method. Code is available at https://github.com/CoderBak/pSVM.
Cardiovascular diseases are a leading cause of mortality worldwide, highlighting the need for accurate diagnostic methods. This study benchmarks centralized and federated machine learning algorithms for heart disease classification using the UCI dataset which includes 920 patient records from four hospitals in the USA, Hungary and Switzerland. Our benchmark is supported by Shapley-value interpretability analysis to quantify features' importance for classification. In the centralized setup, various binary classification algorithms are trained on pooled data, with a support vector machine (SVM) achieving the highest testing accuracy of 83.3\%, surpassing the established benchmark of 78.7\% with logistic regression. Additionally, federated learning algorithms with four clients (hospitals) are explored, leveraging the dataset's natural partition to enhance privacy without sacrificing accuracy. Federated SVM, an uncommon approach in the literature, achieves a top testing accuracy of 73.8\%. Our interpretability analysis aligns with existing medical knowledge of heart disease indicators. Overall, this study establishes a benchmark for efficient and interpretable pre-screening tools for heart disease while maintaining patients' privacy. This work is available at https://github.com/padillma1/Heart-Disease-Classification-on-UCI-dataset-and-Shapley-Interpretability-Analysis.
Being cautious is crucial for enhancing the trustworthiness of machine learning systems integrated into decision-making pipelines. Although calibrated probabilities help in optimal decision-making, perfect calibration remains unattainable, leading to estimates that fluctuate between under- and overconfidence. This becomes a critical issue in high-risk scenarios, where even occasional overestimation can lead to extreme expected costs. In these scenarios, it is important for each predicted probability to lean towards underconfidence, rather than just achieving an average balance. In this study, we introduce the novel concept of cautious calibration in binary classification. This approach aims to produce probability estimates that are intentionally underconfident for each predicted probability. We highlight the importance of this approach in a high-risk scenario and propose a theoretically grounded method for learning cautious calibration maps. Through experiments, we explore and compare our method to various approaches, including methods originally not devised for cautious calibration but applicable in this context. We show that our approach is the most consistent in providing cautious estimates. Our work establishes a strong baseline for further developments in this novel framework.
For centuries, writers have hidden messages in their texts as acrostics, where initial letters of consecutive lines or paragraphs form meaningful words or phrases. Scholars searching for acrostics manually can only focus on a few authors at a time and often favor qualitative arguments in discussing intentionally. We aim to put the study of acrostics on firmer statistical footing by presenting AcrosticSleuth, a first-of-its-kind tool that automatically identifies acrostics and ranks them by the probability that the sequence of characters does not occur by chance (and therefore may have been inserted intentionally). Acrostics are rare, so we formalize the problem as a binary classification task in the presence of extreme class imbalance. To evaluate AcrosticSleuth, we present the Acrostic Identification Dataset (AcrostID), a collection of acrostics from the WikiSource online database. Despite the class imbalance, AcrosticSleuth achieves F1 scores of 0.39, 0.59, and 0.66 on French, English, and Russian subdomains of WikiSource, respectively. We further demonstrate that AcrosticSleuth can identify previously unknown high-profile instances of wordplay, such as the acrostic spelling ARSPOETICA (``art of poetry") by Italian Humanist Albertino Mussato and English philosopher Thomas Hobbes' signature in the opening paragraphs of The Elements of Law.
The ease of access to large language models (LLMs) has enabled a widespread of machine-generated texts, and now it is often hard to tell whether a piece of text was human-written or machine-generated. This raises concerns about potential misuse, particularly within educational and academic domains. Thus, it is important to develop practical systems that can automate the process. Here, we present one such system, LLM-DetectAIve, designed for fine-grained detection. Unlike most previous work on machine-generated text detection, which focused on binary classification, LLM-DetectAIve supports four categories: (i) human-written, (ii) machine-generated, (iii) machine-written, then machine-humanized, and (iv) human-written, then machine-polished. Category (iii) aims to detect attempts to obfuscate the fact that a text was machine-generated, while category (iv) looks for cases where the LLM was used to polish a human-written text, which is typically acceptable in academic writing, but not in education. Our experiments show that LLM-DetectAIve can effectively identify the above four categories, which makes it a potentially useful tool in education, academia, and other domains. LLM-DetectAIve is publicly accessible at https://github.com/mbzuai-nlp/LLM-DetectAIve. The video describing our system is available at https://youtu.be/E8eT_bE7k8c.