Automatic evaluation of retrieval augmented generation (RAG) systems relies on fine-grained dimensions like faithfulness and relevance, as judged by expert human annotators. Meta-evaluation benchmarks support the development of automatic evaluators that correlate well with human judgement. However, existing benchmarks predominantly focus on English or use translated data, which fails to capture cultural nuances. A native approach provides a better representation of the end user experience. In this work, we develop a Multilingual End-to-end Meta-Evaluation RAG benchmark (MEMERAG). Our benchmark builds on the popular MIRACL dataset, using native-language questions and generating responses with diverse large language models (LLMs), which are then assessed by expert annotators for faithfulness and relevance. We describe our annotation process and show that it achieves high inter-annotator agreement. We then analyse the performance of the answer-generating LLMs across languages as per the human evaluators. Finally we apply the dataset to our main use-case which is to benchmark multilingual automatic evaluators (LLM-as-a-judge). We show that our benchmark can reliably identify improvements offered by advanced prompting techniques and LLMs. Our dataset is available at https://github.com/amazon-science/MEMERAG
Glaucoma is a prevalent eye disease that progresses silently without symptoms. If not detected and treated early, it can cause permanent vision loss. Computer-assisted diagnosis systems play a crucial role in timely and efficient identification. This study introduces MaxGlaViT, a lightweight model based on the restructured Multi-Axis Vision Transformer (MaxViT) for early glaucoma detection. First, MaxViT was scaled to optimize block and channel numbers, resulting in a lighter architecture. Second, the stem was enhanced by adding attention mechanisms (CBAM, ECA, SE) after convolution layers to improve feature learning. Third, MBConv structures in MaxViT blocks were replaced by advanced DL blocks (ConvNeXt, ConvNeXtV2, InceptionNeXt). The model was evaluated using the HDV1 dataset, containing fundus images of different glaucoma stages. Additionally, 40 CNN and 40 ViT models were tested on HDV1 to validate MaxGlaViT's efficiency. Among CNN models, EfficientB6 achieved the highest accuracy (84.91%), while among ViT models, MaxViT-Tiny performed best (86.42%). The scaled MaxViT reached 87.93% accuracy. Adding ECA to the stem block increased accuracy to 89.01%. Replacing MBConv with ConvNeXtV2 further improved it to 89.87%. Finally, integrating ECA in the stem and ConvNeXtV2 in MaxViT blocks resulted in 92.03% accuracy. Testing 80 DL models for glaucoma stage classification, this study presents a comprehensive and comparative analysis. MaxGlaViT outperforms experimental and state-of-the-art models, achieving 92.03% accuracy, 92.33% precision, 92.03% recall, 92.13% f1-score, and 87.12% Cohen's kappa score.
Retrieval Augmented Generation (RAG) systems remain vulnerable to hallucinated answers despite incorporating external knowledge sources. We present LettuceDetect a framework that addresses two critical limitations in existing hallucination detection methods: (1) the context window constraints of traditional encoder-based methods, and (2) the computational inefficiency of LLM based approaches. Building on ModernBERT's extended context capabilities (up to 8k tokens) and trained on the RAGTruth benchmark dataset, our approach outperforms all previous encoder-based models and most prompt-based models, while being approximately 30 times smaller than the best models. LettuceDetect is a token-classification model that processes context-question-answer triples, allowing for the identification of unsupported claims at the token level. Evaluations on the RAGTruth corpus demonstrate an F1 score of 79.22% for example-level detection, which is a 14.8% improvement over Luna, the previous state-of-the-art encoder-based architecture. Additionally, the system can process 30 to 60 examples per second on a single GPU, making it more practical for real-world RAG applications.
As machine learning models grow in complexity and increasingly rely on publicly sourced data, such as the human-annotated labels used in training large language models, they become more vulnerable to label poisoning attacks. These attacks, in which adversaries subtly alter the labels within a training dataset, can severely degrade model performance, posing significant risks in critical applications. In this paper, we propose FLORAL, a novel adversarial training defense strategy based on support vector machines (SVMs) to counter these threats. Utilizing a bilevel optimization framework, we cast the training process as a non-zero-sum Stackelberg game between an attacker, who strategically poisons critical training labels, and the model, which seeks to recover from such attacks. Our approach accommodates various model architectures and employs a projected gradient descent algorithm with kernel SVMs for adversarial training. We provide a theoretical analysis of our algorithm's convergence properties and empirically evaluate FLORAL's effectiveness across diverse classification tasks. Compared to robust baselines and foundation models such as RoBERTa, FLORAL consistently achieves higher robust accuracy under increasing attacker budgets. These results underscore the potential of FLORAL to enhance the resilience of machine learning models against label poisoning threats, thereby ensuring robust classification in adversarial settings.
In recent years, generative models have achieved remarkable performance across diverse applications, including image generation, text synthesis, audio creation, video generation, and data augmentation. Diffusion models have emerged as superior alternatives to Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) by addressing their limitations, such as training instability, mode collapse, and poor representation of multimodal distributions. This success has spurred widespread research interest. In the domain of tabular data, diffusion models have begun to showcase similar advantages over GANs and VAEs, achieving significant performance breakthroughs and demonstrating their potential for addressing unique challenges in tabular data modeling. However, while domains like images and time series have numerous surveys summarizing advancements in diffusion models, there remains a notable gap in the literature for tabular data. Despite the increasing interest in diffusion models for tabular data, there has been little effort to systematically review and summarize these developments. This lack of a dedicated survey limits a clear understanding of the challenges, progress, and future directions in this critical area. This survey addresses this gap by providing a comprehensive review of diffusion models for tabular data. Covering works from June 2015, when diffusion models emerged, to December 2024, we analyze nearly all relevant studies, with updates maintained in a \href{https://github.com/Diffusion-Model-Leiden/awesome-diffusion-models-for-tabular-data}{GitHub repository}. Assuming readers possess foundational knowledge of statistics and diffusion models, we employ mathematical formulations to deliver a rigorous and detailed review, aiming to promote developments in this emerging and exciting area.
Diffusion Probabilistic Models (DPMs) have achieved significant success in generative tasks. However, their training and sampling processes suffer from the issue of distribution mismatch. During the denoising process, the input data distributions differ between the training and inference stages, potentially leading to inaccurate data generation. To obviate this, we analyze the training objective of DPMs and theoretically demonstrate that this mismatch can be alleviated through Distributionally Robust Optimization (DRO), which is equivalent to performing robustness-driven Adversarial Training (AT) on DPMs. Furthermore, for the recently proposed Consistency Model (CM), which distills the inference process of the DPM, we prove that its training objective also encounters the mismatch issue. Fortunately, this issue can be mitigated by AT as well. Based on these insights, we propose to conduct efficient AT on both DPM and CM. Finally, extensive empirical studies validate the effectiveness of AT in diffusion-based models. The code is available at https://github.com/kugwzk/AT_Diff.
Humans are influenced by how information is presented, a phenomenon known as the framing effect. Previous work has shown that LLMs may also be susceptible to framing but has done so on synthetic data and did not compare to human behavior. We introduce WildFrame, a dataset for evaluating LLM responses to positive and negative framing, in naturally-occurring sentences, and compare humans on the same data. WildFrame consists of 1,000 texts, first selecting real-world statements with clear sentiment, then reframing them in either positive or negative light, and lastly, collecting human sentiment annotations. By evaluating eight state-of-the-art LLMs on WildFrame, we find that all models exhibit framing effects similar to humans ($r\geq0.57$), with both humans and models being more influenced by positive rather than negative reframing. Our findings benefit model developers, who can either harness framing or mitigate its effects, depending on the downstream application.
Next-generation galaxy surveys promise unprecedented precision in testing gravity at cosmological scales. However, realising this potential requires accurately modelling the non-linear cosmic web. We address this challenge by exploring conditional generative modelling to create 3D dark matter density fields via score-based (diffusion) and flow-based methods. Our results demonstrate the power of diffusion models to accurately reproduce the matter power spectra and bispectra, even for unseen configurations. They also offer a significant speed-up with slightly reduced accuracy, when flow-based reconstructing the probability distribution function, but they struggle with higher-order statistics. To improve conditional generation, we introduce a novel multi-output model to develop feature representations of the cosmological parameters. Our findings offer a powerful tool for exploring deviations from standard gravity, combining high precision with reduced computational cost, thus paving the way for more comprehensive and efficient cosmological analyses
Privacy concerns in machine learning are heightened by regulations such as the GDPR, which enforces the "right to be forgotten" (RTBF), driving the emergence of machine unlearning as a critical research field. Vertical Federated Learning (VFL) enables collaborative model training by aggregating a sample's features across distributed parties while preserving data privacy at each source. This paradigm has seen widespread adoption in healthcare, finance, and other privacy-sensitive domains. However, existing VFL systems lack robust mechanisms to comply with RTBF requirements, as unlearning methodologies for VFL remain underexplored. In this work, we introduce the first VFL framework with theoretically guaranteed unlearning capabilities, enabling the removal of any data at any time. Unlike prior approaches -- which impose restrictive assumptions on model architectures or data types for removal -- our solution is model- and data-agnostic, offering universal compatibility. Moreover, our framework supports asynchronous unlearning, eliminating the need for all parties to be simultaneously online during the forgetting process. These advancements address critical gaps in current VFL systems, ensuring compliance with RTBF while maintaining operational flexibility.We make all our implementations publicly available at https://github.com/wangln19/vertical-federated-unlearning.
Query expansion plays a crucial role in information retrieval, which aims to bridge the semantic gap between queries and documents to improve matching performance. This paper introduces LLM-QE, a novel approach that leverages Large Language Models (LLMs) to generate document-based query expansions, thereby enhancing dense retrieval models. Unlike traditional methods, LLM-QE designs both rank-based and answer-based rewards and uses these reward models to optimize LLMs to align with the ranking preferences of both retrievers and LLMs, thus mitigating the hallucination of LLMs during query expansion. Our experiments on the zero-shot dense retrieval model, Contriever, demonstrate the effectiveness of LLM-QE, achieving an improvement of over 8%. Furthermore, by incorporating answer-based reward modeling, LLM-QE generates more relevant and precise information related to the documents, rather than simply producing redundant tokens to maximize rank-based rewards. Notably, LLM-QE also improves the training process of dense retrievers, achieving a more than 5% improvement after fine-tuning. All codes are available at https://github.com/NEUIR/LLM-QE.
This paper comprehensively evaluates several recently proposed optimizers for 4-bit training, revealing that low-bit precision amplifies sensitivity to learning rates and often causes unstable gradient norms, leading to divergence at higher learning rates. Among these, SPAM, a recent optimizer featuring momentum reset and spike-aware gradient clipping, achieves the best performance across various bit levels, but struggles to stabilize gradient norms, requiring careful learning rate tuning. To address these limitations, we propose Stable-SPAM, which incorporates enhanced gradient normalization and clipping techniques. In particular, Stable-SPAM (1) adaptively updates the clipping threshold for spiked gradients by tracking their historical maxima; (2) normalizes the entire gradient matrix based on its historical $l_2$-norm statistics; and $(3)$ inherits momentum reset from SPAM to periodically reset the first and second moments of Adam, mitigating the accumulation of spiked gradients. Extensive experiments show that Stable-SPAM effectively stabilizes gradient norms in 4-bit LLM training, delivering superior performance compared to Adam and SPAM. Notably, our 4-bit LLaMA-1B model trained with Stable-SPAM outperforms the BF16 LLaMA-1B trained with Adam by up to $2$ perplexity. Furthermore, when both models are trained in 4-bit, Stable-SPAM achieves the same loss as Adam while requiring only about half the training steps. Code is available at https://github.com/TianjinYellow/StableSPAM.git.
As machine learning models become increasingly prevalent in time series applications, Explainable Artificial Intelligence (XAI) methods are essential for understanding their predictions. Within XAI, feature attribution methods aim to identify which input features contribute the most to a model's prediction, with their evaluation typically relying on perturbation-based metrics. Through systematic empirical analysis across multiple datasets, model architectures, and perturbation strategies, we reveal previously overlooked class-dependent effects in these metrics: they show varying effectiveness across classes, achieving strong results for some while remaining less sensitive to others. In particular, we find that the most effective perturbation strategies often demonstrate the most pronounced class differences. Our analysis suggests that these effects arise from the learned biases of classifiers, indicating that perturbation-based evaluation may reflect specific model behaviors rather than intrinsic attribution quality. We propose an evaluation framework with a class-aware penalty term to help assess and account for these effects in evaluating feature attributions, offering particular value for class-imbalanced datasets. Although our analysis focuses on time series classification, these class-dependent effects likely extend to other structured data domains where perturbation-based evaluation is common.
Large-scale physical systems defined on irregular grids pose significant scalability challenges for deep learning methods, especially in the presence of long-range interactions and multi-scale coupling. Traditional approaches that compute all pairwise interactions, such as attention, become computationally prohibitive as they scale quadratically with the number of nodes. We present Erwin, a hierarchical transformer inspired by methods from computational many-body physics, which combines the efficiency of tree-based algorithms with the expressivity of attention mechanisms. Erwin employs ball tree partitioning to organize computation, which enables linear-time attention by processing nodes in parallel within local neighborhoods of fixed size. Through progressive coarsening and refinement of the ball tree structure, complemented by a novel cross-ball interaction mechanism, it captures both fine-grained local details and global features. We demonstrate Erwin's effectiveness across multiple domains, including cosmology, molecular dynamics, and particle fluid dynamics, where it consistently outperforms baseline methods both in accuracy and computational efficiency.
Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While they may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing FADE: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for evaluating feature-description alignment. FADE evaluates alignment across four key metrics - Clarity, Responsiveness, Purity, and Faithfulness - and systematically quantifies the causes for the misalignment of feature and their description. We apply FADE to analyze existing open-source feature descriptions, and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs as compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release FADE as an open-source package at: https://github.com/brunibrun/FADE.
This paper presents Hotter and Colder, a dataset designed to analyze various types of online behavior in Icelandic blog comments. Building on previous work, we used GPT-4o mini to annotate approximately 800,000 comments for 25 tasks, including sentiment analysis, emotion detection, hate speech, and group generalizations. Each comment was automatically labeled on a 5-point Likert scale. In a second annotation stage, comments with high or low probabilities of containing each examined behavior were subjected to manual revision. By leveraging crowdworkers to refine these automatically labeled comments, we ensure the quality and accuracy of our dataset resulting in 12,232 uniquely annotated comments and 19,301 annotations. Hotter and Colder provides an essential resource for advancing research in content moderation and automatically detectiong harmful online behaviors in Icelandic.
Recently, the Muon optimizer based on matrix orthogonalization has demonstrated strong results in training small-scale language models, but the scalability to larger models has not been proven. We identify two crucial techniques for scaling up Muon: (1) adding weight decay and (2) carefully adjusting the per-parameter update scale. These techniques allow Muon to work out-of-the-box on large-scale training without the need of hyper-parameter tuning. Scaling law experiments indicate that Muon achieves $\sim\!2\times$ computational efficiency compared to AdamW with compute optimal training. Based on these improvements, we introduce Moonlight, a 3B/16B-parameter Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon. Our model improves the current Pareto frontier, achieving better performance with much fewer training FLOPs compared to prior models. We open-source our distributed Muon implementation that is memory optimal and communication efficient. We also release the pretrained, instruction-tuned, and intermediate checkpoints to support future research.
We analyse the convergence of one-hidden-layer ReLU networks trained by gradient flow on $n$ data points. Our main contribution leverages the high dimensionality of the ambient space, which implies low correlation of the input samples, to demonstrate that a network with width of order $\log(n)$ neurons suffices for global convergence with high probability. Our analysis uses a Polyak-{\L}ojasiewicz viewpoint along the gradient-flow trajectory, which provides an exponential rate of convergence of $\frac{1}{n}$. When the data are exactly orthogonal, we give further refined characterizations of the convergence speed, proving its asymptotic behavior lies between the orders $\frac{1}{n}$ and $\frac{1}{\sqrt{n}}$, and exhibiting a phase-transition phenomenon in the convergence rate, during which it evolves from the lower bound to the upper, and in a relative time of order $\frac{1}{\log(n)}$.
As Large Language Models (LLMs) continue to advance in understanding and generating long sequences, new safety concerns have been introduced through the long context. However, the safety of LLMs in long-context tasks remains under-explored, leaving a significant gap in both evaluation and improvement of their safety. To address this, we introduce LongSafety, the first comprehensive benchmark specifically designed to evaluate LLM safety in open-ended long-context tasks. LongSafety encompasses 7 categories of safety issues and 6 user-oriented long-context tasks, with a total of 1,543 test cases, averaging 5,424 words per context. Our evaluation towards 16 representative LLMs reveals significant safety vulnerabilities, with most models achieving safety rates below 55%. Our findings also indicate that strong safety performance in short-context scenarios does not necessarily correlate with safety in long-context tasks, emphasizing the unique challenges and urgency of improving long-context safety. Moreover, through extensive analysis, we identify challenging safety issues and task types for long-context models. Furthermore, we find that relevant context and extended input sequences can exacerbate safety risks in long-context scenarios, highlighting the critical need for ongoing attention to long-context safety challenges. Our code and data are available at https://github.com/thu-coai/LongSafety.
Scene flow estimation aims to predict 3D motion from consecutive point cloud frames, which is of great interest in autonomous driving field. Existing methods face challenges such as insufficient spatio-temporal modeling and inherent loss of fine-grained feature during voxelization. However, the success of Mamba, a representative state space model (SSM) that enables global modeling with linear complexity, provides a promising solution. In this paper, we propose MambaFlow, a novel scene flow estimation network with a mamba-based decoder. It enables deep interaction and coupling of spatio-temporal features using a well-designed backbone. Innovatively, we steer the global attention modeling of voxel-based features with point offset information using an efficient Mamba-based decoder, learning voxel-to-point patterns that are used to devoxelize shared voxel representations into point-wise features. To further enhance the model's generalization capabilities across diverse scenarios, we propose a novel scene-adaptive loss function that automatically adapts to different motion patterns.Extensive experiments on the Argoverse 2 benchmark demonstrate that MambaFlow achieves state-of-the-art performance with real-time inference speed among existing works, enabling accurate flow estimation in real-world urban scenarios. The code is available at https://github.com/SCNU-RISLAB/MambaFlow.
We explore \textbf{C}ross-lingual \textbf{B}ackdoor \textbf{AT}tacks (X-BAT) in multilingual Large Language Models (mLLMs), revealing how backdoors inserted in one language can automatically transfer to others through shared embedding spaces. Using toxicity classification as a case study, we demonstrate that attackers can compromise multilingual systems by poisoning data in a single language, with rare and high-occurring tokens serving as specific, effective triggers. Our findings expose a critical vulnerability that influences the model's architecture, resulting in a concealed backdoor effect during the information flow. Our code and data are publicly available https://github.com/himanshubeniwal/X-BAT.