Authors: A. Bochkov
Abstract: The prevailing paradigm for scaling large language models (LLMs) involves monolithic, end-to-end training, a resource-intensive process that lacks flexibility. This paper explores an alternative, constructive approach to model development, built upon the foundation of non-trainable, deterministic input embeddings. In prior [1], we established that high-level semantic reasoning can emerge in Transformers using frozen embeddings derived from the visual structure of Unicode glyphs. Here, we demonstrate that this fixed representational substrate acts as a universal "docking port," enabling two powerful and efficient scaling paradigms: seamless modular composition and progressive layer-wise growth. First, we show that specialist models trained on disparate datasets (e.g., Russian and Chinese text) can be merged into a single, more capable Mixture-of-Experts (MoE) model, post-training, with zero architectural modification. This is achieved by simply averaging their output logits. The resulting MoE model exhibits immediate performance improvements on reasoning benchmarks like MMLU, surpassing its constituent experts without catastrophic forgetting. Second, we introduce a layer-wise constructive training methodology, where a deep Transformer is "grown" by progressively stacking and training one layer at a time. This method demonstrates stable convergence and a clear correlation between model depth and the emergence of complex reasoning abilities, such as those required for SQuAD. Our findings suggest a paradigm shift from monolithic optimization towards a more biological or constructive model of AI development, where complexity is built incrementally and modules can be composed freely. This opens new avenues for resource-efficient scaling, continual learning, and a more democratized ecosystem for building powerful AI systems. We release all code and models to facilitate further research.
Authors: Fran\c{c}ois Gard\`eres, Shizhe Chen, Camille-Sovanneary Gauthier, Jean Ponce
Abstract: The composed image retrieval (CIR) task is to retrieve target images given a reference image and a modification text. Recent methods for CIR leverage large pretrained vision-language models (VLMs) and achieve good performance on general-domain concepts like color and texture. However, they still struggle with application domains like fashion, because the rich and diverse vocabulary used in fashion requires specific fine-grained vision and language understanding. An additional difficulty is the lack of large-scale fashion datasets with detailed and relevant annotations, due to the expensive cost of manual annotation by specialists. To address these challenges, we introduce FACap, a large-scale, automatically constructed fashion-domain CIR dataset. It leverages web-sourced fashion images and a two-stage annotation pipeline powered by a VLM and a large language model (LLM) to generate accurate and detailed modification texts. Then, we propose a new CIR model FashionBLIP-2, which fine-tunes the general-domain BLIP-2 model on FACap with lightweight adapters and multi-head query-candidate matching to better account for fine-grained fashion-specific information. FashionBLIP-2 is evaluated with and without additional fine-tuning on the Fashion IQ benchmark and the enhanced evaluation dataset enhFashionIQ, leveraging our pipeline to obtain higher-quality annotations. Experimental results show that the combination of FashionBLIP-2 and pretraining with FACap significantly improves the model's performance in fashion CIR especially for retrieval with fine-grained modification texts, demonstrating the value of our dataset and approach in a highly demanding environment such as e-commerce websites. Code is available at https://fgxaos.github.io/facap-paper-website/.
Authors: Eric Yeats, Darryl Hannan, Henry Kvinge, Timothy Doster, Scott Mahan
Abstract: Machine unlearning (MU) is a promising cost-effective method to cleanse undesired information (generated concepts, biases, or patterns) from foundational diffusion models. While MU is orders of magnitude less costly than retraining a diffusion model without the undesired information, it can be challenging and labor-intensive to prove that the information has been fully removed from the model. Moreover, MU can damage diffusion model performance on surrounding concepts that one would like to retain, making it unclear if the diffusion model is still fit for deployment. We introduce autoeval-dmun, an automated tool which leverages (vision-) language models to thoroughly assess unlearning in diffusion models. Given a target concept, autoeval-dmun extracts structured, relevant world knowledge from the language model to identify nearby concepts which are likely damaged by unlearning and to circumvent unlearning with adversarial prompts. We use our automated tool to evaluate popular diffusion model unlearning methods, revealing that language models (1) impose semantic orderings of nearby concepts which correlate well with unlearning damage and (2) effectively circumvent unlearning with synthetic adversarial prompts.
Authors: Francesco Ferrini, Veronica Lachi, Antonio Longa, Bruno Lepri, Andrea Passerini
Abstract: Graph Neural Networks (GNNs) often struggle to capture the link-specific structural patterns crucial for accurate link prediction, as their node-centric message-passing schemes overlook the subgraph structures connecting a pair of nodes. Existing methods to inject such structural context either incur high computational cost or rely on simplistic heuristics (e.g., common neighbor counts) that fail to model multi-hop dependencies. We introduce SP4LP (Shortest Path for Link Prediction), a novel framework that combines GNN-based node encodings with sequence modeling over shortest paths. Specifically, SP4LP first applies a GNN to compute representations for all nodes, then extracts the shortest path between each candidate node pair and processes the resulting sequence of node embeddings using a sequence model. This design enables SP4LP to capture expressive multi-hop relational patterns with computational efficiency. Empirically, SP4LP achieves state-of-the-art performance across link prediction benchmarks. Theoretically, we prove that SP4LP is strictly more expressive than standard message-passing GNNs and several state-of-the-art structural features methods, establishing it as a general and principled approach for link prediction in graphs.
Authors: Samin Yeasar Arnob, Zhan Su, Minseon Kim, Oleksiy Ostapenko, Riyasat Ohib, Esra'a Saleh, Doina Precup, Lucas Caccia, Alessandro Sordoni
Abstract: Merging parameter-efficient task experts has recently gained growing attention as a way to build modular architectures that can be rapidly adapted on the fly for specific downstream tasks, without requiring additional fine-tuning. Typically, LoRA serves as the foundational building block of such parameter-efficient modular architectures, leveraging low-rank weight structures to reduce the number of trainable parameters. In this paper, we study the properties of sparse adapters, which train only a subset of weights in the base neural network, as potential building blocks of modular architectures. First, we propose a simple method for training highly effective sparse adapters, which is conceptually simpler than existing methods in the literature and surprisingly outperforms both LoRA and full fine-tuning in our setting. Next, we investigate the merging properties of these sparse adapters by merging adapters for up to 20 natural language processing tasks, thus scaling beyond what is usually studied in the literature. Our findings demonstrate that sparse adapters yield superior in-distribution performance post-merging compared to LoRA or full model merging. Achieving strong held-out performance remains a challenge for all methods considered.
Authors: Dongxiao He, Yongqi Huang, Jitao Zhao, Xiaobao Wang, Zhen Wang
Abstract: Graph Contrastive Learning (GCL) is a widely adopted approach in self-supervised graph representation learning, applying contrastive objectives to produce effective representations. However, current GCL methods primarily focus on capturing implicit semantic relationships, often overlooking the structural commonsense embedded within the graph's structure and attributes, which contains underlying knowledge crucial for effective representation learning. Due to the lack of explicit information and clear guidance in general graph, identifying and integrating such structural commonsense in GCL poses a significant challenge. To address this gap, we propose a novel framework called Structural Commonsense Unveiling in Graph Contrastive Learning (Str-GCL). Str-GCL leverages first-order logic rules to represent structural commonsense and explicitly integrates them into the GCL framework. It introduces topological and attribute-based rules without altering the original graph and employs a representation alignment mechanism to guide the encoder in effectively capturing this commonsense. To the best of our knowledge, this is the first attempt to directly incorporate structural commonsense into GCL. Extensive experiments demonstrate that Str-GCL outperforms existing GCL methods, providing a new perspective on leveraging structural commonsense in graph representation learning.
Authors: Karthik Pappu, Prathamesh Dinesh Joshi, Raj Abhijit Dandekar, Rajat Dandekar, Sreedath Panat
Abstract: Accurately modeling malware propagation is essential for designing effective cybersecurity defenses, particularly against adaptive threats that evolve in real time. While traditional epidemiological models and recent neural approaches offer useful foundations, they often fail to fully capture the nonlinear feedback mechanisms present in real-world networks. In this work, we apply scientific machine learning to malware modeling by evaluating three approaches: classical Ordinary Differential Equations (ODEs), Universal Differential Equations (UDEs), and Neural ODEs. Using data from the Code Red worm outbreak, we show that the UDE approach substantially reduces prediction error compared to both traditional and neural baselines by 44%, while preserving interpretability. We introduce a symbolic recovery method that transforms the learned neural feedback into explicit mathematical expressions, revealing suppression mechanisms such as network saturation, security response, and malware variant evolution. Our results demonstrate that hybrid physics-informed models can outperform both purely analytical and purely neural approaches, offering improved predictive accuracy and deeper insight into the dynamics of malware spread. These findings support the development of early warning systems, efficient outbreak response strategies, and targeted cyber defense interventions.
Authors: Zhaojing Zhou, Xunchao Li, Minghao Li, Handi Zhang, Haoshuang Wang, Wenbin Chang, Yiqun Liu, Qingqing Dang, Dianhai Yu, Yanjun Ma, Haifeng Wang
Abstract: The rapid scaling of Large Language Models (LLMs) elevates inference costs and compounds substantial deployment barriers. While quantization to 8 or 4 bits mitigates this, sub-3-bit methods face severe accuracy, scalability, and efficiency degradation. We propose Convolutional Code Quantization (CCQ), an inference-optimized quantization approach compressing LLMs to 2.0-2.75 bits with minimal accuracy loss. Departing from error-prone scalar quantization or slow vector quantization, CCQ integrates a hardware-aware bit-shift encoding and decoding solution with Convolutional Code, Hybrid Encoding, and Code Cluster, jointly overcoming accuracy-speed bottlenecks. We construct a lookup-free encoding space, enabling a linear mapping between the codebook and weight vectors, thereby optimizing inference performance. Meanwhile, by drawing on the concept of data mapping from vector quantization, we minimize the performance degradation of the model under extremely low-bit conditions. Experiments demonstrate that CCQ achieves outstanding performance on LLMs across various benchmarks. We compress DeepSeek-V3 (671B total parameters) to 184GB and ERNIE-4.5-300B-A47B to 89GB, enabling single-GPU deployment of ERNIE 4.5 and eliminating inter-card communication. The 2-bit ERNIE-4.5-300B-A47B model and inference engine have been open-sourced.
Authors: Zixuan Huang, Kecheng Huang, Lihao Yin, Bowei He, Huiling Zhen, Mingxuan Yuan, Zili Shao
Abstract: Large Language Models (LLMs) have gained widespread popularity and are increasingly integrated into various applications. However, their capabilities can be exploited for both benign and harmful purposes. Despite rigorous training and fine-tuning for safety, LLMs remain vulnerable to jailbreak attacks. Recently, multi-turn attacks have emerged, exacerbating the issue. Unlike single-turn attacks, multi-turn attacks gradually escalate the dialogue, making them more difficult to detect and mitigate, even after they are identified. In this study, we propose G-Guard, an innovative attention-aware GNN-based input classifier designed to defend against multi-turn jailbreak attacks on LLMs. G-Guard constructs an entity graph for multi-turn queries, explicitly capturing relationships between harmful keywords and queries even when those keywords appear only in previous queries. Additionally, we introduce an attention-aware augmentation mechanism that retrieves the most similar single-turn query based on the multi-turn conversation. This retrieved query is treated as a labeled node in the graph, enhancing the ability of GNN to classify whether the current query is harmful. Evaluation results demonstrate that G-Guard outperforms all baselines across all datasets and evaluation metrics.
Authors: Sua Lee, Kyubum Shin, Jung Ho Park
Abstract: Recent advances in pre-trained Vision Language Models (VLM) have shown promising potential for effectively adapting to downstream tasks through prompt learning, without the need for additional annotated paired datasets. To supplement the text information in VLM trained on correlations with vision data, new approaches leveraging Large Language Models (LLM) in prompts have been proposed, enhancing robustness to unseen and diverse data. Existing methods typically extract text-based responses (i.e., descriptions) from LLM to incorporate into prompts; however, this approach suffers from high variability and low reliability. In this work, we propose Description-free Multi-prompt Learning(DeMul), a novel method that eliminates the process of extracting descriptions and instead directly distills knowledge from LLM into prompts. By adopting a description-free approach, prompts can encapsulate richer semantics while still being represented as continuous vectors for optimization, thereby eliminating the need for discrete pre-defined templates. Additionally, in a multi-prompt setting, we empirically demonstrate the potential of prompt weighting in reflecting the importance of different prompts during training. Experimental results show that our approach achieves superior performance across 11 recognition datasets.
Authors: Huibo Xu, Runlong Yu, Likang Wu, Xianquan Wang, Qi Liu
Abstract: Diffusion models, a type of generative model, have shown promise in time series forecasting. But they face limitations like rigid source distributions and limited sampling paths, which hinder their performance. Flow matching offers faster generation, higher-quality outputs, and greater flexibility, while also possessing the ability to utilize valuable information from the prediction errors of prior models, which were previously inaccessible yet critically important. To address these challenges and fully unlock the untapped potential of flow matching, we propose Conditional Guided Flow Matching (CGFM). CGFM extends flow matching by incorporating the outputs of an auxiliary model, enabling a previously unattainable capability in the field: learning from the errors of the auxiliary model. For time series forecasting tasks, it integrates historical data as conditions and guidance, constructs two-sided conditional probability paths, and uses a general affine path to expand the space of probability paths, ultimately leading to improved predictions. Extensive experiments show that CGFM consistently enhances and outperforms state-of-the-art models, highlighting its effectiveness in advancing forecasting methods.
Authors: Elia Piccoli, Malio Li, Giacomo Carf\`i, Vincenzo Lomonaco, Davide Bacciu
Abstract: The recent focus and release of pre-trained models have been a key components to several advancements in many fields (e.g. Natural Language Processing and Computer Vision), as a matter of fact, pre-trained models learn disparate latent embeddings sharing insightful representations. On the other hand, Reinforcement Learning (RL) focuses on maximizing the cumulative reward obtained via agent's interaction with the environment. RL agents do not have any prior knowledge about the world, and they either learn from scratch an end-to-end mapping between the observation and action spaces or, in more recent works, are paired with monolithic and computationally expensive Foundational Models. How to effectively combine and leverage the hidden information of different pre-trained models simultaneously in RL is still an open and understudied question. In this work, we propose Weight Sharing Attention (WSA), a new architecture to combine embeddings of multiple pre-trained models to shape an enriched state representation, balancing the tradeoff between efficiency and performance. We run an extensive comparison between several combination modes showing that WSA obtains comparable performance on multiple Atari games compared to end-to-end models. Furthermore, we study the generalization capabilities of this approach and analyze how scaling the number of models influences agents' performance during and after training.
Authors: Florian Redhardt, Yassir Akram, Simon Schug
Abstract: Can neural networks systematically capture discrete, compositional task structure despite their continuous, distributed nature? The impressive capabilities of large-scale neural networks suggest that the answer to this question is yes. However, even for the most capable models, there are still frequent failure cases that raise doubts about their compositionality. Here, we seek to understand what it takes for a standard neural network to generalize over tasks that share compositional structure. We find that simply scaling data and model size leads to compositional generalization. We show that this holds across different task encodings as long as the training distribution sufficiently covers the task space. In line with this finding, we prove that standard multilayer perceptrons can approximate a general class of compositional task families to arbitrary precision using only a linear number of neurons with respect to the number of task modules. Finally, we uncover that if networks successfully compositionally generalize, the constituents of a task can be linearly decoded from their hidden activations. We show that this metric correlates with failures of text-to-image generation models to compose known concepts.
Authors: Yunyi Li, Maria De-Arteaga, Maytal Saar-Tsechansky
Abstract: Reliable data is a cornerstone of modern organizational systems. A notable data integrity challenge stems from label bias, which refers to systematic errors in a label, a covariate that is central to a quantitative analysis, such that its quality differs across social groups. This type of bias has been conceptually and empirically explored and is widely recognized as a pressing issue across critical domains. However, effective methodologies for addressing it remain scarce. In this work, we propose Decoupled Confident Learning (DeCoLe), a principled machine learning based framework specifically designed to detect mislabeled instances in datasets affected by label bias, enabling bias aware mislabelling detection and facilitating data quality improvement. We theoretically justify the effectiveness of DeCoLe and evaluate its performance in the impactful context of hate speech detection, a domain where label bias is a well documented challenge. Empirical results demonstrate that DeCoLe excels at bias aware mislabeling detection, consistently outperforming alternative approaches for label error detection. Our work identifies and addresses the challenge of bias aware mislabeling detection and offers guidance on how DeCoLe can be integrated into organizational data management practices as a powerful tool to enhance data reliability.
Authors: Minchan Jeong, J. Jon Ryu, Se-Young Yun, Gregory W. Wornell
Abstract: The Koopman operator provides a principled framework for analyzing nonlinear dynamical systems through linear operator theory. Recent advances in dynamic mode decomposition (DMD) have shown that trajectory data can be used to identify dominant modes of a system in a data-driven manner. Building on this idea, deep learning methods such as VAMPnet and DPNet have been proposed to learn the leading singular subspaces of the Koopman operator. However, these methods require backpropagation through potentially numerically unstable operations on empirical second moment matrices, such as singular value decomposition and matrix inversion, during objective computation, which can introduce biased gradient estimates and hinder scalability to large systems. In this work, we propose a scalable and conceptually simple method for learning the top-k singular functions of the Koopman operator for stochastic dynamical systems based on the idea of low-rank approximation. Our approach eliminates the need for unstable linear algebraic operations and integrates easily into modern deep learning pipelines. Empirical results demonstrate that the learned singular subspaces are both reliable and effective for downstream tasks such as eigen-analysis and multi-step prediction.
Authors: Maya Kruse, Majid Afshar, Saksham Khatwani, Anoop Mayampurath, Guanhua Chen, Yanjun Gao
Abstract: Large language models (LLMs) often behave inconsistently across inputs, indicating uncertainty and motivating the need for its quantification in high-stakes settings. Prior work on calibration and uncertainty quantification often focuses on individual models, overlooking the potential of model diversity. We hypothesize that LLMs make complementary predictions due to differences in training and the Zipfian nature of language, and that aggregating their outputs leads to more reliable uncertainty estimates. To leverage this, we propose MUSE (Multi-LLM Uncertainty via Subset Ensembles), a simple information-theoretic method that uses Jensen-Shannon Divergence to identify and aggregate well-calibrated subsets of LLMs. Experiments on binary prediction tasks demonstrate improved calibration and predictive performance compared to single-model and naive ensemble baselines.
Authors: Erfan Hamdi, Emma Lejeune
Abstract: Data driven approaches have the potential to make modeling complex, nonlinear physical phenomena significantly more computationally tractable. For example, computational modeling of fracture is a core challenge where machine learning techniques have the potential to provide a much needed speedup that would enable progress in areas such as mutli-scale modeling and uncertainty quantification. Currently, phase field modeling (PFM) of fracture is one such approach that offers a convenient variational formulation to model crack nucleation, branching and propagation. To date, machine learning techniques have shown promise in approximating PFM simulations. However, most studies rely on overly simple benchmarks that do not reflect the true complexity of the fracture processes where PFM excels as a method. To address this gap, we introduce a challenging dataset based on PFM simulations designed to benchmark and advance ML methods for fracture modeling. This dataset includes three energy decomposition methods, two boundary conditions, and 1,000 random initial crack configurations for a total of 6,000 simulations. Each sample contains 100 time steps capturing the temporal evolution of the crack field. Alongside this dataset, we also implement and evaluate Physics Informed Neural Networks (PINN), Fourier Neural Operators (FNO) and UNet models as baselines, and explore the impact of ensembling strategies on prediction accuracy. With this combination of our dataset and baseline models drawn from the literature we aim to provide a standardized and challenging benchmark for evaluating machine learning approaches to solid mechanics. Our results highlight both the promise and limitations of popular current models, and demonstrate the utility of this dataset as a testbed for advancing machine learning in fracture mechanics research.
Authors: Zhengyu Tian, Anantha Padmanaban Krishna Kumar, Hemant Krishnakumar, Reza Rawassizadeh
Abstract: As large language models (LLMs) and visual language models (VLMs) grow in scale and application, attention mechanisms have become a central computational bottleneck due to their high memory and time complexity. While many efficient attention variants have been proposed, there remains a lack of rigorous evaluation on their actual energy usage and hardware resource demands during training. In this work, we benchmark eight attention mechanisms in training GPT-2 architecture, measuring key metrics including training time, GPU memory usage, FLOPS, CPU usage, and power consumption. Our results reveal that attention mechanisms with optimized kernel implementations, including Flash Attention, Locality-Sensitive Hashing (LSH) Attention, and Multi-Head Latent Attention (MLA), achieve the best energy efficiency. We further show that lower GPU power alone does not guarantee reduced energy use, as training time plays an equally important role. Our study highlights the importance of energy-aware benchmarking in attention design and provides a practical insight for selecting resource-efficient mechanisms. All our codes are available at GitHub.
Authors: Giulio Rossolini, Fabio Brau, Alessandro Biondi, Battista Biggio, Giorgio Buttazzo
Abstract: As machine learning models become increasingly deployed across the edge of internet of things environments, a partitioned deep learning paradigm in which models are split across multiple computational nodes introduces a new dimension of security risk. Unlike traditional inference setups, these distributed pipelines span the model computation across heterogeneous nodes and communication layers, thereby exposing a broader attack surface to potential adversaries. Building on these motivations, this work explores a previously overlooked vulnerability: even when both the edge and cloud components of the model are inaccessible (i.e., black-box), an adversary who intercepts the intermediate features transmitted between them can still pose a serious threat. We demonstrate that, under these mild and realistic assumptions, an attacker can craft highly transferable proxy models, making the entire deep learning system significantly more vulnerable to evasion attacks. In particular, the intercepted features can be effectively analyzed and leveraged to distill surrogate models capable of crafting highly transferable adversarial examples against the target model. To this end, we propose an exploitation strategy specifically designed for distributed settings, which involves reconstructing the original tensor shape from vectorized transmitted features using simple statistical analysis, and adapting surrogate architectures accordingly to enable effective feature distillation. A comprehensive and systematic experimental evaluation has been conducted to demonstrate that surrogate models trained with the proposed strategy, i.e., leveraging intermediate features, tremendously improve the transferability of adversarial attacks. These findings underscore the urgent need to account for intermediate feature leakage in the design of secure distributed deep learning systems.
Authors: Chunzhuo Wang, Hans Hallez, Bart Vanrumste
Abstract: Automated food intake gesture detection plays a vital role in dietary monitoring, enabling objective and continuous tracking of eating behaviors to support better health outcomes. Wrist-worn inertial measurement units (IMUs) have been widely used for this task with promising results. More recently, contactless radar sensors have also shown potential. This study explores whether combining wearable and contactless sensing modalities through multimodal learning can further improve detection performance. We also address a major challenge in multimodal learning: reduced robustness when one modality is missing. To this end, we propose a robust multimodal temporal convolutional network with cross-modal attention (MM-TCN-CMA), designed to integrate IMU and radar data, enhance gesture detection, and maintain performance under missing modality conditions. A new dataset comprising 52 meal sessions (3,050 eating gestures and 797 drinking gestures) from 52 participants is developed and made publicly available. Experimental results show that the proposed framework improves the segmental F1-score by 4.3% and 5.2% over unimodal Radar and IMU models, respectively. Under missing modality scenarios, the framework still achieves gains of 1.3% and 2.4% for missing radar and missing IMU inputs. This is the first study to demonstrate a robust multimodal learning framework that effectively fuses IMU and radar data for food intake gesture detection.
Authors: Julianna Piskorz, Krzysztof Kacprzyk, Mihaela van der Schaar
Abstract: The Average Treatment Effect (ATE) is a foundational metric in causal inference, widely used to assess intervention efficacy in randomized controlled trials (RCTs). However, in many applications -- particularly in healthcare -- this static summary fails to capture the nuanced dynamics of treatment effects that vary with both dose and time. We propose a framework for modelling treatment effect trajectories as smooth surfaces over dose and time, enabling the extraction of clinically actionable insights such as onset time, peak effect, and duration of benefit. To ensure interpretability, robustness, and verifiability -- key requirements in high-stakes domains -- we adapt SemanticODE, a recent framework for interpretable trajectory modelling, to the causal setting where treatment effects are never directly observed. Our approach decouples the estimation of trajectory shape from the specification of clinically relevant properties (e.g., maxima, inflection points), supporting domain-informed priors, post-hoc editing, and transparent analysis. We show that our method yields accurate, interpretable, and editable models of treatment dynamics, facilitating both rigorous causal analysis and practical decision-making.
Authors: Aaron Foote, Danny Krizanc
Abstract: Along with accurate prediction, understanding the contribution of each feature to the making of the prediction, i.e., the importance of the feature, is a desirable and arguably necessary component of a machine learning model. For a complex model such as a random forest, such importances are not innate -- as they are, e.g., with linear regression. Efficient methods have been created to provide such capabilities, with one of the most popular among them being permutation feature importance due to its efficiency, model-agnostic nature, and perceived intuitiveness. However, permutation feature importance has been shown to be misleading in the presence of dependent features as a result of the creation of unrealistic observations when permuting the dependent features. In this work, we develop TRIP (Test for Reliable Interpretation via Permutation), a test requiring minimal assumptions that is able to detect unreliable permutation feature importance scores that are the result of model extrapolation. To build on this, we demonstrate how the test can be complemented in order to allow its use in high dimensional settings. Through testing on simulated data and applications, our results show that the test can be used to reliably detect when permutation feature importance scores are unreliable.
Authors: Pierre Osselin, Masaki Adachi, Xiaowen Dong, Michael A. Osborne
Abstract: Zeroth-order local optimisation algorithms are essential for solving real-valued black-box optimisation problems. Among these, Natural Evolution Strategies (NES) represent a prominent class, particularly well-suited for scenarios where prior distributions are available. By optimising the objective function in the space of search distributions, NES algorithms naturally integrate prior knowledge during initialisation, making them effective in settings such as semi-supervised learning and user-prior belief frameworks. However, due to their reliance on random sampling and Monte Carlo estimates, NES algorithms can suffer from limited sample efficiency. In this paper, we introduce a novel class of algorithms, termed Probabilistic Natural Evolutionary Strategy Algorithms (ProbNES), which enhance the NES framework with Bayesian quadrature. We show that ProbNES algorithms consistently outperforms their non-probabilistic counterparts as well as global sample efficient methods such as Bayesian Optimisation (BO) or $\pi$BO across a wide range of tasks, including benchmark test functions, data-driven optimisation tasks, user-informed hyperparameter tuning tasks and locomotion tasks.
Authors: Paola Causin, Alessio Marta
Abstract: High-dimensional datasets often exhibit low-dimensional geometric structures, as suggested by the manifold hypothesis, which implies that data lie on a smooth manifold embedded in a higher-dimensional ambient space. While this insight underpins many advances in machine learning and inverse problems, fully leveraging it requires to deal with three key tasks: estimating the intrinsic dimension (ID) of the manifold, constructing appropriate local coordinates, and learning mappings between ambient and manifold spaces. In this work, we propose a framework that addresses all these challenges using a Mixture of Variational Autoencoders (VAEs) and tools from Riemannian geometry. We specifically focus on estimating the ID of datasets by analyzing the numerical rank of the VAE decoder pullback metric. The estimated ID guides the construction of an atlas of local charts using a mixture of invertible VAEs, enabling accurate manifold parameterization and efficient inference. We how this approach enhances solutions to ill-posed inverse problems, particularly in biomedical imaging, by enforcing that reconstructions lie on the learned manifold. Lastly, we explore the impact of network pruning on manifold geometry and reconstruction quality, showing that the intrinsic dimension serves as an effective proxy for monitoring model capacity.
Authors: Jacob Hauck, Yanzhi Zhang
Abstract: We develop a new and general encode-approximate-reconstruct operator learning model that leverages learned neural representations of bases for input and output function distributions. We introduce the concepts of \textit{numerical operator learning} and \textit{discretization independence}, which clarify the relationship between theoretical formulations and practical realizations of operator learning models. Our model is discretization-independent, making it particularly effective for multifidelity learning. We establish theoretical approximation guarantees, demonstrating uniform universal approximation under strong assumptions on the input functions and statistical approximation under weaker conditions. To our knowledge, this is the first comprehensive study that investigates how discretization independence enables robust and efficient multifidelity operator learning. We validate our method through extensive numerical experiments involving both local and nonlocal PDEs, including time-independent and time-dependent problems. The results show that multifidelity training significantly improves accuracy and computational efficiency. Moreover, multifidelity training further enhances empirical discretization independence.
Authors: Alan Malek, Jiawei Ge, Jiawei Ge, Chi Jin, Andr\'as Gy\"orgy, Csaba Szepesv\'ari
Abstract: While state-of-the-art large language models (LLMs) demonstrate advanced reasoning capabilities-achieving remarkable performance on challenging competitive math and coding benchmarks-they also frequently fail on tasks that are easy for humans. This work studies the performance of frontier LLMs on a broad set of such "easy" reasoning problems. By extending previous work in the literature, we create a suite of procedurally generated simple reasoning tasks, including counting, first-order logic, proof trees, and travel planning, with changeable parameters (such as document length. or the number of variables in a math problem) that can arbitrarily increase the amount of computation required to produce the answer while preserving the fundamental difficulty. While previous work showed that traditional, non-thinking models can be made to fail on such problems, we demonstrate that even state-of-the-art thinking models consistently fail on such problems and for similar reasons (e.g. statistical shortcuts, errors in intermediate steps, and difficulties in processing long contexts). To further understand the behavior of the models, we introduce the unpuzzles dataset, a different "easy" benchmark consisting of trivialized versions of well-known math and logic puzzles. Interestingly, while modern LLMs excel at solving the original puzzles, they tend to fail on the trivialized versions, exhibiting several systematic failure patterns related to memorizing the originals. We show that this happens even if the models are otherwise able to solve problems with different descriptions but requiring the same logic. Our results highlight that out-of-distribution generalization is still problematic for frontier language models and the new generation of thinking models, even for simple reasoning tasks, and making tasks easier does not necessarily imply improved performance.
Authors: Md Abrar Jahin, Taufikur Rahman Fuad, M. F. Mridha, Nafiz Fahad, Md. Jakir Hossen
Abstract: Federated Learning (FL) faces inherent challenges in balancing model performance, privacy preservation, and communication efficiency, especially in non-IID decentralized environments. Recent approaches either sacrifice formal privacy guarantees, incur high overheads, or overlook quantum-enhanced expressivity. We introduce AdeptHEQ-FL, a unified hybrid classical-quantum FL framework that integrates (i) a hybrid CNN-PQC architecture for expressive decentralized learning, (ii) an adaptive accuracy-weighted aggregation scheme leveraging differentially private validation accuracies, (iii) selective homomorphic encryption (HE) for secure aggregation of sensitive model layers, and (iv) dynamic layer-wise adaptive freezing to minimize communication overhead while preserving quantum adaptability. We establish formal privacy guarantees, provide convergence analysis, and conduct extensive experiments on the CIFAR-10, SVHN, and Fashion-MNIST datasets. AdeptHEQ-FL achieves a $\approx 25.43\%$ and $\approx 14.17\%$ accuracy improvement over Standard-FedQNN and FHE-FedQNN, respectively, on the CIFAR-10 dataset. Additionally, it reduces communication overhead by freezing less important layers, demonstrating the efficiency and practicality of our privacy-preserving, resource-aware design for FL.
Authors: Dongyu Wei, Xiaoren Xu, Shiwen Mao, Mingzhe Chen
Abstract: In this paper, a secure and communication-efficient clustered federated learning (CFL) design is proposed. In our model, several base stations (BSs) with heterogeneous task-handling capabilities and multiple users with non-independent and identically distributed (non-IID) data jointly perform CFL training incorporating differential privacy (DP) techniques. Since each BS can process only a subset of the learning tasks and has limited wireless resource blocks (RBs) to allocate to users for federated learning (FL) model parameter transmission, it is necessary to jointly optimize RB allocation and user scheduling for CFL performance optimization. Meanwhile, our considered CFL method requires devices to use their limited data and FL model information to determine their task identities, which may introduce additional communication overhead. We formulate an optimization problem whose goal is to minimize the training loss of all learning tasks while considering device clustering, RB allocation, DP noise, and FL model transmission delay. To solve the problem, we propose a novel dynamic penalty function assisted value decomposed multi-agent reinforcement learning (DPVD-MARL) algorithm that enables distributed BSs to independently determine their connected users, RBs, and DP noise of the connected users but jointly minimize the training loss of all learning tasks across all BSs. Different from the existing MARL methods that assign a large penalty for invalid actions, we propose a novel penalty assignment scheme that assigns penalty depending on the number of devices that cannot meet communication constraints (e.g., delay), which can guide the MARL scheme to quickly find valid actions, thus improving the convergence speed. Simulation results show that the DPVD-MARL can improve the convergence rate by up to 20% and the ultimate accumulated rewards by 15% compared to independent Q-learning.
Authors: Dongyu Wei, Xiaoren Xu, Yuchen Liu, H. Vincent Poor, Mingzhe Chen
Abstract: In this paper, deceptive signal-assisted private split learning is investigated. In our model, several edge devices jointly perform collaborative training, and some eavesdroppers aim to collect the model and data information from devices. To prevent the eavesdroppers from collecting model and data information, a subset of devices can transmit deceptive signals. Therefore, it is necessary to determine the subset of devices used for deceptive signal transmission, the subset of model training devices, and the models assigned to each model training device. This problem is formulated as an optimization problem whose goal is to minimize the information leaked to eavesdroppers while meeting the model training energy consumption and delay constraints. To solve this problem, we propose a soft actor-critic deep reinforcement learning framework with intrinsic curiosity module and cross-attention (ICM-CA) that enables a centralized agent to determine the model training devices, the deceptive signal transmission devices, the transmit power, and sub-models assigned to each model training device without knowing the position and monitoring probability of eavesdroppers. The proposed method uses an ICM module to encourage the server to explore novel actions and states and a CA module to determine the importance of each historical state-action pair thus improving training efficiency. Simulation results demonstrate that the proposed method improves the convergence rate by up to 3x and reduces the information leaked to eavesdroppers by up to 13% compared to the traditional SAC algorithm.
Authors: Malikussaid, Hilal Hudan Nuha
Abstract: Large Language Models (LLMs) often generate scientifically plausible but factually invalid information, a challenge we term the "plausibility-validity gap," particularly in specialized domains like chemistry. This paper presents a systematic methodology to bridge this gap by developing a specialized scientific assistant. We utilized the Magistral Small model, noted for its integrated reasoning capabilities, and fine-tuned it using Low-Rank Adaptation (LoRA). A key component of our approach was the creation of a "dual-domain dataset," a comprehensive corpus curated from various sources encompassing both molecular properties and chemical reactions, which was standardized to ensure quality. Our evaluation demonstrates that the fine-tuned model achieves significant improvements over the baseline model in format adherence, chemical validity of generated molecules, and the feasibility of proposed synthesis routes. The results indicate a hierarchical learning pattern, where syntactic correctness is learned more readily than chemical possibility and synthesis feasibility. While a comparative analysis with human experts revealed competitive performance in areas like chemical creativity and reasoning, it also highlighted key limitations, including persistent errors in stereochemistry, a static knowledge cutoff, and occasional reference hallucination. This work establishes a viable framework for adapting generalist LLMs into reliable, specialized tools for chemical research, while also delineating critical areas for future improvement.
Authors: Ankit Jyothish, Ali Jannesari
Abstract: Graph transformers typically embed every node in a single Euclidean space, blurring heterogeneous topologies. We prepend a lightweight Riemannian mixture-of-experts layer that routes each node to various kinds of manifold, mixture of spherical, flat, hyperbolic - best matching its local structure. These projections provide intrinsic geometric explanations to the latent space. Inserted into a state-of-the-art ensemble graph transformer, this projector lifts accuracy by up to 3% on four node-classification benchmarks. The ensemble makes sure that both euclidean and non-euclidean features are captured. Explicit, geometry-aware projection thus sharpens predictive power while making graph representations more interpretable.
Authors: James Chapman, Kedar Karhadkar, Guido Montufar
Abstract: Deep reinforcement learning (DRL) has achieved remarkable success across multiple domains, including competitive games, natural language processing, and robotics. Despite these advancements, policies trained via DRL often struggle to generalize to evaluation environments with different parameters. This challenge is typically addressed by training with multiple contexts and/or by leveraging additional structure in the problem. However, obtaining sufficient training data across diverse contexts can be impractical in real-world applications. In this work, we consider contextual Markov decision processes (CMDPs) with transition and reward functions that exhibit regularity in context parameters. We introduce the context-enhanced Bellman equation (CEBE) to improve generalization when training on a single context. We prove both analytically and empirically that the CEBE yields a first-order approximation to the Q-function trained across multiple contexts. We then derive context sample enhancement (CSE) as an efficient data augmentation method for approximating the CEBE in deterministic control environments. We numerically validate the performance of CSE in simulation environments, showcasing its potential to improve generalization in DRL.
Authors: Farnam Mansouri, Shai Ben-David
Abstract: PU (Positive Unlabeled) learning is a variant of supervised classification learning in which the only labels revealed to the learner are of positively labeled instances. PU learning arises in many real-world applications. Most existing work relies on the simplifying assumptions that the positively labeled training data is drawn from the restriction of the data generating distribution to positively labeled instances and/or that the proportion of positively labeled points (a.k.a. the class prior) is known apriori to the learner. This paper provides a theoretical analysis of the statistical complexity of PU learning under a wider range of setups. Unlike most prior work, our study does not assume that the class prior is known to the learner. We prove upper and lower bounds on the required sample sizes (of both the positively labeled and the unlabeled samples).
Authors: Zheyu Zhang (University of Michigan), Jiayuan Dong (University of Michigan), Jie Liu (University of Michigan), Xun Huan (University of Michigan)
Abstract: We present GO-CBED, a goal-oriented Bayesian framework for sequential causal experimental design. Unlike conventional approaches that select interventions aimed at inferring the full causal model, GO-CBED directly maximizes the expected information gain (EIG) on user-specified causal quantities of interest, enabling more targeted and efficient experimentation. The framework is both non-myopic, optimizing over entire intervention sequences, and goal-oriented, targeting only model aspects relevant to the causal query. To address the intractability of exact EIG computation, we introduce a variational lower bound estimator, optimized jointly through a transformer-based policy network and normalizing flow-based variational posteriors. The resulting policy enables real-time decision-making via an amortized network. We demonstrate that GO-CBED consistently outperforms existing baselines across various causal reasoning and discovery tasks-including synthetic structural causal models and semi-synthetic gene regulatory networks-particularly in settings with limited experimental budgets and complex causal mechanisms. Our results highlight the benefits of aligning experimental design objectives with specific research goals and of forward-looking sequential planning.
Authors: Irsyad Adam, Steven Swee, Erika Yilin, Ethan Ji, William Speier, Dean Wang, Alex Bui, Wei Wang, Karol Watson, Peipei Ping
Abstract: In this work, we study the problem pertaining to personalized classification of subclinical atherosclerosis by developing a hierarchical graph neural network framework to leverage two characteristic modalities of a patient: clinical features within the context of the cohort, and molecular data unique to individual patients. Current graph-based methods for disease classification detect patient-specific molecular fingerprints, but lack consistency and comprehension regarding cohort-wide features, which are an essential requirement for understanding pathogenic phenotypes across diverse atherosclerotic trajectories. Furthermore, understanding patient subtypes often considers clinical feature similarity in isolation, without integration of shared pathogenic interdependencies among patients. To address these challenges, we introduce ATHENA: Atherosclerosis Through Hierarchical Explainable Neural Network Analysis, which constructs a novel hierarchical network representation through integrated modality learning; subsequently, it optimizes learned patient-specific molecular fingerprints that reflect individual omics data, enforcing consistency with cohort-wide patterns. With a primary clinical dataset of 391 patients, we demonstrate that this heterogeneous alignment of clinical features with molecular interaction patterns has significantly boosted subclinical atherosclerosis classification performance across various baselines by up to 13% in area under the receiver operating curve (AUC) and 20% in F1 score. Taken together, ATHENA enables mechanistically-informed patient subtype discovery through explainable AI (XAI)-driven subnetwork clustering; this novel integration framework strengthens personalized intervention strategies, thereby improving the prediction of atherosclerotic disease progression and management of their clinical actionable outcomes.
Authors: Zhiwei Zhang, Hui Liu, Xiaomin Li, Zhenwei Dai, Jingying Zeng, Fali Wang, Minhua Lin, Ramraj Chandradevan, Zhen Li, Chen Luo, Xianfeng Tang, Qi He, Suhang Wang
Abstract: Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the reward function rather than genuinely learning the intended behavior. Although significant efforts have been made to mitigate reward hacking, they predominantly focus on and evaluate in-distribution scenarios, where the training and testing data for the reward model share the same distribution. In this paper, we empirically show that state-of-the-art methods struggle in more challenging out-of-distribution (OOD) settings. We further demonstrate that incorporating fine-grained multi-attribute scores helps address this challenge. However, the limited availability of high-quality data often leads to weak performance of multi-objective reward functions, which can negatively impact overall performance and become the bottleneck. To address this issue, we propose a unified reward modeling framework that jointly trains Bradley--Terry (BT) single-objective and multi-objective regression-based reward functions using a shared embedding space. We theoretically establish a connection between the BT loss and the regression objective and highlight their complementary benefits. Specifically, the regression task enhances the single-objective reward function's ability to mitigate reward hacking in challenging OOD settings, while BT-based training improves the scoring capability of the multi-objective reward function, enabling a 7B model to outperform a 70B baseline. Extensive experimental results demonstrate that our framework significantly improves both the robustness and the scoring performance of reward models.
Authors: Zesheng Liu, Maryam Rahnemoonfar
Abstract: Gaining a deeper understanding of the thickness and variability of internal ice layers in Radar imagery is essential in monitoring the snow accumulation, better evaluating ice dynamics processes, and minimizing uncertainties in climate models. Radar sensors, capable of penetrating ice, capture detailed radargram images of internal ice layers. In this work, we introduce GRIT, graph transformer for ice layer thickness. GRIT integrates an inductive geometric graph learning framework with an attention mechanism, designed to map the relationships between shallow and deeper ice layers. Compared to baseline graph neural networks, GRIT demonstrates consistently lower prediction errors. These results highlight the attention mechanism's effectiveness in capturing temporal changes across ice layers, while the graph transformer combines the strengths of transformers for learning long-range dependencies with graph neural networks for capturing spatial patterns, enabling robust modeling of complex spatiotemporal dynamics.
Authors: Zesheng Liu, Maryam Rahnemoonfar
Abstract: Understanding the thickness and variability of internal ice layers in radar imagery is crucial for monitoring snow accumulation, assessing ice dynamics, and reducing uncertainties in climate models. Radar sensors, capable of penetrating ice, provide detailed radargram images of these internal layers. In this work, we present ST-GRIT, a spatio-temporal graph transformer for ice layer thickness, designed to process these radargrams and capture the spatiotemporal relationships between shallow and deep ice layers. ST-GRIT leverages an inductive geometric graph learning framework to extract local spatial features as feature embeddings and employs a series of temporal and spatial attention blocks separately to model long-range dependencies effectively in both dimensions. Experimental evaluation on radargram data from the Greenland ice sheet demonstrates that ST-GRIT consistently outperforms current state-of-the-art methods and other baseline graph neural networks by achieving lower root mean-squared error. These results highlight the advantages of self-attention mechanisms on graphs over pure graph neural networks, including the ability to handle noise, avoid oversmoothing, and capture long-range dependencies. Moreover, the use of separate spatial and temporal attention blocks allows for distinct and robust learning of spatial relationships and temporal patterns, providing a more comprehensive and effective approach.
Authors: Seonghyun Park, Kiyoung Seong, Soojung Yang, Rafael G\'omez-Bombarelli, Sungsoo Ahn
Abstract: Rare events such as state transitions are difficult to observe directly with molecular dynamics simulations due to long timescales. Enhanced sampling techniques overcome this by introducing biases along carefully chosen low-dimensional features, known as collective variables (CVs), which capture the slow degrees of freedom. Machine learning approaches (MLCVs) have automated CV discovery, but existing methods typically focus on discriminating meta-stable states without fully encoding the detailed dynamics essential for accurate sampling. We propose TLC, a framework that learns CVs directly from time-lagged conditions of a generative model. Instead of modeling the static Boltzmann distribution, TLC models a time-lagged conditional distribution yielding CVs to capture the slow dynamic behavior. We validate TLC on the Alanine Dipeptide system using two CV-based enhanced sampling tasks: (i) steered molecular dynamics (SMD) and (ii) on-the-fly probability enhanced sampling (OPES), demonstrating equal or superior performance compared to existing MLCV methods in both transition path sampling and state discrimination.
Authors: Yuntian Liu, Tao Zhu, Xiaoyang Liu, Yu Chen, Zhaoxuan Liu, Qingfeng Guo, Jiashuo Zhang, Kangjie Bao, Tao Luo
Abstract: Statement autoformalization, the automated translation of statement from natural language into formal languages, has become a subject of extensive research, yet the development of robust automated evaluation metrics remains limited. Existing evaluation methods often lack semantic understanding, face challenges with high computational costs, and are constrained by the current progress of automated theorem proving. To address these issues, we propose GTED (Generalized Tree Edit Distance), a novel evaluation framework that first standardizes formal statements and converts them into operator trees, then determines the semantic similarity using the eponymous GTED metric. On the miniF2F and ProofNet benchmarks, GTED outperforms all baseline metrics by achieving the highest accuracy and Kappa scores, thus providing the community with a more faithful metric for automated evaluation. The code and experimental results are available at https://github.com/XiaoyangLiu-sjtu/GTED.
Authors: Pengfei Jiao, Jialong Ni, Di Jin, Xuan Guo, Huan Liu, Hongjiang Chen, Yanxian Bi
Abstract: The pre-training and fine-tuning methods have gained widespread attention in the field of heterogeneous graph neural networks due to their ability to leverage large amounts of unlabeled data during the pre-training phase, allowing the model to learn rich structural features. However, these methods face the issue of a mismatch between the pre-trained model and downstream tasks, leading to suboptimal performance in certain application scenarios. Prompt learning methods have emerged as a new direction in heterogeneous graph tasks, as they allow flexible adaptation of task representations to address target inconsistency. Building on this idea, this paper proposes a novel multi-task prompt framework for the heterogeneous graph domain, named HGMP. First, to bridge the gap between the pre-trained model and downstream tasks, we reformulate all downstream tasks into a unified graph-level task format. Next, we address the limitations of existing graph prompt learning methods, which struggle to integrate contrastive pre-training strategies in the heterogeneous graph domain. We design a graph-level contrastive pre-training strategy to better leverage heterogeneous information and enhance performance in multi-task scenarios. Finally, we introduce heterogeneous feature prompts, which enhance model performance by refining the representation of input graph features. Experimental results on public datasets show that our proposed method adapts well to various tasks and significantly outperforms baseline methods.
Authors: Paul M. Riechers, Thomas J. Elliott, Adam S. Shai
Abstract: We show that deep neural networks, including transformers and RNNs, pretrained as usual on next-token prediction, intrinsically discover and represent beliefs over 'quantum' and 'post-quantum' low-dimensional generative models of their training data -- as if performing iterative Bayesian updates over the latent state of this world model during inference as they observe more context. Notably, neural nets easily find these representation whereas there is no finite classical circuit that would do the job. The corresponding geometric relationships among neural activations induced by different input sequences are found to be largely independent of neural-network architecture. Each point in this geometry corresponds to a history-induced probability density over all possible futures, and the relative displacement of these points reflects the difference in mechanism and magnitude for how these distinct pasts affect the future.
Authors: Nawaf Alampara, Anagha Aneesh, Marti\~no R\'ios-Garc\'ia, Adrian Mirza, Mara Schilling-Wilhelmi, Ali Asghar Aghajani, Meiling Sun, Gordan Prastalo, Kevin Maik Jablonka
Abstract: Data-driven techniques have a large potential to transform and accelerate the chemical sciences. However, chemical sciences also pose the unique challenge of very diverse, small, fuzzy datasets that are difficult to leverage in conventional machine learning approaches completely. A new class of models, general-purpose models (GPMs) such as large language models, have shown the ability to solve tasks they have not been directly trained on, and to flexibly operate with low amounts of data in different formats. In this review, we discuss fundamental building principles of GPMs and review recent applications of those models in the chemical sciences across the entire scientific process. While many of these applications are still in the prototype phase, we expect that the increasing interest in GPMs will make many of them mature in the coming years.
Authors: Wooseong Jeong, Kuk-Jin Yoon
Abstract: Multi-Task Learning (MTL) enables multiple tasks to be learned within a shared network, but differences in objectives across tasks can cause negative transfer, where the learning of one task degrades another task's performance. While pre-trained transformers significantly improve MTL performance, their fixed network capacity and rigid structure limit adaptability. Previous dynamic network architectures attempt to address this but are inefficient as they directly convert shared parameters into task-specific ones. We propose Dynamic Token Modulation and Expansion (DTME-MTL), a framework applicable to any transformer-based MTL architecture. DTME-MTL enhances adaptability and reduces overfitting by identifying gradient conflicts in token space and applying adaptive solutions based on conflict type. Unlike prior methods that mitigate negative transfer by duplicating network parameters, DTME-MTL operates entirely in token space, enabling efficient adaptation without excessive parameter growth. Extensive experiments demonstrate that DTME-MTL consistently improves multi-task performance with minimal computational overhead, offering a scalable and effective solution for enhancing transformer-based MTL models.
Authors: Joris Suurmeijer, Ivo Pascal de Jong, Matias Valdenegro-Toro, Andreea Ioana Sburlea
Abstract: Brain-computer interfaces (BCIs) turn brain signals into functionally useful output, but they are not always accurate. A good Machine Learning classifier should be able to indicate how confident it is about a given classification, by giving a probability for its classification. Standard classifiers for Motor Imagery BCIs do give such probabilities, but research on uncertainty quantification has been limited to Deep Learning. We compare the uncertainty quantification ability of established BCI classifiers using Common Spatial Patterns (CSP-LDA) and Riemannian Geometry (MDRM) to specialized methods in Deep Learning (Deep Ensembles and Direct Uncertainty Quantification) as well as standard Convolutional Neural Networks (CNNs). We found that the overconfidence typically seen in Deep Learning is not a problem in CSP-LDA and MDRM. We found that MDRM is underconfident, which we solved by adding Temperature Scaling (MDRM-T). CSP-LDA and MDRM-T give the best uncertainty estimates, but Deep Ensembles and standard CNNs give the best classifications. We show that all models are able to separate between easy and difficult estimates, so that we can increase the accuracy of a Motor Imagery BCI by rejecting samples that are ambiguous.
Authors: Berkant Turan, Suhrab Asadulla, David Steinmann, Wolfgang Stammer, Sebastian Pokutta
Abstract: While Prover-Verifier Games (PVGs) offer a promising path toward verifiability in nonlinear classification models, they have not yet been applied to complex inputs such as high-dimensional images. Conversely, Concept Bottleneck Models (CBMs) effectively translate such data into interpretable concepts but are limited by their reliance on low-capacity linear predictors. In this work, we introduce the Neural Concept Verifier (NCV), a unified framework combining PVGs with concept encodings for interpretable, nonlinear classification in high-dimensional settings. NCV achieves this by utilizing recent minimally supervised concept discovery models to extract structured concept encodings from raw inputs. A prover then selects a subset of these encodings, which a verifier -- implemented as a nonlinear predictor -- uses exclusively for decision-making. Our evaluations show that NCV outperforms CBM and pixel-based PVG classifier baselines on high-dimensional, logically complex datasets and also helps mitigate shortcut behavior. Overall, we demonstrate NCV as a promising step toward performative, verifiable AI.
Authors: Amirhossein Sadough, Mahyar Shahsavari, Mark Wijtvliet, Marcel van Gerven
Abstract: Anomaly detection (AD) plays a vital role across a wide range of real-world domains by identifying data instances that deviate from expected patterns, potentially signaling critical events such as system failures, fraudulent activities, or rare medical conditions. The demand for real-time AD has surged with the rise of the (Industrial) Internet of Things, where massive volumes of multivariate sensor data must be processed instantaneously. Real-time AD requires methods that not only handle high-dimensional streaming data but also operate in a single-pass manner, without the burden of storing historical instances, thereby ensuring minimal memory usage and fast decision-making. We propose DAD, a novel real-time decorrelation-based anomaly detection method for multivariate time series, based on an online decorrelation learning approach. Unlike traditional proximity-based or reconstruction-based detectors that process entire data or windowed instances, DAD dynamically learns and monitors the correlation structure of data sample by sample in a single pass, enabling efficient and effective detection. To support more realistic benchmarking practices, we also introduce a practical hyperparameter tuning strategy tailored for real-time anomaly detection scenarios. Extensive experiments on widely used benchmark datasets demonstrate that DAD achieves the most consistent and superior performance across diverse anomaly types compared to state-of-the-art methods. Crucially, its robustness to increasing dimensionality makes it particularly well-suited for real-time, high-dimensional data streams. Ultimately, DAD not only strikes an optimal balance between detection efficacy and computational efficiency but also sets a new standard for real-time, memory-constrained anomaly detection.
Authors: Uliana Parkina, Maxim Rakhuba
Abstract: Recent studies suggest that context-aware low-rank approximation is a useful tool for compression and fine-tuning of modern large-scale neural networks. In this type of approximation, a norm is weighted by a matrix of input activations, significantly improving metrics over the unweighted case. Nevertheless, existing methods for neural networks suffer from numerical instabilities due to their reliance on classical formulas involving explicit Gram matrix computation and their subsequent inversion. We demonstrate that this can degrade the approximation quality or cause numerically singular matrices. To address these limitations, we propose a novel inversion-free regularized framework that is based entirely on stable decompositions and overcomes the numerical pitfalls of prior art. Our method can handle possible challenging scenarios: (1) when calibration matrices exceed GPU memory capacity, (2) when input activation matrices are nearly singular, and even (3) when insufficient data prevents unique approximation. For the latter, we prove that our solution converges to a desired approximation and derive explicit error bounds.
Authors: Michail Kalntis, Fernando A. Kuipers, George Iosifidis
Abstract: Handovers (HOs) are the cornerstone of modern cellular networks for enabling seamless connectivity to a vast and diverse number of mobile users. However, as mobile networks become more complex with more diverse users and smaller cells, traditional HOs face significant challenges, such as prolonged delays and increased failures. To mitigate these issues, 3GPP introduced conditional handovers (CHOs), a new type of HO that enables the preparation (i.e., resource allocation) of multiple cells for a single user to increase the chance of HO success and decrease the delays in the procedure. Despite its advantages, CHO introduces new challenges that must be addressed, including efficient resource allocation and managing signaling/communication overhead from frequent cell preparations and releases. This paper presents a novel framework aligned with the O-RAN paradigm that leverages meta-learning for CHO optimization, providing robust dynamic regret guarantees and demonstrating at least 180% superior performance than other 3GPP benchmarks in volatile signal conditions.
Authors: Iago Xabier V\'azquez Garc\'ia, Damla Partanaz, Emrullah Fatih Yetkin
Abstract: In this study, we focused on proposing an optimal clustering mechanism for the occupations defined in the well-known US-based occupational database, O*NET. Even though all occupations are defined according to well-conducted surveys in the US, their definitions can vary for different firms and countries. Hence, if one wants to expand the data that is already collected in O*NET for the occupations defined with different tasks, a map between the definitions will be a vital requirement. We proposed a pipeline using several BERT-based techniques with various clustering approaches to obtain such a map. We also examined the effect of dimensionality reduction approaches on several metrics used in measuring performance of clustering algorithms. Finally, we improved our results by using a specialized silhouette approach. This new clustering-based mapping approach with dimensionality reduction may help distinguish the occupations automatically, creating new paths for people wanting to change their careers.
Authors: Arpana Sinhal, Anay Sinhal, Amit Sinhal
Abstract: Healthcare professionals, particularly nurses, face elevated occupational stress, a concern amplified during the COVID-19 pandemic. While wearable sensors offer promising avenues for real-time stress monitoring, existing studies often lack comprehensive datasets and robust analytical frameworks. This study addresses these gaps by introducing a multimodal dataset comprising physiological signals, electrodermal activity, heart rate and skin temperature. A systematic literature review identified limitations in prior stress-detection methodologies, particularly in handling class imbalance and optimizing model generalizability. To overcome these challenges, the dataset underwent preprocessing with the Synthetic Minority Over sampling Technique (SMOTE), ensuring balanced representation of stress states. Advanced machine learning models including Random Forest, XGBoost and a Multi-Layer Perceptron (MLP) were evaluated and combined into a Stacking Classifier to leverage their collective predictive strengths. By using a publicly accessible dataset and a reproducible analytical pipeline, this work advances the development of deployable stress-monitoring systems, offering practical implications for safeguarding healthcare workers' mental health. Future research directions include expanding demographic diversity and exploring edge-computing implementations for low latency stress alerts.
Authors: Sebastian Lotter, Elisabeth Mohr, Andrina Rutsch, Lukas Brand, Francesca Ronchi, Laura D\'iaz-Marug\'an
Abstract: Synthetic molecular communication (SMC) is a key enabler for future healthcare systems in which Internet of Bio-Nano-Things (IoBNT) devices facilitate the continuous monitoring of a patient's biochemical signals. To close the loop between sensing and actuation, both the detection and the generation of in-body molecular communication (MC) signals is key. However, generating signals inside the human body, e.g., via synthetic nanodevices, poses a challenge in SMC, due to technological obstacles as well as legal, safety, and ethical issues. Hence, this paper considers an SMC system in which signals are generated indirectly via the modulation of a natural in-body MC system, namely the gut-brain axis (GBA). Therapeutic GBA modulation is already established as treatment for neurological diseases, e.g., drug refractory epilepsy (DRE), and performed via the administration of nutritional supplements or specific diets. However, the molecular signaling pathways that mediate the effect of such treatments are mostly unknown. Consequently, existing treatments are standardized or designed heuristically and able to help only some patients while failing to help others. In this paper, we propose to leverage personal health data, e.g., gathered by in-body IoBNT devices, to design more versatile and robust GBA modulation-based treatments as compared to the existing ones. To show the feasibility of our approach, we define a catalog of theoretical requirements for therapeutic GBA modulation. Then, we propose a machine learning model to verify these requirements for practical scenarios when only limited data on the GBA modulation exists. By evaluating the proposed model on several datasets, we confirm its excellent accuracy in identifying different modulators of the GBA. Finally, we utilize the proposed model to identify specific modulatory pathways that play an important role for therapeutic GBA modulation.
Authors: Davide Domini, Laura Erhan, Gianluca Aguzzi, Lucia Cavallaro, Amirhossein Douzandeh Zenoozi, Antonio Liotta, Mirko Viroli
Abstract: Federated Learning offers privacy-preserving collaborative intelligence but struggles to meet the sustainability demands of emerging IoT ecosystems necessary for Society 5.0-a human-centered technological future balancing social advancement with environmental responsibility. The excessive communication bandwidth and computational resources required by traditional FL approaches make them environmentally unsustainable at scale, creating a fundamental conflict with green AI principles as billions of resource-constrained devices attempt to participate. To this end, we introduce Sparse Proximity-based Self-Federated Learning (SParSeFuL), a resource-aware approach that bridges this gap by combining aggregate computing for self-organization with neural network sparsification to reduce energy and bandwidth consumption.
Authors: Junyu Luo, Yuhao Tang, Yiwei Fu, Xiao Luo, Zhizhuo Kou, Zhiping Xiao, Wei Ju, Wentao Zhang, Ming Zhang
Abstract: Unsupervised Graph Domain Adaptation (UGDA) leverages labeled source domain graphs to achieve effective performance in unlabeled target domains despite distribution shifts. However, existing methods often yield suboptimal results due to the entanglement of causal-spurious features and the failure of global alignment strategies. We propose SLOGAN (Sparse Causal Discovery with Generative Intervention), a novel approach that achieves stable graph representation transfer through sparse causal modeling and dynamic intervention mechanisms. Specifically, SLOGAN first constructs a sparse causal graph structure, leveraging mutual information bottleneck constraints to disentangle sparse, stable causal features while compressing domain-dependent spurious correlations through variational inference. To address residual spurious correlations, we innovatively design a generative intervention mechanism that breaks local spurious couplings through cross-domain feature recombination while maintaining causal feature semantic consistency via covariance constraints. Furthermore, to mitigate error accumulation in target domain pseudo-labels, we introduce a category-adaptive dynamic calibration strategy, ensuring stable discriminative learning. Extensive experiments on multiple real-world datasets demonstrate that SLOGAN significantly outperforms existing baselines.
Authors: Federico Del Pup, Riccardo Brun, Filippo Iotti, Edoardo Paccagnella, Mattia Pezzato, Sabrina Bertozzo, Andrea Zanola, Louis Fabrice Tshimanga, Henning M\"uller, Manfredo Atzori
Abstract: Electroencephalography (EEG) is establishing itself as an important, low-cost, noninvasive diagnostic tool for the early detection of Parkinson's Disease (PD). In this context, EEG-based Deep Learning (DL) models have shown promising results due to their ability to discover highly nonlinear patterns within the signal. However, current state-of-the-art DL models suffer from poor generalizability caused by high inter-subject variability. This high variability underscores the need for enhancing model generalizability by developing new architectures better tailored to EEG data. This paper introduces TransformEEG, a hybrid Convolutional-Transformer designed for Parkinson's disease detection using EEG data. Unlike transformer models based on the EEGNet structure, TransformEEG incorporates a depthwise convolutional tokenizer. This tokenizer is specialized in generating tokens composed by channel-specific features, which enables more effective feature mixing within the self-attention layers of the transformer encoder. To evaluate the proposed model, four public datasets comprising 290 subjects (140 PD patients, 150 healthy controls) were harmonized and aggregated. A 10-outer, 10-inner Nested-Leave-N-Subjects-Out (N-LNSO) cross-validation was performed to provide an unbiased comparison against seven other consolidated EEG deep learning models. TransformEEG achieved the highest balanced accuracy's median (78.45%) as well as the lowest interquartile range (6.37%) across all the N-LNSO partitions. When combined with data augmentation and threshold correction, median accuracy increased to 80.10%, with an interquartile range of 5.74%. In conclusion, TransformEEG produces more consistent and less skewed results. It demonstrates a substantial reduction in variability and more reliable PD detection using EEG data compared to the other investigated models.
Authors: Carlos Beis Penedo, Rebeca P. D\'iaz Redondo, Ana Fern\'andez Vilas, Manuel Fern\'andez Veiga, Francisco Troncoso Pastoriza
Abstract: Collaborative machine learning in sensitive domains demands scalable, privacy preserving solutions for enterprise deployment. Conventional Federated Learning (FL) relies on a central server, introducing single points of failure and privacy risks, while Split Learning (SL) partitions models for privacy but scales poorly due to sequential training. We present a decentralized architecture that combines Federated Split Learning (FSL) with the permissioned blockchain Hyperledger Fabric (HLF). Our chaincode orchestrates FSL's split model execution and peer-to-peer aggregation without any central coordinator, leveraging HLF's transient fields and Private Data Collections (PDCs) to keep raw data and model activations private. On CIFAR-10 and MNIST benchmarks, HLF-FSL matches centralized FSL accuracy while reducing per epoch training time compared to Ethereum-based works. Performance and scalability tests show minimal blockchain overhead and preserved accuracy, demonstrating enterprise grade viability.
Authors: Darshan Makwana
Abstract: We analyze the layerwise effective dimension (rank of the feature matrix) in fully-connected ReLU networks of finite width. Specifically, for a fixed batch of $m$ inputs and random Gaussian weights, we derive closed-form expressions for the expected rank of the \$m\times n\$ hidden activation matrices. Our main result shows that $\mathbb{E}[EDim(\ell)]=m[1-(1-2/\pi)^\ell]+O(e^{-c m})$ so that the rank deficit decays geometrically with ratio $1-2 / \pi \approx 0.3634$. We also prove a sub-Gaussian concentration bound, and identify the "revival" depths at which the expected rank attains local maxima. In particular, these peaks occur at depths $\ell_k^*\approx(k+1/2)\pi/\log(1/\rho)$ with height $\approx (1-e^{-\pi/2}) m \approx 0.79m$. We further show that this oscillatory rank behavior is a finite-width phenomenon: under orthogonal weight initialization or strong negative-slope leaky-ReLU, the rank remains (nearly) full. These results provide a precise characterization of how random ReLU layers alternately collapse and partially revive the subspace of input variations, adding nuance to prior work on expressivity of deep networks.
Authors: Zhuang Qi, Lei Meng, Han Yu
Abstract: Federated Class Incremental Learning (FCIL) aims to collaboratively process continuously increasing incoming tasks across multiple clients. Among various approaches, data replay has become a promising solution, which can alleviate forgetting by reintroducing representative samples from previous tasks. However, their performance is typically limited by class imbalance, both within the replay buffer due to limited global awareness and between replayed and newly arrived classes. To address this issue, we propose a class wise balancing data replay method for FCIL (FedCBDR), which employs a global coordination mechanism for class-level memory construction and reweights the learning objective to alleviate the aforementioned imbalances. Specifically, FedCBDR has two key components: 1) the global-perspective data replay module reconstructs global representations of prior task in a privacy-preserving manner, which then guides a class-aware and importance-sensitive sampling strategy to achieve balanced replay; 2) Subsequently, to handle class imbalance across tasks, the task aware temperature scaling module adaptively adjusts the temperature of logits at both class and instance levels based on task dynamics, which reduces the model's overconfidence in majority classes while enhancing its sensitivity to minority classes. Experimental results verified that FedCBDR achieves balanced class-wise sampling under heterogeneous data distributions and improves generalization under task imbalance between earlier and recent tasks, yielding a 2%-15% Top-1 accuracy improvement over six state-of-the-art methods.
Authors: Peiyan Zhang, Haibo Jin, Liying Kang, Haohan Wang
Abstract: Jailbreak attacks reveal critical vulnerabilities in Large Language Models (LLMs) by causing them to generate harmful or unethical content. Evaluating these threats is particularly challenging due to the evolving nature of LLMs and the sophistication required in effectively probing their vulnerabilities. Current benchmarks and evaluation methods struggle to fully address these challenges, leaving gaps in the assessment of LLM vulnerabilities. In this paper, we review existing jailbreak evaluation practices and identify three assumed desiderata for an effective jailbreak evaluation protocol. To address these challenges, we introduce GuardVal, a new evaluation protocol that dynamically generates and refines jailbreak prompts based on the defender LLM's state, providing a more accurate assessment of defender LLMs' capacity to handle safety-critical situations. Moreover, we propose a new optimization method that prevents stagnation during prompt refinement, ensuring the generation of increasingly effective jailbreak prompts that expose deeper weaknesses in the defender LLMs. We apply this protocol to a diverse set of models, from Mistral-7b to GPT-4, across 10 safety domains. Our findings highlight distinct behavioral patterns among the models, offering a comprehensive view of their robustness. Furthermore, our evaluation process deepens the understanding of LLM behavior, leading to insights that can inform future research and drive the development of more secure models.
Authors: Tomu Hirata, Undral Byambadalai, Tatsushi Oka, Shota Yasui, Shingo Uto
Abstract: We propose a novel multi-task neural network approach for estimating distributional treatment effects (DTE) in randomized experiments. While DTE provides more granular insights into the experiment outcomes over conventional methods focusing on the Average Treatment Effect (ATE), estimating it with regression adjustment methods presents significant challenges. Specifically, precision in the distribution tails suffers due to data imbalance, and computational inefficiencies arise from the need to solve numerous regression problems, particularly in large-scale datasets commonly encountered in industry. To address these limitations, our method leverages multi-task neural networks to estimate conditional outcome distributions while incorporating monotonic shape constraints and multi-threshold label learning to enhance accuracy. To demonstrate the practical effectiveness of our proposed method, we apply our method to both simulated and real-world datasets, including a randomized field experiment aimed at reducing water consumption in the US and a large-scale A/B test from a leading streaming platform in Japan. The experimental results consistently demonstrate superior performance across various datasets, establishing our method as a robust and practical solution for modern causal inference applications requiring a detailed understanding of treatment effect heterogeneity.
Authors: Jaeheun Jung, Bosung Jung, Suhyun Bae, Donghun Lee
Abstract: Machine unlearning seeks to remove the influence of particular data or class from trained models to meet privacy, legal, or ethical requirements. Existing unlearning methods tend to forget shallowly: phenomenon of an unlearned model pretend to forget by adjusting only the model response, while its internal representations retain information sufficiently to restore the forgotten data or behavior. We empirically confirm the widespread shallowness by reverting the forgetting effect of various unlearning methods via training-free performance recovery attack and gradient-inversion-based data reconstruction attack. To address this vulnerability fundamentally, we define a theoretical criterion of ``deep forgetting'' based on one-point-contraction of feature representations of data to forget. We also propose an efficient approximation algorithm, and use it to construct a novel general-purpose unlearning algorithm: One-Point-Contraction (OPC). Empirical evaluations on image classification unlearning benchmarks show that OPC achieves not only effective unlearning performance but also superior resilience against both performance recovery attack and gradient-inversion attack. The distinctive unlearning performance of OPC arises from the deep feature forgetting enforced by its theoretical foundation, and recaps the need for improved robustness of machine unlearning methods.
Authors: Tejaswini Medi, Steffen Jung, Margret Keuper
Abstract: Adversarial Training (AT) is a widely adopted defense against adversarial examples. However, existing approaches typically apply a uniform training objective across all classes, overlooking disparities in class-wise vulnerability. This results in adversarial unfairness: classes with well distinguishable features (strong classes) tend to become more robust, while classes with overlapping or shared features(weak classes) remain disproportionately susceptible to adversarial attacks. We observe that strong classes do not require strong adversaries during training, as their non-robust features are quickly suppressed. In contrast, weak classes benefit from stronger adversaries to effectively reduce their vulnerabilities. Motivated by this, we introduce TRIX, a feature-aware adversarial training framework that adaptively assigns weaker targeted adversaries to strong classes, promoting feature diversity via uniformly sampled targets, and stronger untargeted adversaries to weak classes, enhancing their focused robustness. TRIX further incorporates per-class loss weighting and perturbation strength adjustments, building on prior work, to emphasize weak classes during the optimization. Comprehensive experiments on standard image classification benchmarks, including evaluations under strong attacks such as PGD and AutoAttack, demonstrate that TRIX significantly improves worst-case class accuracy on both clean and adversarial data, reducing inter-class robustness disparities, and preserves overall accuracy. Our results highlight TRIX as a practical step toward fair and effective adversarial defense.
Authors: Ruohong Liu, Jack Umenberger, Yize Chen
Abstract: Recent years have seen significant advancements in designing reinforcement learning (RL)-based agents for building energy management. While individual success is observed in simulated or controlled environments, the scalability of RL approaches in terms of efficiency and generalization across building dynamics and operational scenarios remains an open question. In this work, we formally characterize the generalization space for the cross-environment, multi-objective building energy management task, and formulate the multi-objective contextual RL problem. Such a formulation helps understand the challenges of transferring learned policies across varied operational contexts such as climate and heat convection dynamics under multiple control objectives such as comfort level and energy consumption. We provide a principled way to parameterize such contextual information in realistic building RL environments, and construct a novel benchmark to facilitate the evaluation of generalizable RL algorithms in practical building control tasks. Our results show that existing multi-objective RL methods are capable of achieving reasonable trade-offs between conflicting objectives. However, their performance degrades under certain environment variations, underscoring the importance of incorporating dynamics-dependent contextual information into the policy learning process.
Authors: Wooseong Jeong, Jegyeong Cho, Youngho Yoon, Kuk-Jin Yoon
Abstract: Generalizing neural networks to unseen target domains is a significant challenge in real-world deployments. Test-time training (TTT) addresses this by using an auxiliary self-supervised task to reduce the domain gap caused by distribution shifts between the source and target. However, we find that when models are required to perform multiple tasks under domain shifts, conventional TTT methods suffer from unsynchronized task behavior, where the adaptation steps needed for optimal performance in one task may not align with the requirements of other tasks. To address this, we propose a novel TTT approach called Synchronizing Tasks for Test-time Training (S4T), which enables the concurrent handling of multiple tasks. The core idea behind S4T is that predicting task relations across domain shifts is key to synchronizing tasks during test time. To validate our approach, we apply S4T to conventional multi-task benchmarks, integrating it with traditional TTT protocols. Our empirical results show that S4T outperforms state-of-the-art TTT methods across various benchmarks.
Authors: Hermann Klein, Max Heinz Herkersdorf, Oliver Nelles
Abstract: The state space dynamics representation is the most general approach for nonlinear systems and often chosen for system identification. During training, the state trajectory can deform significantly leading to poor data coverage of the state space. This can cause significant issues for space-oriented training algorithms which e.g. rely on grid structures, tree partitioning, or similar. Besides hindering training, significant state trajectory deformations also deteriorate interpretability and robustness properties. This paper proposes a new type of space-filling regularization that ensures a favorable data distribution in state space via introducing a data-distribution-based penalty. This method is demonstrated in local model network architectures where good interpretability is a major concern. The proposed approach integrates ideas from modeling and design of experiments for state space structures. This is why we present two regularization techniques for the data point distributions of the state trajectories for local affine state space models. Beyond that, we demonstrate the results on a widely known system identification benchmark.
Authors: Alba Garrido, Alejandro Almod\'ovar, Patricia A. Apell\'aniz, Juan Parras, Santiago Zazo
Abstract: Accurate survival prediction is critical in oncology for prognosis and treatment planning. Traditional approaches often rely on a single data modality, limiting their ability to capture the complexity of tumor biology. To address this challenge, we introduce a multimodal deep learning framework for survival analysis capable of modeling both single and competing risks scenarios, evaluating the impact of integrating multiple medical data sources on survival predictions. We propose SAMVAE (Survival Analysis Multimodal Variational Autoencoder), a novel deep learning architecture designed for survival prediction that integrates six data modalities: clinical variables, four molecular profiles, and histopathological images. SAMVAE leverages modality specific encoders to project inputs into a shared latent space, enabling robust survival prediction while preserving modality specific information. Its parametric formulation enables the derivation of clinically meaningful statistics from the output distributions, providing patient-specific insights through interactive multimedia that contribute to more informed clinical decision-making and establish a foundation for interpretable, data-driven survival analysis in oncology. We evaluate SAMVAE on two cancer cohorts breast cancer and lower grade glioma applying tailored preprocessing, dimensionality reduction, and hyperparameter optimization. The results demonstrate the successful integration of multimodal data for both standard survival analysis and competing risks scenarios across different datasets. Our model achieves competitive performance compared to state-of-the-art multimodal survival models. Notably, this is the first parametric multimodal deep learning architecture to incorporate competing risks while modeling continuous time to a specific event, using both tabular and image data.
Authors: Nikolay Yudin, Alexander Gaponov, Sergei Kudriashov, Maxim Rakhuba
Abstract: We present a novel local Lipschitz bound for self-attention blocks of transformers. This bound is based on a refined closed-form expression for the spectral norm of the softmax function. The resulting bound is not only more accurate than in the prior art, but also unveils the dependence of the Lipschitz constant on attention score maps. Based on the new findings, we suggest an explanation of the way distributions inside the attention map affect the robustness from the Lipschitz constant perspective. We also introduce a new lightweight regularization term called JaSMin (Jacobian Softmax norm Minimization), which boosts the transformer's robustness and decreases local Lipschitz constants of the whole network.
Authors: Erfan Mirzaei, Andreas Maurer, Vladimir R. Kostic, Massimiliano Pontil
Abstract: Learning from non-independent and non-identically distributed data poses a persistent challenge in statistical learning. In this study, we introduce data-dependent Bernstein inequalities tailored for vector-valued processes in Hilbert space. Our inequalities apply to both stationary and non-stationary processes and exploit the potential rapid decay of correlations between temporally separated variables to improve estimation. We demonstrate the utility of these bounds by applying them to covariance operator estimation in the Hilbert-Schmidt norm and to operator learning in dynamical systems, achieving novel risk bounds. Finally, we perform numerical experiments to illustrate the practical implications of these bounds in both contexts.
Authors: Martin Mr\'az, Breenda Das, Anshul Gupta, Lennart Purucker, Frank Hutter
Abstract: Foundation models for tabular data are rapidly evolving, with increasing interest in extending them to support additional modalities such as free-text features. However, existing benchmarks for tabular data rarely include textual columns, and identifying real-world tabular datasets with semantically rich text features is non-trivial. We propose a series of simple yet effective ablation-style strategies for incorporating text into conventional tabular pipelines. Moreover, we benchmark how state-of-the-art tabular foundation models can handle textual data by manually curating a collection of real-world tabular datasets with meaningful textual features. Our study is an important step towards improving benchmarking of foundation models for tabular data with text.
Authors: Giovanni Dispoto, Paolo Bonetti, Marcello Restelli
Abstract: Recent advances in Reinforcement Learning (RL) largely benefit from the inclusion of Deep Neural Networks, boosting the number of novel approaches proposed in the field of Deep Reinforcement Learning (DRL). These techniques demonstrate the ability to tackle complex games such as Atari, Go, and other real-world applications, including financial trading. Nevertheless, a significant challenge emerges from the lack of interpretability, particularly when attempting to comprehend the underlying patterns learned, the relative importance of the state features, and how they are integrated to generate the policy's output. For this reason, in mission-critical and real-world settings, it is often preferred to deploy a simpler and more interpretable algorithm, although at the cost of performance. In this paper, we propose a novel algorithm, supported by theoretical guarantees, that can extract an interpretable policy (e.g., a linear policy) without disregarding the peculiarities of expert behavior. This result is obtained by considering the advantage function, which includes information about why an action is superior to the others. In contrast to previous works, our approach enables the training of an interpretable policy using previously collected experience. The proposed algorithm is empirically evaluated on classic control environments and on a financial trading scenario, demonstrating its ability to extract meaningful information from complex expert policies.
Authors: Haichen Hu, David Simchi-Levi
Abstract: We study a sequential contextual decision-making problem in which certain covariates are missing but can be imputed using a pre-trained AI model. From a theoretical perspective, we analyze how the presence of such a model influences the regret of the decision-making process. We introduce a novel notion called "model elasticity", which quantifies the sensitivity of the reward function to the discrepancy between the true covariate and its imputed counterpart. This concept provides a unified way to characterize the regret incurred due to model imputation, regardless of the underlying missingness mechanism. More surprisingly, we show that under the missing at random (MAR) setting, it is possible to sequentially calibrate the pre-trained model using tools from orthogonal statistical learning and doubly robust regression. This calibration significantly improves the quality of the imputed covariates, leading to much better regret guarantees. Our analysis highlights the practical value of having an accurate pre-trained model in sequential decision-making tasks and suggests that model elasticity may serve as a fundamental metric for understanding and improving the integration of pre-trained models in a wide range of data-driven decision-making problems.
Authors: Navish Kumar, Thomas M\"ollenhoff, Mohammad Emtiyaz Khan, Aurelien Lucchi
Abstract: Variational inference with natural-gradient descent often shows fast convergence in practice, but its theoretical convergence guarantees have been challenging to establish. This is true even for the simplest cases that involve concave log-likelihoods and use a Gaussian approximation. We show that the challenge can be circumvented for such cases using a square-root parameterization for the Gaussian covariance. This approach establishes novel convergence guarantees for natural-gradient variational-Gaussian inference and its continuous-time gradient flow. Our experiments demonstrate the effectiveness of natural gradient methods and highlight their advantages over algorithms that use Euclidean or Wasserstein geometries.
Authors: Zizhou Zhang, Qinyan Shen, Zhuohuan Hu, Qianying Liu, Huijie Shen
Abstract: Small and Medium-sized Enterprises (SMEs) are vital to the modern economy, yet their credit risk analysis often struggles with scarce data, especially for online lenders lacking direct credit records. This paper introduces a Graph Neural Network (GNN)-based framework, leveraging SME interactions from transaction and social data to map spatial dependencies and predict loan default risks. Tests on real-world datasets from Discover and Ant Credit (23.4M nodes for supply chain analysis, 8.6M for default prediction) show the GNN surpasses traditional and other GNN baselines, with AUCs of 0.995 and 0.701 for supply chain mining and default prediction, respectively. It also helps regulators model supply chain disruption impacts on banks, accurately forecasting loan defaults from material shortages, and offers Federal Reserve stress testers key data for CCAR risk buffers. This approach provides a scalable, effective tool for assessing SME credit risk.
Authors: Wenxuan Zhou, Shujian Zhang, Brice Magdalou, John Lambert, Ehsan Amid, Richard Nock, Andrew Hard
Abstract: In this paper, we show that direct preference optimization (DPO) is a very specific form of a connection between two major theories in the ML context of learning from preferences: loss functions (Savage) and stochastic choice (Doignon-Falmagne and Machina). The connection is established for all of Savage's losses and at this level of generality, (i) it includes support for abstention on the choice theory side, (ii) it includes support for non-convex objectives on the ML side, and (iii) it allows to frame for free some notable extensions of the DPO setting, including margins and corrections for length. Getting to understand how DPO operates from a general principled perspective is crucial because of the huge and diverse application landscape of models, because of the current momentum around DPO, but also -- and importantly -- because many state of the art variations on DPO definitely occupy a small region of the map that we cover. It also helps to understand the pitfalls of departing from this map, and figure out workarounds.
Authors: Tianang Leng, Fangping Wan, Marcelo Der Torossian Torres, Cesar de la Fuente-Nunez
Abstract: Antimicrobial resistance (AMR) is escalating and outpacing current antibiotic development. Thus, discovering antibiotics effective against emerging pathogens is becoming increasingly critical. However, existing approaches cannot rapidly identify effective molecules against novel pathogens or emerging drug-resistant strains. Here, we introduce ApexOracle, an artificial intelligence (AI) model that both predicts the antibacterial potency of existing compounds and designs de novo molecules active against strains it has never encountered. Departing from models that rely solely on molecular features, ApexOracle incorporates pathogen-specific context through the integration of molecular features captured via a foundational discrete diffusion language model and a dual-embedding framework that combines genomic- and literature-derived strain representations. Across diverse bacterial species and chemical modalities, ApexOracle consistently outperformed state-of-the-art approaches in activity prediction and demonstrated reliable transferability to novel pathogens with little or no antimicrobial data. Its unified representation-generation architecture further enables the in silico creation of "new-to-nature" molecules with high predicted efficacy against priority threats. By pairing rapid activity prediction with targeted molecular generation, ApexOracle offers a scalable strategy for countering AMR and preparing for future infectious-disease outbreaks.
Authors: Wei-Tse Hsu, Savva Grevtsev, Thomas Douglas, Aniket Magarkar, Philip C. Biggin
Abstract: We evaluate the feasibility of using co-folding models for synthetic data augmentation in training machine learning-based scoring functions (MLSFs) for binding affinity prediction. Our results show that performance gains depend critically on the structural quality of augmented data. In light of this, we established simple heuristics for identifying high-quality co-folding predictions without reference structures, enabling them to substitute for experimental structures in MLSF training. Our study informs future data augmentation strategies based on co-folding models.
Authors: Hao Ban, Gokul Ram Subramani, Kaiyi Ji
Abstract: Multi-task learning (MTL) enables a joint model to capture commonalities across multiple tasks, reducing computation costs and improving data efficiency. However, a major challenge in MTL optimization is task conflicts, where the task gradients differ in direction or magnitude, limiting model performance compared to single-task counterparts. Sharpness-aware minimization (SAM) minimizes task loss while simultaneously reducing the sharpness of the loss landscape. Our empirical observations show that SAM effectively mitigates task conflicts in MTL. Motivated by these findings, we explore integrating SAM into MTL but face two key challenges. While both the average loss gradient and individual task gradients-referred to as global and local information-contribute to SAM, how to combine them remains unclear. Moreover, directly computing each task gradient introduces significant computational and memory overheads. To address these challenges, we propose SAMO, a lightweight \textbf{S}harpness-\textbf{A}ware \textbf{M}ulti-task \textbf{O}ptimization approach, that leverages a joint global-local perturbation. The local perturbations are approximated using only forward passes and are layerwise normalized to improve efficiency. Extensive experiments on a suite of multi-task benchmarks demonstrate both the effectiveness and efficiency of our method. Code is available at https://github.com/OptMN-Lab/SAMO.
Authors: Ashe Neth, Sawinder kaur, Mohammad Nur Hossain Khan, Subrata Biswas, Asif Salekin, Bashima Islam
Abstract: Existing pruning methods are typically applied during training or compile time and often rely on structured sparsity. While compatible with low-power microcontrollers (MCUs), structured pruning underutilizes the opportunity for fine-grained efficiency on devices without SIMD support or parallel compute. To address these limitations, we introduce UnIT (Unstructured Inference-Time pruning), a lightweight method that dynamically identifies and skips unnecessary multiply-accumulate (MAC) operations during inference, guided by input-specific activation patterns. Unlike structured pruning, UnIT embraces irregular sparsity and does not require retraining or hardware specialization. It transforms pruning decisions into lightweight comparisons, replacing multiplications with threshold checks and approximated divisions. UnIT further optimizes compute by reusing threshold computations across multiple connections and applying layer- and group-specific pruning sensitivity. We present three fast, hardware-friendly division approximations tailored to the capabilities of common embedded platforms. Demonstrated on the MSP430 microcontroller, UnIT achieves 11.02% to 82.03% MAC reduction, 27.30% to 84.19% faster inference, and 27.33% to 84.38% lower energy consumption compared to training-time pruned models, while maintaining accuracy with 0.48-7%. Under domain shift, UnIT matches or exceeds the accuracy of retrained models while requiring significantly fewer MACs. These results establish unstructured inference-time pruning as a viable and practical solution for efficient, retraining-free deployment of deep neural networks on MCUs.
Authors: Mohammad Fesanghary, Achintya Gopal
Abstract: In this study, we present a novel constraint-based algorithm for causal structure learning specifically designed for nonlinear autoregressive time series. Our algorithm significantly reduces computational complexity compared to existing methods, making it more efficient and scalable to larger problems. We rigorously evaluate its performance on synthetic datasets, demonstrating that our algorithm not only outperforms current techniques, but also excels in scenarios with limited data availability. These results highlight its potential for practical applications in fields requiring efficient and accurate causal inference from nonlinear time series data.
Authors: Anant Gupta, Rajarshi Bhowmik, Geoffrey Gunow
Abstract: Tracking the strategic focus of companies through topics in their earnings calls is a key task in financial analysis. However, as industries evolve, traditional topic modeling techniques struggle to dynamically capture emerging topics and their relationships. In this work, we propose an LLM-agent driven approach to discover and retrieve emerging topics from quarterly earnings calls. We propose an LLM-agent to extract topics from documents, structure them into a hierarchical ontology, and establish relationships between new and existing topics through a topic ontology. We demonstrate the use of extracted topics to infer company-level insights and emerging trends over time. We evaluate our approach by measuring ontology coherence, topic evolution accuracy, and its ability to surface emerging financial trends.
Authors: Jakub \v{C}ern\'y, Ji\v{r}\'i N\v{e}me\v{c}ek, Ivan Dovica, Jakub Mare\v{c}ek
Abstract: Explanations play a variety of roles in various recommender systems, from a legally mandated afterthought, through an integral element of user experience, to a key to persuasiveness. A natural and useful form of an explanation is the Counterfactual Explanation (CE). We present a method for generating highly plausible CEs in recommender systems and evaluate it both numerically and with a user study.
Authors: Sol Yarkoni, Roi Livni
Abstract: The recent advances in generative models such as diffusion models have raised several risks and concerns related to privacy, copyright infringements and data stewardship. To better understand and control the risks, various researchers have created techniques, experiments and attacks that reconstruct images, or part of images, from the training set. While these techniques already establish that data from the training set can be reconstructed, they often rely on high-resources, excess to the training set as well as well-engineered and designed prompts. In this work, we devise a new attack that requires low resources, assumes little to no access to the actual training set, and identifies, seemingly, benign prompts that lead to potentially-risky image reconstruction. This highlights the risk that images might even be reconstructed by an uninformed user and unintentionally. For example, we identified that, with regard to one existing model, the prompt ``blue Unisex T-Shirt'' can generate the face of a real-life human model. Our method builds on an intuition from previous works which leverages domain knowledge and identifies a fundamental vulnerability that stems from the use of scraped data from e-commerce platforms, where templated layouts and images are tied to pattern-like prompts.
Authors: Sukjun Hwang, Brandon Wang, Albert Gu
Abstract: Despite incredible progress in language models (LMs) in recent years, largely resulting from moving away from specialized models designed for specific tasks to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechanism which automatically learns content -- and context -- dependent segmentation strategies learned jointly with the rest of the model. Incorporating this into an explicit hierarchical network (H-Net) allows replacing the (implicitly hierarchical) tokenization-LM-detokenization pipeline with a single model learned fully end-to-end. When compute- and data- matched, an H-Net with one stage of hierarchy operating at the byte level outperforms a strong Transformer language model operating over BPE tokens. Iterating the hierarchy to multiple stages further increases its performance by modeling multiple levels of abstraction, demonstrating significantly better scaling with data and matching a token-based Transformer of twice its size. H-Nets pretrained on English show significantly increased character-level robustness, and qualitatively learn meaningful data-dependent chunking strategies without any heuristics or explicit supervision. Finally, the H-Net's improvement over tokenized pipelines is further increased in languages and modalities with weaker tokenization heuristics, such as Chinese and code, or DNA sequences (nearly 4x improvement in data efficiency over baselines), showing the potential of true end-to-end models that learn and scale better from unprocessed data.
Authors: Yuxin Bai, Cecelia Shuai, Ashwin De Silva, Siyu Yu, Pratik Chaudhari, Joshua T. Vogelstein
Abstract: In most real-world applications of artificial intelligence, the distributions of the data and the goals of the learners tend to change over time. The Probably Approximately Correct (PAC) learning framework, which underpins most machine learning algorithms, fails to account for dynamic data distributions and evolving objectives, often resulting in suboptimal performance. Prospective learning is a recently introduced mathematical framework that overcomes some of these limitations. We build on this framework to present preliminary results that improve the algorithm and numerical results, and extend prospective learning to sequential decision-making scenarios, specifically foraging. Code is available at: https://github.com/neurodata/prolearn2.
Authors: Qiyang Li, Zhiyuan Zhou, Sergey Levine
Abstract: We present Q-chunking, a simple yet effective recipe for improving reinforcement learning (RL) algorithms for long-horizon, sparse-reward tasks. Our recipe is designed for the offline-to-online RL setting, where the goal is to leverage an offline prior dataset to maximize the sample-efficiency of online learning. Effective exploration and sample-efficient learning remain central challenges in this setting, as it is not obvious how the offline data should be utilized to acquire a good exploratory policy. Our key insight is that action chunking, a technique popularized in imitation learning where sequences of future actions are predicted rather than a single action at each timestep, can be applied to temporal difference (TD)-based RL methods to mitigate the exploration challenge. Q-chunking adopts action chunking by directly running RL in a 'chunked' action space, enabling the agent to (1) leverage temporally consistent behaviors from offline data for more effective online exploration and (2) use unbiased $n$-step backups for more stable and efficient TD learning. Our experimental results demonstrate that Q-chunking exhibits strong offline performance and online sample efficiency, outperforming prior best offline-to-online methods on a range of long-horizon, sparse-reward manipulation tasks.
Authors: Perry Dong, Qiyang Li, Dorsa Sadigh, Chelsea Finn
Abstract: We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propagation from actions to policy parameters when optimizing against some value function. Our key insight is that we can address stable value maximization by avoiding direct optimization over value with the expressive policy and instead construct an on-the-fly RL policy to maximize Q-value. We propose Expressive Policy Optimization (EXPO), a sample-efficient online RL algorithm that utilizes an on-the-fly policy to maximize value with two parameterized policies -- a larger expressive base policy trained with a stable imitation learning objective and a light-weight Gaussian edit policy that edits the actions sampled from the base policy toward a higher value distribution. The on-the-fly policy optimizes the actions from the base policy with the learned edit policy and chooses the value maximizing action from the base and edited actions for both sampling and temporal-difference (TD) backup. Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods both in the setting of fine-tuning a pretrained policy given offline data and in leveraging offline data to train online.
Authors: Ziyue Li, Yang Li, Tianyi Zhou
Abstract: Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample. This compositional space greatly expands the scope of existing works on looped/recurrent pretrained modules, layer pruning, or early-exit networks. We develop a Monte Carlo Tree Search (MCTS) protocol to explore and identify the optimal CoLa for each sample from math and commonsense reasoning benchmarks. Compared to a static model of a fixed depth, CoLa allows shortcut paths (fast thinking), recurrence of the same layer(s) (slow thinking), and combining both, offering more flexible, dynamic architectures for different inputs. We conduct an extensive analysis of the MCTS-optimized CoLa, which leads to two key findings: (1) For >75% of samples with correct predictions by the original LLM, we can find shorter CoLa, suggesting a large space for improving inference efficiency; (2) For >60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions, suggesting a large space of performance enhancement. Our results highlight the shortcomings of using a fixed architecture of pre-trained LLMs for inference on different samples and pave the way to unlock the generalization power of test-time depth adaptation.
Authors: Zhiwei Hu, V\'ictor Guti\'errez-Basulto, Zhiliang Xiang, Ru Li, Jeff Z. Pan
Abstract: Multimodal Entity Linking (MEL) aims to link ambiguous mentions within multimodal contexts to associated entities in a multimodal knowledge base. Existing approaches to MEL introduce multimodal interaction and fusion mechanisms to bridge the modality gap and enable multi-grained semantic matching. However, they do not address two important problems: (i) mention ambiguity, i.e., the lack of semantic content caused by the brevity and omission of key information in the mention's textual context; (ii) dynamic selection of modal content, i.e., to dynamically distinguish the importance of different parts of modal information. To mitigate these issues, we propose a Multi-level Mixture of Experts (MMoE) model for MEL. MMoE has four components: (i) the description-aware mention enhancement module leverages large language models to identify the WikiData descriptions that best match a mention, considering the mention's textual context; (ii) the multimodal feature extraction module adopts multimodal feature encoders to obtain textual and visual embeddings for both mentions and entities; (iii)-(iv) the intra-level mixture of experts and inter-level mixture of experts modules apply a switch mixture of experts mechanism to dynamically and adaptively select features from relevant regions of information. Extensive experiments demonstrate the outstanding performance of MMoE compared to the state-of-the-art. MMoE's code is available at: https://github.com/zhiweihu1103/MEL-MMoE.
Authors: Erez Weintraub, Ron Banner, Ariel Orda
Abstract: State-of-the-art language and vision models are routinely trained across thousands of GPUs, often spanning multiple data-centers, yet today's distributed frameworks still assume reliable connections (e.g., InfiniBand or RoCE). The resulting acknowledgment traffic and retransmissions inflate tail latencies and limit scalability. Leveraging unreliable connections will reduce latency but may sacrifice model accuracy and convergence once packets are dropped. A principled, end-to-end solution that preserves accuracy and convergence guarantees under genuine packet loss has previously been missing. We address this critical gap by introducing a novel distributed training framework capable of operating over unreliable connections, offering unbiased gradient aggregation and bounded parameter drift without modifying model code or optimizers. The key insight is a two-stage defense against missing messages: (i) Unbiased gradient aggregation: each worker reconstructs a consistent gradient estimate from whatever packets arrive, guaranteeing expectation-level correctness; and (ii) Bounded-drift parameter broadcasts: we prove the inter-worker model discrepancy remains O(1) even after arbitrarily many iterations, preventing the unbounded divergence typical of asynchronous setups. Analytical bounds are matched by experiments on the LLAMA2 7B model with 64 GPUs: tolerating 10% random packet loss yields at most 0.8% perplexity change. This work bridges the gap between communication-efficient datacenter protocols and the accuracy and generalization guarantees demanded by modern large-model training, enabling robust, high-throughput learning on commodity or wide-area networks.
Authors: Zelin Zhu, Kai Yang, Rui Zhang
Abstract: Wireless localization and sensing technologies are essential in modern wireless networks, supporting applications in smart cities, the Internet of Things (IoT), and autonomous systems. High-performance localization and sensing systems are critical for both network efficiency and emerging intelligent applications. Integrating channel state information (CSI) with deep learning has recently emerged as a promising solution. Recent works have leveraged the spatial diversity of multiple input multiple output (MIMO) systems and the frequency granularity of orthogonal frequency division multiplexing (OFDM) waveforms to improve spatial resolution. Nevertheless, the joint modeling of localization and sensing under the high-dimensional CSI characteristics of MIMO-OFDM systems remains insufficiently investigated. This work aims to jointly model and optimize localization and sensing tasks to harness their potential synergy. We first formulate localization and sensing as a mixed-integer bilevel deep learning problem and then propose a novel stochastic proximal gradient-based mixed-integer bilevel optimization (SPG-MIBO) algorithm. SPG-MIBO is well-suited for high-dimensional and large-scale datasets, leveraging mini-batch training at each step for computational and memory efficiency. The algorithm is also supported by theoretical convergence guarantees. Extensive experiments on multiple datasets validate its effectiveness and highlight the performance gains from joint localization and sensing optimization.
Authors: Zihan Zhang, Leon Wong, Blesson Varghese
Abstract: A Federated Learning (FL) system collaboratively trains neural networks across devices and a server but is limited by significant on-device computation costs. Split Federated Learning (SFL) systems mitigate this by offloading a block of layers of the network from the device to a server. However, in doing so, it introduces large communication overheads due to frequent exchanges of intermediate activations and gradients between devices and the server and reduces model accuracy for non-IID data. We propose Ampere, a novel collaborative training system that simultaneously minimizes on-device computation and device-server communication while improving model accuracy. Unlike SFL, which uses a global loss by iterative end-to-end training, Ampere develops unidirectional inter-block training to sequentially train the device and server block with a local loss, eliminating the transfer of gradients. A lightweight auxiliary network generation method decouples training between the device and server, reducing frequent intermediate exchanges to a single transfer, which significantly reduces the communication overhead. Ampere mitigates the impact of data heterogeneity by consolidating activations generated by the trained device block to train the server block, in contrast to SFL, which trains on device-specific, non-IID activations. Extensive experiments on multiple CNNs and transformers show that, compared to state-of-the-art SFL baseline systems, Ampere (i) improves model accuracy by up to 13.26% while reducing training time by up to 94.6%, (ii) reduces device-server communication overhead by up to 99.1% and on-device computation by up to 93.13%, and (iii) reduces standard deviation of accuracy by 53.39% for various non-IID degrees highlighting superior performance when faced with heterogeneous data.
Authors: Mathieu Tuli, Kaveh Kamali, David B. Lindell
Abstract: We introduce the task of generative panoramic image stitching, which aims to synthesize seamless panoramas that are faithful to the content of multiple reference images containing parallax effects and strong variations in lighting, camera capture settings, or style. In this challenging setting, traditional image stitching pipelines fail, producing outputs with ghosting and other artifacts. While recent generative models are capable of outpainting content consistent with multiple reference images, they fail when tasked with synthesizing large, coherent regions of a panorama. To address these limitations, we propose a method that fine-tunes a diffusion-based inpainting model to preserve a scene's content and layout based on multiple reference images. Once fine-tuned, the model outpaints a full panorama from a single reference image, producing a seamless and visually coherent result that faithfully integrates content from all reference images. Our approach significantly outperforms baselines for this task in terms of image quality and the consistency of image structure and scene layout when evaluated on captured datasets.
Authors: Mridula Vijendran, Shuang Chen, Jingjing Deng, Hubert P. H. Shum
Abstract: The pervasive issue of bias in AI presents a significant challenge to painting classification, and is getting more serious as these systems become increasingly integrated into tasks like art curation and restoration. Biases, often arising from imbalanced datasets where certain artistic styles dominate, compromise the fairness and accuracy of model predictions, i.e., classifiers are less accurate on rarely seen paintings. While prior research has made strides in improving classification performance, it has largely overlooked the critical need to address these underlying biases, that is, when dealing with out-of-distribution (OOD) data. Our insight highlights the necessity of a more robust approach to bias mitigation in AI models for art classification on biased training data. We propose a novel OOD-informed model bias adaptive sampling method called BOOST (Bias-Oriented OOD Sampling and Tuning). It addresses these challenges by dynamically adjusting temperature scaling and sampling probabilities, thereby promoting a more equitable representation of all classes. We evaluate our proposed approach to the KaoKore and PACS datasets, focusing on the model's ability to reduce class-wise bias. We further propose a new metric, Same-Dataset OOD Detection Score (SODC), designed to assess class-wise separation and per-class bias reduction. Our method demonstrates the ability to balance high performance with fairness, making it a robust solution for unbiasing AI models in the art domain.
Authors: Renyang Liu, Guanlin Li, Tianwei Zhang, See-Kiong Ng
Abstract: Recent advances in image generation models (IGMs), particularly diffusion-based architectures such as Stable Diffusion (SD), have markedly enhanced the quality and diversity of AI-generated visual content. However, their generative capability has also raised significant ethical, legal, and societal concerns, including the potential to produce harmful, misleading, or copyright-infringing content. To mitigate these concerns, machine unlearning (MU) emerges as a promising solution by selectively removing undesirable concepts from pretrained models. Nevertheless, the robustness and effectiveness of existing unlearning techniques remain largely unexplored, particularly in the presence of multi-modal adversarial inputs. To bridge this gap, we propose Recall, a novel adversarial framework explicitly designed to compromise the robustness of unlearned IGMs. Unlike existing approaches that predominantly rely on adversarial text prompts, Recall exploits the intrinsic multi-modal conditioning capabilities of diffusion models by efficiently optimizing adversarial image prompts with guidance from a single semantically relevant reference image. Extensive experiments across ten state-of-the-art unlearning methods and diverse tasks show that Recall consistently outperforms existing baselines in terms of adversarial effectiveness, computational efficiency, and semantic fidelity with the original textual prompt. These findings reveal critical vulnerabilities in current unlearning mechanisms and underscore the need for more robust solutions to ensure the safety and reliability of generative models. Code and data are publicly available at \textcolor{blue}{https://github.com/ryliu68/RECALL}.
Authors: Getamesay Haile Dagnaw, Yanming Zhu, Muhammad Hassan Maqsood, Wencheng Yang, Xingshuai Dong, Xuefei Yin, Alan Wee-Chung Liew
Abstract: Explainable artificial intelligence (XAI) has become increasingly important in biomedical image analysis to promote transparency, trust, and clinical adoption of DL models. While several surveys have reviewed XAI techniques, they often lack a modality-aware perspective, overlook recent advances in multimodal and vision-language paradigms, and provide limited practical guidance. This survey addresses this gap through a comprehensive and structured synthesis of XAI methods tailored to biomedical image analysis.We systematically categorize XAI methods, analyzing their underlying principles, strengths, and limitations within biomedical contexts. A modality-centered taxonomy is proposed to align XAI methods with specific imaging types, highlighting the distinct interpretability challenges across modalities. We further examine the emerging role of multimodal learning and vision-language models in explainable biomedical AI, a topic largely underexplored in previous work. Our contributions also include a summary of widely used evaluation metrics and open-source frameworks, along with a critical discussion of persistent challenges and future directions. This survey offers a timely and in-depth foundation for advancing interpretable DL in biomedical image analysis.
Authors: Renyuan Liu (Richard), Yuyang Leng (Richard), Kaiyan Liu (Richard), Shaohan Hu (Richard), Chun-Fu (Richard), Chen, Peijun Zhao, Heechul Yun, Shuochao Yao
Abstract: Recent advancements in on-device training for deep neural networks have underscored the critical need for efficient activation compression to overcome the memory constraints of mobile and edge devices. As activations dominate memory usage during training and are essential for gradient computation, compressing them without compromising accuracy remains a key research challenge. While existing methods for dynamic activation quantization promise theoretical memory savings, their practical deployment is impeded by system-level challenges such as computational overhead and memory fragmentation. To address these challenges, we introduce DAF, a Dynamic Activation Framework that enables scalable and efficient on-device training through system-level optimizations. DAF achieves both memory- and time-efficient dynamic quantization training by addressing key system bottlenecks. It develops hybrid reduction operations tailored to the memory hierarchies of mobile and edge SoCs, leverages collaborative CPU-GPU bit-packing for efficient dynamic quantization, and implements an importance-aware paging memory management scheme to reduce fragmentation and support dynamic memory adjustments. These optimizations collectively enable DAF to achieve substantial memory savings and speedup without compromising model training accuracy. Evaluations on various deep learning models across embedded and mobile platforms demonstrate up to a $22.9\times$ reduction in memory usage and a $3.2\times$ speedup, making DAF a scalable and practical solution for resource-constrained environments.
Authors: Jean-Baptiste Fermanian (IMAG, IROKO), Mohamed Hebiri (LAMA), Joseph Salmon (IMAG, IROKO)
Abstract: Conformal prediction methods are statistical tools designed to quantify uncertainty and generate predictive sets with guaranteed coverage probabilities. This work introduces an innovative refinement to these methods for classification tasks, specifically tailored for scenarios where multiple observations (multi-inputs) of a single instance are available at prediction time. Our approach is particularly motivated by applications in citizen science, where multiple images of the same plant or animal are captured by individuals. Our method integrates the information from each observation into conformal prediction, enabling a reduction in the size of the predicted label set while preserving the required class-conditional coverage guarantee. The approach is based on the aggregation of conformal p-values computed from each observation of a multi-input. By exploiting the exact distribution of these p-values, we propose a general aggregation framework using an abstract scoring function, encompassing many classical statistical tools. Knowledge of this distribution also enables refined versions of standard strategies, such as majority voting. We evaluate our method on simulated and real data, with a particular focus on Pl@ntNet, a prominent citizen science platform that facilitates the collection and identification of plant species through user-submitted images.
Authors: Nicole Abreu, Parker B. Edwards, Francis Motta
Abstract: Supervised machine learning pipelines trained on features derived from persistent homology have been experimentally observed to ignore much of the information contained in a persistence diagram. Computing persistence diagrams is often the most computationally demanding step in such a pipeline, however. To explore this, we introduce several methods to generate topological feature vectors from unreduced boundary matrices. We compared the performance of pipelines trained on vectorizations of unreduced PDs to vectorizations of fully-reduced PDs across several data and task types. Our results indicate that models trained on PDs built from unreduced diagrams can perform on par and even outperform those trained on fully-reduced diagrams on some tasks. This observation suggests that machine learning pipelines which incorporate topology-based features may benefit in terms of computational cost and performance by utilizing information contained in unreduced boundary matrices.
Authors: Arshak Rezvani, Ali Akbari, Kosar Sanjar Arani, Maryam Mirian, Emad Arasteh, Martin J. McKeown
Abstract: Decoding visual experience from brain signals offers exciting possibilities for neuroscience and interpretable AI. While EEG is accessible and temporally precise, its limitations in spatial detail hinder image reconstruction. Our model bypasses direct EEG-to-image generation by aligning EEG signals with multilevel semantic captions -- ranging from object-level to abstract themes -- generated by a large language model. A transformer-based EEG encoder maps brain activity to these captions through contrastive learning. During inference, caption embeddings retrieved via projection heads condition a pretrained latent diffusion model for image generation. This text-mediated framework yields state-of-the-art visual decoding on the EEGCVPR dataset, with interpretable alignment to known neurocognitive pathways. Dominant EEG-caption associations reflected the importance of different semantic levels extracted from perceived images. Saliency maps and t-SNE projections reveal semantic topography across the scalp. Our model demonstrates how structured semantic mediation enables cognitively aligned visual decoding from EEG.
Authors: Nishan Ranabhat, Behnam Javanparast, David Goerz, Estelle Inack
Abstract: Portfolio optimization is a routine asset management operation conducted in financial institutions around the world. However, under real-world constraints such as turnover limits and transaction costs, its formulation becomes a mixed-integer nonlinear program that current mixed-integer optimizers often struggle to solve. We propose mapping this problem onto a classical Ising-like Hamiltonian and solving it with Variational Neural Annealing (VNA), via its classical formulation implemented using autoregressive neural networks. We demonstrate that VNA can identify near-optimal solutions for portfolios comprising more than 2,000 assets and yields performance comparable to that of state-of-the-art optimizers, such as Mosek, while exhibiting faster convergence on hard instances. Finally, we present a dynamical finite-size scaling analysis applied to the S&P 500, Russell 1000, and Russell 3000 indices, revealing universal behavior and polynomial annealing time scaling of the VNA algorithm on portfolio optimization problems.
Authors: Itay Itzhak, Yonatan Belinkov, Gabriel Stanovsky
Abstract: Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over $30$ cognitive biases. Second, we introduce \emph{cross-tuning} -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.
Authors: Dong Xu, Zhangfan Yang, Sisi Yuan, Jenna Xinyi Yao, Jiangqiang Li, Junkai Ji
Abstract: Three-dimensional molecular generators based on diffusion models can now reach near-crystallographic accuracy, yet they remain fragmented across tasks. SMILES-only inputs, two-stage pretrain-finetune pipelines, and one-task-one-model practices hinder stereochemical fidelity, task alignment, and zero-shot transfer. We introduce MODA, a diffusion framework that unifies fragment growing, linker design, scaffold hopping, and side-chain decoration with a Bayesian mask scheduler. During training, a contiguous spatial fragment is masked and then denoised in one pass, enabling the model to learn shared geometric and chemical priors across tasks. Multi-task training yields a universal backbone that surpasses six diffusion baselines and three training paradigms on substructure, chemical property, interaction, and geometry. Model-C reduces ligand-protein clashes and substructure divergences while maintaining Lipinski compliance, whereas Model-B preserves similarity but trails in novelty and binding affinity. Zero-shot de novo design and lead-optimisation tests confirm stable negative Vina scores and high improvement rates without force-field refinement. These results demonstrate that a single-stage multi-task diffusion routine can replace two-stage workflows for structure-based molecular design.
Authors: Zili Wang, Frank Montabon, Kristin Yvonne Rozier
Abstract: Supply chain networks are complex systems that are challenging to analyze; this problem is exacerbated when there are illicit activities involved in the supply chain, such as counterfeit parts, forced labor, or human trafficking. While machine learning (ML) can find patterns in complex systems like supply chains, traditional ML techniques require large training data sets. However, illicit supply chains are characterized by very sparse data, and the data that is available is often (purposely) corrupted or unreliable in order to hide the nature of the activities. We need to be able to automatically detect new patterns that correlate with such illegal activity over complex, even temporal data, without requiring large training data sets. We explore neurosymbolic methods for identifying instances of illicit activity in supply chains and compare the effectiveness of manual and automated feature extraction from news articles accurately describing illicit activities uncovered by authorities. We propose a question tree approach for querying a large language model (LLM) to identify and quantify the relevance of articles. This enables a systematic evaluation of the differences between human and machine classification of news articles related to forced labor in supply chains.
Authors: Aaron Goldstein, Ayan Dutta
Abstract: Traditional recommendation algorithms are not designed to provide personalized recommendations based on user preferences provided through text, e.g., "I enjoy light-hearted comedies with a lot of humor". Large Language Models (LLMs) have emerged as one of the most promising tools for natural language processing in recent years. This research proposes a novel framework that mimics how a close friend would recommend items based on their knowledge of an individual's tastes. We leverage LLMs to enhance movie recommendation systems by refining traditional algorithm outputs and integrating them with language-based user preference inputs. We employ Singular Value Decomposition (SVD) or SVD++ algorithms to generate initial movie recommendations, implemented using the Surprise Python library and trained on the MovieLens-Latest-Small dataset. We compare the performance of the base algorithms with our LLM-enhanced versions using leave-one-out validation hit rates and cumulative hit rates. Additionally, to compare the performance of our framework against the current state-of-the-art recommendation systems, we use rating and ranking metrics with an item-based stratified 0.75 train, 0.25 test split. Our framework can generate preference profiles automatically based on users' favorite movies or allow manual preference specification for more personalized results. Using an automated approach, our framework overwhelmingly surpassed SVD and SVD++ on every evaluation metric used (e.g., improvements of up to ~6x in cumulative hit rate, ~3.7x in NDCG, etc.), albeit at the cost of a slight increase in computational overhead.
Authors: Marcel Hudiani
Abstract: We study the almost sure convergence rate for the last iterate of stochastic gradient descent (SGD) and stochastic heavy ball (SHB) in the parametric setting when the objective function $F$ is globally convex or non-convex whose gradient is $\gamma$-H\"{o}lder. Using only discrete Gronwall's inequality without Robbins-Siegmund theorem nor martingale convergence theory, we recover results for both SGD and SHB: $\min_{s\leq t} \|\nabla F(w_s)\|^2 = o(t^{p-1})$ for non-convex objectives and $F(w_t) - F_* = o(t^{2\gamma/(1+\gamma) \cdot \max(p-1,-2p+1)-\epsilon})$ for $\beta \in (0, 1)$ and $\min_{s \leq t} F(w_s) - F_* = o(t^{p-1})$ almost surely for convex objectives. In addition, we proved that SHB with constant momentum parameter $\beta \in (0, 1)$ attains a convergence rate of $F(w_t) - F_* = O(t^{\max(p-1,-2p+1)} \log^2 \frac{t}{\delta})$ with probability at least $1-\delta$ when $F$ is convex and $\gamma = 1$ and step size $\alpha_t = \Theta(t^{-p})$ with $p \in (\frac{1}{2}, 1)$.
Authors: Juejing Liu, Haydn Anderson, Noah I. Waxman, Vsevolod Kovalev, Byron Fisher, Elizabeth Li, Xiaofeng Guo
Abstract: New discoveries in chemistry and materials science, with increasingly expanding volume of requisite knowledge and experimental workload, provide unique opportunities for machine learning (ML) to take critical roles in accelerating research efficiency. Here, we demonstrate (1) the use of large language models (LLMs) for automated literature reviews, and (2) the training of an ML model to predict chemical knowledge (thermodynamic parameters). Our LLM-based literature review tool (LMExt) successfully extracted chemical information and beyond into a machine-readable structure, including stability constants for metal cation-ligand interactions, thermodynamic properties, and other broader data types (medical research papers, and financial reports), effectively overcoming the challenges inherent in each domain. Using the autonomous acquisition of thermodynamic data, an ML model was trained using the CatBoost algorithm for accurately predicting thermodynamic parameters (e.g., enthalpy of formation) of minerals. This work highlights the transformative potential of integrated ML approaches to reshape chemistry and materials science research.
Authors: Ben A. Marconi
Abstract: Financial time series forecasting presents significant challenges due to complex nonlinear relationships, temporal dependencies, variable interdependencies and limited data availability, particularly for tasks involving low-frequency data, newly listed instruments, or emerging market assets. Time Series Foundation Models (TSFMs) offer a promising solution through pretraining on diverse time series corpora followed by task-specific adaptation. This study evaluates two TSFMs (Tiny Time Mixers (TTM) and Chronos) across three financial forecasting tasks: US 10-year Treasury yield changes, EUR/USD volatility, and equity spread prediction. Results demonstrate that TTM exhibits strong transferability. When fine-tuning both the pretrained version of TTM and an untrained model with the same architecture, the pretrained version achieved 25-50% better performance when fine-tuned on limited data and 15-30% improvements even when fine-tuned on lengthier datasets. Notably, TTM's zero-shot performance outperformed naive benchmarks in volatility forecasting and equity spread prediction, with the latter demonstrating that TSFMs can surpass traditional benchmark models without fine-tuning. The pretrained model consistently required 3-10 fewer years of data to achieve comparable performance levels compared to the untrained model, demonstrating significant sample-efficiency gains. However, while TTM outperformed naive baselines, traditional specialised models matched or exceeded its performance in two of three tasks, suggesting TSFMs prioritise breadth over task-specific optimisation. These findings indicate that TSFMs, though still nascent, offer substantial promise for financial forecasting-particularly in noisy, data-constrained tasks-but achieving competitive performance likely requires domain-specific pretraining and architectural refinements tailored to financial time series characteristics.
Authors: Muhammad Kazim, Harun Pirim, Chau Le, Trung Le, Om Prakash Yadav
Abstract: Unplanned power outages cost the US economy over $150 billion annually, partly due to predictive maintenance (PdM) models that overlook spatial, temporal, and causal dependencies in grid failures. This study introduces a multilayer Graph Neural Network (GNN) framework to enhance PdM and enable resilience-based substation clustering. Using seven years of incident data from Oklahoma Gas & Electric (292,830 records across 347 substations), the framework integrates Graph Attention Networks (spatial), Graph Convolutional Networks (temporal), and Graph Isomorphism Networks (causal), fused through attention-weighted embeddings. Our model achieves a 30-day F1-score of 0.8935 +/- 0.0258, outperforming XGBoost and Random Forest by 3.2% and 2.7%, and single-layer GNNs by 10 to 15 percent. Removing the causal layer drops performance to 0.7354 +/- 0.0418. For resilience analysis, HDBSCAN clustering on HierarchicalRiskGNN embeddings identifies eight operational risk groups. The highest-risk cluster (Cluster 5, 44 substations) shows 388.4 incidents/year and 602.6-minute recovery time, while low-risk groups report fewer than 62 incidents/year. ANOVA (p < 0.0001) confirms significant inter-cluster separation. Our clustering outperforms K-Means and Spectral Clustering with a Silhouette Score of 0.626 and Davies-Bouldin index of 0.527. This work supports proactive grid management through improved failure prediction and risk-aware substation clustering.
Authors: Nick Polson, Vadim Sokolov
Abstract: Double descent is a phenomenon of over-parameterized statistical models. Our goal is to view double descent from a Bayesian perspective. Over-parameterized models such as deep neural networks have an interesting re-descending property in their risk characteristics. This is a recent phenomenon in machine learning and has been the subject of many studies. As the complexity of the model increases, there is a U-shaped region corresponding to the traditional bias-variance trade-off, but then as the number of parameters equals the number of observations and the model becomes one of interpolation, the risk can become infinite and then, in the over-parameterized region, it re-descends -- the double descent effect. We show that this has a natural Bayesian interpretation. Moreover, we show that it is not in conflict with the traditional Occam's razor that Bayesian models possess, in that they tend to prefer simpler models when possible. We illustrate the approach with an example of Bayesian model selection in neural networks. Finally, we conclude with directions for future research.
Authors: Yingtao Luo, Reza Skandari, Carlos Martinez, Arman Kilic, Rema Padman
Abstract: Decisions about managing patients on the heart transplant waitlist are currently made by committees of doctors who consider multiple factors, but the process remains largely ad-hoc. With the growing volume of longitudinal patient, donor, and organ data collected by the United Network for Organ Sharing (UNOS) since 2018, there is increasing interest in analytical approaches to support clinical decision-making at the time of organ availability. In this study, we benchmark machine learning models that leverage longitudinal waitlist history data for time-dependent, time-to-event modeling of waitlist mortality. We train on 23,807 patient records with 77 variables and evaluate both survival prediction and discrimination at a 1-year horizon. Our best model achieves a C-Index of 0.94 and AUROC of 0.89, significantly outperforming previous models. Key predictors align with known risk factors while also revealing novel associations. Our findings can support urgency assessment and policy refinement in heart transplant decision making.
Authors: James P. Crutchfield
Abstract: We show that mixtures comprised of multicomponent systems typically are much more structurally complex than the sum of their parts; sometimes, infinitely more complex. We contrast this with the more familiar notion of statistical mixtures, demonstrating how statistical mixtures miss key aspects of emergent hierarchical organization. This leads us to identify a new kind of structural complexity inherent in multicomponent systems and to draw out broad consequences for system ergodicity.
Authors: Lo\"ic Pottier, Konstantia Georgouli, Timothy S. Carpenter, Fikret Aydin, Jeremy O. B. Tempkin, Dwight V. Nissley, Frederick H. Streitz, Thomas R. W. Scogland, Peer-Timo Bremer, Felice C. Lightstone, Helgi I. Ing\'olfsson
Abstract: Computational models have become one of the prevalent methods to model complex phenomena. To accurately model complex interactions, such as detailed biomolecular interactions, scientists often rely on multiscale models comprised of several internal models operating at difference scales, ranging from microscopic to macroscopic length and time scales. Bridging the gap between different time and length scales has historically been challenging but the advent of newer machine learning (ML) approaches has shown promise for tackling that task. Multiscale models require massive amounts of computational power and a powerful workflow management system. Orchestrating ML-driven multiscale studies on parallel systems with thousands of nodes is challenging, the workflow must schedule, allocate and control thousands of simulations operating at different scales. Here, we discuss the massively parallel Multiscale Machine-Learned Modeling Infrastructure (MuMMI), a multiscale workflow management infrastructure, that can orchestrate thousands of molecular dynamics (MD) simulations operating at different timescales, spanning from millisecond to nanosecond. More specifically, we introduce a novel version of MuMMI called "mini-MuMMI". Mini-MuMMI is a curated version of MuMMI designed to run on modest HPC systems or even laptops whereas MuMMI requires larger HPC systems. We demonstrate mini-MuMMI utility by exploring RAS-RAF membrane interactions and discuss the different challenges behind the generalization of multiscale workflows and how mini-MuMMI can be leveraged to target a broader range of applications outside of MD and RAS-RAF interactions.
Authors: Erika Yilin Zheng, Yu Yan, Baradwaj Simha Sankar, Ethan Ji, Steven Swee, Irsyad Adam, Ding Wang, Alexander Russell Pelletier, Alex Bui, Wei Wang, Peipei Ping
Abstract: Existing machine learning methods for molecular (e.g., gene) embeddings are restricted to specific tasks or data modalities, limiting their effectiveness within narrow domains. As a result, they fail to capture the full breadth of gene functions and interactions across diverse biological contexts. In this study, we have systematically evaluated knowledge representations of biomolecules across multiple dimensions representing a task-agnostic manner spanning three major data sources, including omics experimental data, literature-derived text data, and knowledge graph-based representations. To distinguish between meaningful biological signals from chance correlations, we devised an adjusted variant of Singular Vector Canonical Correlation Analysis (SVCCA) that quantifies signal redundancy and complementarity across different data modalities and sources. These analyses reveal that existing embeddings capture largely non-overlapping molecular signals, highlighting the value of embedding integration. Building on this insight, we propose Platform for Representation and Integration of multimodal Molecular Embeddings (PRISME), a machine learning based workflow using an autoencoder to integrate these heterogeneous embeddings into a unified multimodal representation. We validated this approach across various benchmark tasks, where PRISME demonstrated consistent performance, and outperformed individual embedding methods in missing value imputations. This new framework supports comprehensive modeling of biomolecules, advancing the development of robust, broadly applicable multimodal embeddings optimized for downstream biomedical machine learning applications.
Authors: Zhanhong Jiang, Dylan Shah, Hsin-Jung Yang, Soumik Sarkar
Abstract: Precise kinematic modeling is critical in calibration and controller design for soft robots, yet remains a challenging issue due to their highly nonlinear and complex behaviors. To tackle the issue, numerous data-driven machine learning approaches have been proposed for modeling nonlinear dynamics. However, these models suffer from prediction uncertainty that can negatively affect modeling accuracy, and uncertainty quantification for kinematic modeling in soft robots is underexplored. In this work, using limited simulation and real-world data, we first investigate multiple linear and nonlinear machine learning models commonly used for kinematic modeling of soft robots. The results reveal that nonlinear ensemble methods exhibit the most robust generalization performance. We then develop a conformal kinematic modeling framework for soft robots by utilizing split conformal prediction to quantify predictive position uncertainty, ensuring distribution-free prediction intervals with a theoretical guarantee.
Authors: Zeyang Song, Shimin Zhang, Yuhong Chou, Jibin Wu, Haizhou Li
Abstract: Spiking Neural Networks (SNNs), inspired by biological neural mechanisms, represent a promising neuromorphic computing paradigm that offers energy-efficient alternatives to traditional Artificial Neural Networks (ANNs). Despite proven effectiveness, SNN architectures have struggled to achieve competitive performance on large-scale speech processing task. Two key challenges hinder progress: (1) the high computational overhead during training caused by multi-timestep spike firing, and (2) the absence of large-scale SNN architectures tailored to speech processing tasks. To overcome the issues, we introduce Input-aware Multi-Level Spikeformer, i.e. IML-Spikeformer, a spiking Transformer architecture specifically designed for large-scale speech processing. Central to our design is the Input-aware Multi-Level Spike (IMLS) mechanism, which simulate multi-timestep spike firing within a single timestep using an adaptive, input-aware thresholding scheme. IML-Spikeformer further integrates a Reparameterized Spiking Self-Attention (RepSSA) module with a Hierarchical Decay Mask (HDM), forming the HD-RepSSA module. This module enhances the precision of attention maps and enables modeling of multi-scale temporal dependencies in speech signals. Experiments demonstrate that IML-Spikeformer achieves word error rates of 6.0\% on AiShell-1 and 3.4\% on Librispeech-960, comparable to conventional ANN transformers while reducing theoretical inference energy consumption by 4.64$\times$ and 4.32$\times$ respectively. IML-Spikeformer marks an advance of scalable SNN architectures for large-scale speech processing in both task performance and energy efficiency.
Authors: Jikesh Thapa, Gurrehmat Chahal, Serban Voinea Gabreanu, Yazan Otoum
Abstract: Phishing attacks are becoming increasingly sophisticated, underscoring the need for detection systems that strike a balance between high accuracy and computational efficiency. This paper presents a comparative evaluation of traditional Machine Learning (ML), Deep Learning (DL), and quantized small-parameter Large Language Models (LLMs) for phishing detection. Through experiments on a curated dataset, we show that while LLMs currently underperform compared to ML and DL methods in terms of raw accuracy, they exhibit strong potential for identifying subtle, context-based phishing cues. We also investigate the impact of zero-shot and few-shot prompting strategies, revealing that LLM-rephrased emails can significantly degrade the performance of both ML and LLM-based detectors. Our benchmarking highlights that models like DeepSeek R1 Distill Qwen 14B (Q8_0) achieve competitive accuracy, above 80%, using only 17GB of VRAM, supporting their viability for cost-efficient deployment. We further assess the models' adversarial robustness and cost-performance tradeoffs, and demonstrate how lightweight LLMs can provide concise, interpretable explanations to support real-time decision-making. These findings position optimized LLMs as promising components in phishing defence systems and offer a path forward for integrating explainable, efficient AI into modern cybersecurity frameworks.
Authors: Mohammad F. Al-Hammouri, Yazan Otoum, Rasha Atwa, Amiya Nayak
Abstract: This paper presents a novel approach to intrusion detection by integrating traditional signature-based methods with the contextual understanding capabilities of the GPT-2 Large Language Model (LLM). As cyber threats become increasingly sophisticated, particularly in distributed, heterogeneous, and resource-constrained environments such as those enabled by the Internet of Things (IoT), the need for dynamic and adaptive Intrusion Detection Systems (IDSs) becomes increasingly urgent. While traditional methods remain effective for detecting known threats, they often fail to recognize new and evolving attack patterns. In contrast, GPT-2 excels at processing unstructured data and identifying complex semantic relationships, making it well-suited to uncovering subtle, zero-day attack vectors. We propose a hybrid IDS framework that merges the robustness of signature-based techniques with the adaptability of GPT-2-driven semantic analysis. Experimental evaluations on a representative intrusion dataset demonstrate that our model enhances detection accuracy by 6.3%, reduces false positives by 9.0%, and maintains near real-time responsiveness. These results affirm the potential of language model integration to build intelligent, scalable, and resilient cybersecurity defences suited for modern connected environments.
Authors: Jenifer Paulraj, Brindha Raghuraman, Nagarani Gopalakrishnan, Yazan Otoum
Abstract: Critical infrastructure systems, including energy grids, healthcare facilities, transportation networks, and water distribution systems, are pivotal to societal stability and economic resilience. However, the increasing interconnectivity of these systems exposes them to various cyber threats, including ransomware, Denial-of-Service (DoS) attacks, and Advanced Persistent Threats (APTs). This paper examines cybersecurity vulnerabilities in critical infrastructure, highlighting the threat landscape, attack vectors, and the role of Artificial Intelligence (AI) in mitigating these risks. We propose a hybrid AI-driven cybersecurity framework to enhance real-time vulnerability detection, threat modelling, and automated remediation. This study also addresses the complexities of adversarial AI, regulatory compliance, and integration. Our findings provide actionable insights to strengthen the security and resilience of critical infrastructure systems against emerging cyber threats.
Authors: Abdelrahman S. Abdelrahman, Shuvro Chowdhury, Flaviano Morone, Kerem Y. Camsari
Abstract: We introduce a generalized \textit{Probabilistic Approximate Optimization Algorithm (PAOA)}, a classical variational Monte Carlo framework that extends and formalizes prior work by Weitz \textit{et al.}~\cite{Combes_2023}, enabling parameterized and fast sampling on present-day Ising machines and probabilistic computers. PAOA operates by iteratively modifying the couplings of a network of binary stochastic units, guided by cost evaluations from independent samples. We establish a direct correspondence between derivative-free updates and the gradient of the full $2^N \times 2^N$ Markov flow, showing that PAOA admits a principled variational formulation. Simulated annealing emerges as a limiting case under constrained parameterizations, and we implement this regime on an FPGA-based probabilistic computer with on-chip annealing to solve large 3D spin-glass problems. Benchmarking PAOA against QAOA on the canonical 26-spin Sherrington-Kirkpatrick model with matched parameters reveals superior performance for PAOA. We show that PAOA naturally extends simulated annealing by optimizing multiple temperature profiles, leading to improved performance over SA on heavy-tailed problems such as SK-L\'evy.
Authors: Joshua Murphy, Conor Rosato, Andrew Millard, Lee Devlin, Paul Horridge, Simon Maskell
Abstract: When performing Bayesian inference using Sequential Monte Carlo (SMC) methods, two considerations arise: the accuracy of the posterior approximation and computational efficiency. To address computational demands, Sequential Monte Carlo Squared (SMC$^2$) is well-suited for high-performance computing (HPC) environments. The design of the proposal distribution within SMC$^2$ can improve accuracy and exploration of the posterior as poor proposals may lead to high variance in importance weights and particle degeneracy. The Metropolis-Adjusted Langevin Algorithm (MALA) uses gradient information so that particles preferentially explore regions of higher probability. In this paper, we extend this idea by incorporating second-order information, specifically the Hessian of the log-target. While second-order proposals have been explored previously in particle Markov Chain Monte Carlo (p-MCMC) methods, we are the first to introduce them within the SMC$^2$ framework. Second-order proposals not only use the gradient (first-order derivative), but also the curvature (second-order derivative) of the target distribution. Experimental results on synthetic models highlight the benefits of our approach in terms of step-size selection and posterior approximation accuracy when compared to other proposals.
Authors: Haojie Liu, Zihan Lin
Abstract: Time-series models like ARIMA remain widely used for forecasting but limited to linear assumptions and high computational cost in large and complex datasets. We propose Galerkin-ARIMA that generalizes the AR component of ARIMA and replace it with a flexible spline-based function estimated by Galerkin projection. This enables the model to capture nonlinear dependencies in lagged values and retain the MA component and Gaussian noise assumption. We derive a closed-form OLS estimator for the Galerkin coefficients and show the model is asymptotically unbiased and consistent under standard conditions. Our method bridges classical time-series modeling and nonparametric regression, which offering improved forecasting performance and computational efficiency.
Authors: Kaiqu Liang, Haimin Hu, Xuandong Zhao, Dawn Song, Thomas L. Griffiths, Jaime Fern\'andez Fisac
Abstract: Bullshit, as conceptualized by philosopher Harry Frankfurt, refers to statements made without regard to their truth value. While previous work has explored large language model (LLM) hallucination and sycophancy, we propose machine bullshit as an overarching conceptual framework that can allow researchers to characterize the broader phenomenon of emergent loss of truthfulness in LLMs and shed light on its underlying mechanisms. We introduce the Bullshit Index, a novel metric quantifying LLMs' indifference to truth, and propose a complementary taxonomy analyzing four qualitative forms of bullshit: empty rhetoric, paltering, weasel words, and unverified claims. We conduct empirical evaluations on the Marketplace dataset, the Political Neutrality dataset, and our new BullshitEval benchmark (2,400 scenarios spanning 100 AI assistants) explicitly designed to evaluate machine bullshit. Our results demonstrate that model fine-tuning with reinforcement learning from human feedback (RLHF) significantly exacerbates bullshit and inference-time chain-of-thought (CoT) prompting notably amplify specific bullshit forms, particularly empty rhetoric and paltering. We also observe prevalent machine bullshit in political contexts, with weasel words as the dominant strategy. Our findings highlight systematic challenges in AI alignment and provide new insights toward more truthful LLM behavior.
Authors: Marie-Christine Pali, Christina Schwaiger, Malik Galijasevic, Valentin K. Ladenhauf, Stephanie Mangesius, Elke R. Gizewski
Abstract: The analysis of carotid arteries, particularly plaques, in multi-sequence Magnetic Resonance Imaging (MRI) data is crucial for assessing the risk of atherosclerosis and ischemic stroke. In order to evaluate metrics and radiomic features, quantifying the state of atherosclerosis, accurate segmentation is important. However, the complex morphology of plaques and the scarcity of labeled data poses significant challenges. In this work, we address these problems and propose a semi-supervised deep learning-based approach designed to effectively integrate multi-sequence MRI data for the segmentation of carotid artery vessel wall and plaque. The proposed algorithm consists of two networks: a coarse localization model identifies the region of interest guided by some prior knowledge on the position and number of carotid arteries, followed by a fine segmentation model for precise delineation of vessel walls and plaques. To effectively integrate complementary information across different MRI sequences, we investigate different fusion strategies and introduce a multi-level multi-sequence version of U-Net architecture. To address the challenges of limited labeled data and the complexity of carotid artery MRI, we propose a semi-supervised approach that enforces consistency under various input transformations. Our approach is evaluated on 52 patients with arteriosclerosis, each with five MRI sequences. Comprehensive experiments demonstrate the effectiveness of our approach and emphasize the role of fusion point selection in U-Net-based architectures. To validate the accuracy of our results, we also include an expert-based assessment of model performance. Our findings highlight the potential of fusion strategies and semi-supervised learning for improving carotid artery segmentation in data-limited MRI applications.
Authors: Keqin Bao, Nuo Chen, Xiaoyuan Li, Binyuan Hui, Bowen Yu, Fuli Feng, Junyang Lin, Xiangnan He, Dayiheng Liu
Abstract: Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.
Authors: Binxu Li, Minkai Xu, Meihua Dang, Stefano Ermon
Abstract: Diffusion models have achieved remarkable success in generating realistic and versatile images from text prompts. Inspired by the recent advancements of language models, there is an increasing interest in further improving the models by aligning with human preferences. However, we investigate alignment from a divergence minimization perspective and reveal that existing preference optimization methods are typically trapped in suboptimal mean-seeking optimization. In this paper, we introduce Divergence Minimization Preference Optimization (DMPO), a novel and principled method for aligning diffusion models by minimizing reverse KL divergence, which asymptotically enjoys the same optimization direction as original RL. We provide rigorous analysis to justify the effectiveness of DMPO and conduct comprehensive experiments to validate its empirical strength across both human evaluations and automatic metrics. Our extensive results show that diffusion models fine-tuned with DMPO can consistently outperform or match existing techniques, specifically outperforming all existing diffusion alignment baselines by at least 64.6% in PickScore across all evaluation datasets, demonstrating the method's superiority in aligning generative behavior with desired outputs. Overall, DMPO unlocks a robust and elegant pathway for preference alignment, bridging principled theory with practical performance in diffusion models.
Authors: Oliver Eberle, Thomas McGee, Hamza Giaffar, Taylor Webb, Ida Momennejad
Abstract: What algorithms do LLMs actually learn and use to solve problems? Studies addressing this question are sparse, as research priorities are focused on improving performance through scale, leaving a theoretical and empirical gap in understanding emergent algorithms. This position paper proposes AlgEval: a framework for systematic research into the algorithms that LLMs learn and use. AlgEval aims to uncover algorithmic primitives, reflected in latent representations, attention, and inference-time compute, and their algorithmic composition to solve task-specific problems. We highlight potential methodological paths and a case study toward this goal, focusing on emergent search algorithms. Our case study illustrates both the formation of top-down hypotheses about candidate algorithms, and bottom-up tests of these hypotheses via circuit-level analysis of attention patterns and hidden states. The rigorous, systematic evaluation of how LLMs actually solve tasks provides an alternative to resource-intensive scaling, reorienting the field toward a principled understanding of underlying computations. Such algorithmic explanations offer a pathway to human-understandable interpretability, enabling comprehension of the model's internal reasoning performance measures. This can in turn lead to more sample-efficient methods for training and improving performance, as well as novel architectures for end-to-end and multi-agent systems.
Authors: Mohamed Siala, Jordi Planes, Joao Marques-Silva
Abstract: A task of interest in machine learning (ML) is that of ascribing explanations to the predictions made by ML models. Furthermore, in domains deemed high risk, the rigor of explanations is paramount. Indeed, incorrect explanations can and will mislead human decision makers. As a result, and even if interpretability is acknowledged as an elusive concept, so-called interpretable models are employed ubiquitously in high-risk uses of ML and data mining (DM). This is the case for rule-based ML models, which encompass decision trees, diagrams, sets and lists. This paper relates explanations with well-known undesired facets of rule-based ML models, which include negative overlap and several forms of redundancy. The paper develops algorithms for the analysis of these undesired facets of rule-based systems, and concludes that well-known and widely used tools for learning rule-based ML models will induce rule sets that exhibit one or more negative facets.
Authors: Cooper Doyle
Abstract: We reveal a hidden Bayesian core of discrete-diffusion language models by showing that the expected denoiser output under the forward masking distribution recovers the exact posterior over clean tokens. Under minimal assumptions, Monte Carlo marginalization over K independent corruptions converges to this posterior at rate O(1/sqrt(K)), yielding a simple proof of consistency and finite-sample error bounds. Building on this insight, we introduce a lightweight inference-time ensemble that averages K mask-and-denoise passes to obtain posterior-aware token probabilities and uncertainty estimates at no extra training cost. On WikiText-2, our method achieves test perplexity 8.8 with K=8, versus 20.3 for GPT-2 Small, despite using a model of comparable size. Code is available at https://github.com/mercury0100/bayesradd.
Authors: Rados{\l}aw Adamczak
Abstract: We prove concentration inequalities for several models of non-linear random matrices. As corollaries we obtain estimates for linear spectral statistics of the conjugate kernel of neural networks and non-commutative polynomials in (possibly dependent) random matrices.
Authors: Joyeeta Datta, Niclas Doll, Qusai Ramadan, Zeyd Boukhers
Abstract: Large Language Models (LLMs) have demonstrated outstanding performance across a range of NLP tasks, however, their computational demands hinder their deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using Knowledge Distillation (KD) while maintaining strong performance on Question Answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90% of their teacher models' performance while reducing parameter counts by up to 57.1%. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for resource-constrained applications.
Authors: Seyed Reza Nabavi, Zonglin Guo, Zhiyuan Wang
Abstract: This study presents an integrated modeling and optimization framework for a steam methane reforming (SMR) reactor, combining a mathematical model, artificial neural network (ANN)-based hybrid modeling, advanced multi-objective optimization (MOO) and multi-criteria decision-making (MCDM) techniques. A one-dimensional fixed-bed reactor model accounting for internal mass transfer resistance was employed to simulate reactor performance. To reduce the high computational cost of the mathematical model, a hybrid ANN surrogate was constructed, achieving a 93.8% reduction in average simulation time while maintaining high predictive accuracy. The hybrid model was then embedded into three MOO scenarios using the non-dominated sorting genetic algorithm II (NSGA-II) solver: 1) maximizing methane conversion and hydrogen output; 2) maximizing hydrogen output while minimizing carbon dioxide emissions; and 3) a combined three-objective case. The optimal trade-off solutions were further ranked and selected using two MCDM methods: technique for order of preference by similarity to ideal solution (TOPSIS) and simplified preference ranking on the basis of ideal-average distance (sPROBID). Optimal results include a methane conversion of 0.863 with 4.556 mol/s hydrogen output in the first case, and 0.988 methane conversion with 3.335 mol/s hydrogen and 0.781 mol/s carbon dioxide in the third. This comprehensive methodology offers a scalable and effective strategy for optimizing complex catalytic reactor systems with multiple, often conflicting, objectives.
Authors: Felix Frohnert, Denny Lane B. Sombrillo, Evert van Nieuwenburg, Patrick Emonts
Abstract: Matching theoretical predictions to experimental data remains a central challenge in hadron spectroscopy. In particular, the identification of new hadronic states is difficult, as exotic signals near threshold can arise from a variety of physical mechanisms. A key diagnostic in this context is the pole structure of the scattering amplitude, but different configurations can produce similar signatures. The mapping between pole configurations and line shapes is especially ambiguous near the mass threshold, where analytic control is limited. In this work, we introduce an uncertainty-aware machine learning approach for classifying pole structures in $S$-matrix elements. Our method is based on an ensemble of classifier chains that provide both epistemic and aleatoric uncertainty estimates. We apply a rejection criterion based on predictive uncertainty, achieving a validation accuracy of nearly $95\%$ while discarding only a small fraction of high-uncertainty predictions. Trained on synthetic data with known pole structures, the model generalizes to previously unseen experimental data, including enhancements associated with the $P_{c\bar{c}}(4312)^+$ state observed by LHCb. In this, we infer a four-pole structure, representing the presence of a genuine compact pentaquark in the presence of a higher channel virtual state pole with non-vanishing width. While evaluated on this particular state, our framework is broadly applicable to other candidate hadronic states and offers a scalable tool for pole structure inference in scattering amplitudes.
Authors: Jude Haris, Jos\'e Cano
Abstract: Transposed Convolutions (TCONV) enable the up-scaling mechanism within generative Artificial Intelligence (AI) models. However, the predominant Input-Oriented Mapping (IOM) method for implementing TCONV has complex output mapping, overlapping sums, and ineffectual computations. These inefficiencies further exacerbate the performance bottleneck of TCONV and generative models on resource-constrained edge devices. To address this problem, in this paper we propose MM2IM, a hardware-software co-designed accelerator that combines Matrix Multiplication (MatMul) with col2IM to process TCONV layers on resource-constrained edge devices efficiently. Using the SECDA-TFLite design toolkit, we implement MM2IM and evaluate its performance across 261 TCONV problem configurations, achieving an average speedup of 1.9x against a dual-thread ARM Neon optimized CPU baseline. We then evaluate the performance of MM2IM on a range of TCONV layers from well-known generative models achieving up to 4.2x speedup, and compare it against similar resource-constrained TCONV accelerators, outperforming them by at least 2x GOPs/DSP. Finally, we evaluate MM2IM on the DCGAN and pix2pix GAN models, achieving up to 3x speedup and 2.4x energy reduction against the CPU baseline.
Authors: Shin'ya Yamaguchi, Kosuke Nishida, Daiki Chijiwa
Abstract: Large vision-language models (LVLMs) have demonstrated remarkable capabilities by integrating pre-trained vision encoders with large language models (LLMs). Similar to single-modal LLMs, chain-of-thought (CoT) prompting has been adapted for LVLMs to enhance multi-modal reasoning by generating intermediate rationales based on visual and textual inputs. While CoT is assumed to improve grounding and accuracy in LVLMs, our experiments reveal a key challenge: existing LVLMs often ignore the contents of generated rationales in CoT reasoning. To address this, we re-formulate multi-modal CoT reasoning as a KL-constrained reward maximization focused on rationale-conditional log-likelihood. As the optimal solution, we propose rationale-enhanced decoding (RED), a novel plug-and-play inference-time decoding strategy. RED harmonizes visual and rationale information by multiplying distinct image-conditional and rationale-conditional next token distributions. Extensive experiments show that RED consistently and significantly improves reasoning over standard CoT and other decoding methods across multiple benchmarks and LVLMs. Our work offers a practical and effective approach to improve both the faithfulness and accuracy of CoT reasoning in LVLMs, paving the way for more reliable rationale-grounded multi-modal systems.
Authors: Julio Garrido, Javier Vales, Diego Silva-Mu\~niz, Enrique Riveiro, Pablo L\'opez-Matencio, Josu\'e Rivera-Andrade
Abstract: Cable-Driven Parallel Robots (CDPRs) are increasingly used for load manipulation tasks involving predefined toolpaths with intermediate stops. At each stop, where the platform maintains a fixed pose and the motors keep the cables under tension, the system must evaluate whether it is safe to proceed by detecting anomalies that could compromise performance (e.g., wind gusts or cable impacts). This paper investigates whether anomalies can be detected using only motor torque data, without additional sensors. It introduces an adaptive, unsupervised outlier detection algorithm based on Gaussian Mixture Models (GMMs) to identify anomalies from torque signals. The method starts with a brief calibration period, just a few seconds, during which a GMM is fit on known anomaly-free data. Real-time torque measurements are then evaluated using Mahalanobis distance from the GMM, with statistically derived thresholds triggering anomaly flags. Model parameters are periodically updated using the latest segments identified as anomaly-free to adapt to changing conditions. Validation includes 14 long-duration test sessions simulating varied wind intensities. The proposed method achieves a 100% true positive rate and 95.4% average true negative rate, with 1-second detection latency. Comparative evaluation against power threshold and non-adaptive GMM methods indicates higher robustness to drift and environmental variation.
Authors: Jakub Kry\'s, Yashvardhan Sharma, Janet Egan
Abstract: Advances in low-communication training algorithms are enabling a shift from centralised model training to compute setups that are either distributed across multiple clusters or decentralised via community-driven contributions. This paper distinguishes these two scenarios - distributed and decentralised training - which are little understood and often conflated in policy discourse. We discuss how they could impact technical AI governance through an increased risk of compute structuring, capability proliferation, and the erosion of detectability and shutdownability. While these trends foreshadow a possible new paradigm that could challenge key assumptions of compute governance, we emphasise that certain policy levers, like export controls, remain relevant. We also acknowledge potential benefits of decentralised AI, including privacy-preserving training runs that could unlock access to more data, and mitigating harmful power concentration. Our goal is to support more precise policymaking around compute, capability proliferation, and decentralised AI development.
Authors: Shuying Huang, Junpeng Li, Changchun Hua, Yana Yang
Abstract: To alleviate the annotation burden in supervised learning, N-tuples learning has recently emerged as a powerful weakly-supervised method. While existing N-tuples learning approaches extend pairwise learning to higher-order comparisons and accommodate various real-world scenarios, they often rely on task-specific designs and lack a unified theoretical foundation. In this paper, we propose a general N-tuples learning framework based on empirical risk minimization, which systematically integrates pointwise unlabeled data to enhance learning performance. This paper first unifies the data generation processes of N-tuples and pointwise unlabeled data under a shared probabilistic formulation. Based on this unified view, we derive an unbiased empirical risk estimator that generalizes a broad class of existing N-tuples models. We further establish a generalization error bound for theoretical support. To demonstrate the flexibility of the framework, we instantiate it in four representative weakly supervised scenarios, each recoverable as a special case of our general model. Additionally, to address overfitting issues arising from negative risk terms, we adopt correction functions to adjust the empirical risk. Extensive experiments on benchmark datasets validate the effectiveness of the proposed framework and demonstrate that leveraging pointwise unlabeled data consistently improves generalization across various N-tuples learning tasks.
Authors: Egor Bakaev, Florestan Brunck, Amir Yehudayoff
Abstract: We study approximations of polytopes in the standard model for computing polytopes using Minkowski sums and (convex hulls of) unions. Specifically, we study the ability to approximate a target polytope by polytopes of a given depth. Our main results imply that simplices can only be ``trivially approximated''. On the way, we obtain a characterization of simplices as the only ``outer additive'' convex bodies.
Authors: Anwoy Chatterjee, H S V N S Kowndinya Renduchintala, Sumit Bhatia, Tanmoy Chakraborty
Abstract: Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.
Authors: Dimitrios Bralios, Jonah Casebeer, Paris Smaragdis
Abstract: Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most are trained to maximize reconstruction fidelity, often neglecting the specific latent structure necessary for optimal performance in diverse downstream applications. We propose a simple, post-hoc framework to address this by modifying the bottleneck of a pre-trained autoencoder. Our method introduces a "Re-Bottleneck", an inner bottleneck trained exclusively through latent space losses to instill user-defined structure. We demonstrate the framework's effectiveness in three experiments. First, we enforce an ordering on latent channels without sacrificing reconstruction quality. Second, we align latents with semantic embeddings, analyzing the impact on downstream diffusion modeling. Third, we introduce equivariance, ensuring that a filtering operation on the input waveform directly corresponds to a specific transformation in the latent space. Ultimately, our Re-Bottleneck framework offers a flexible and efficient way to tailor representations of neural audio models, enabling them to seamlessly meet the varied demands of different applications with minimal additional training.
Authors: Toluwani Aremu, Noor Hussein, Munachiso Nwadike, Samuele Poppi, Jie Zhang, Karthik Nandakumar, Neil Gong, Nils Lukas
Abstract: Watermarking offers a promising solution for GenAI providers to establish the provenance of their generated content. A watermark is a hidden signal embedded in the generated content, whose presence can later be verified using a secret watermarking key. A threat to GenAI providers are \emph{watermark stealing} attacks, where users forge a watermark into content that was \emph{not} generated by the provider's models without access to the secret key, e.g., to falsely accuse the provider. Stealing attacks collect \emph{harmless} watermarked samples from the provider's model and aim to maximize the expected success rate of generating \emph{harmful} watermarked samples. Our work focuses on mitigating stealing attacks while treating the underlying watermark as a black-box. Our contributions are: (i) Proposing a multi-key extension to mitigate stealing attacks that can be applied post-hoc to any watermarking method across any modality. (ii) We provide theoretical guarantees and demonstrate empirically that our method makes forging substantially less effective across multiple datasets, and (iii) we formally define the threat of watermark forging as the task of generating harmful, watermarked content and model this threat via security games.
Authors: Daniel Betschinske, Steven Peters
Abstract: The safety validation of automatic emergency braking system (AEBS) requires accurately distinguishing between false positive (FP) and true positive (TP) system activations. While simulations allow straightforward differentiation by comparing scenarios with and without interventions, analyzing activations from open-loop resimulations - such as those from field operational testing (FOT) - is more complex. This complexity arises from scenario parameter uncertainty and the influence of driver interventions in the recorded data. Human labeling is frequently used to address these challenges, relying on subjective assessments of intervention necessity or situational criticality, potentially introducing biases and limitations. This work proposes a rule-based classification approach leveraging the Prediction Divergence Principle (PDP) to address those issues. Applied to a simplified AEBS, the proposed method reveals key strengths, limitations, and system requirements for effective implementation. The findings suggest that combining this approach with human labeling may enhance the transparency and consistency of classification, thereby improving the overall validation process. While the rule set for classification derived in this work adopts a conservative approach, the paper outlines future directions for refinement and broader applicability. Finally, this work highlights the potential of such methods to complement existing practices, paving the way for more reliable and reproducible AEBS validation frameworks.
Authors: Chen Feng, Yicheng Lin, Shaojie Zhuo, Chenzheng Su, Ramchalam Kinattinkara Ramakrishnan, Zhaocong Yuan, Xiaopeng Zhang
Abstract: Recent advances in Automatic Speech Recognition (ASR) have demonstrated remarkable accuracy and robustness in diverse audio applications, such as live transcription and voice command processing. However, deploying these models on resource constrained edge devices (e.g., IoT device, wearables) still presents substantial challenges due to strict limits on memory, compute and power. Quantization, particularly Post-Training Quantization (PTQ), offers an effective way to reduce model size and inference cost without retraining. Despite its importance, the performance implications of various advanced quantization methods and bit-width configurations on ASR models remain unclear. In this work, we present a comprehensive benchmark of eight state-of-the-art (SOTA) PTQ methods applied to two leading edge-ASR model families, Whisper and Moonshine. We systematically evaluate model performances (i.e., accuracy, memory I/O and bit operations) across seven diverse datasets from the open ASR leaderboard, analyzing the impact of quantization and various configurations on both weights and activations. Built on an extension of the LLM compression toolkit, our framework integrates edge-ASR models, diverse advanced quantization algorithms, a unified calibration and evaluation data pipeline, and detailed analysis tools. Our results characterize the trade-offs between efficiency and accuracy, demonstrating that even 3-bit quantization can succeed on high capacity models when using advanced PTQ techniques. These findings provide valuable insights for optimizing ASR models on low-power, always-on edge devices.
Authors: Francesca Mignacco, Francesco Mori
Abstract: Learning is a complex dynamical process shaped by a range of interconnected decisions. Careful design of hyperparameter schedules for artificial neural networks or efficient allocation of cognitive resources by biological learners can dramatically affect performance. Yet, theoretical understanding of optimal learning strategies remains sparse, especially due to the intricate interplay between evolving meta-parameters and nonlinear learning dynamics. The search for optimal protocols is further hindered by the high dimensionality of the learning space, often resulting in predominantly heuristic, difficult to interpret, and computationally demanding solutions. Here, we combine statistical physics with control theory in a unified theoretical framework to identify optimal protocols in prototypical neural network models. In the high-dimensional limit, we derive closed-form ordinary differential equations that track online stochastic gradient descent through low-dimensional order parameters. We formulate the design of learning protocols as an optimal control problem directly on the dynamics of the order parameters with the goal of minimizing the generalization error at the end of training. This framework encompasses a variety of learning scenarios, optimization constraints, and control budgets. We apply it to representative cases, including optimal curricula, adaptive dropout regularization and noise schedules in denoising autoencoders. We find nontrivial yet interpretable strategies highlighting how optimal protocols mediate crucial learning tradeoffs, such as maximizing alignment with informative input directions while minimizing noise fitting. Finally, we show how to apply our framework to real datasets. Our results establish a principled foundation for understanding and designing optimal learning protocols and suggest a path toward a theory of meta-learning grounded in statistical physics.
Authors: Juan Pablo Oberhauser, Daniel Grzenda
Abstract: Continuous, automated monitoring of laboratory mice enables more accurate data collection and improves animal welfare through real-time insights. Researchers can achieve a more dynamic and clinically relevant characterization of disease progression and therapeutic effects by integrating behavioral and physiological monitoring in the home cage. However, providing individual mouse metrics is difficult because of their housing density, similar appearances, high mobility, and frequent interactions. To address these challenges, we develop a real-time identification (ID) algorithm that accurately assigns ID predictions to mice wearing custom ear tags in digital home cages monitored by cameras. Our pipeline consists of three parts: (1) a custom multiple object tracker (MouseTracks) that combines appearance and motion cues from mice; (2) a transformer-based ID classifier (Mouseformer); and (3) a tracklet associator linear program to assign final ID predictions to tracklets (MouseMap). Our models assign an animal ID based on custom ear tags at 30 frames per second with 24/7 cage coverage. We show that our custom tracking and ID pipeline improves tracking efficiency and lowers ID switches across mouse strains and various environmental factors compared to current mouse tracking methods.
Authors: Sizhen Bian, Mengxi Liu, Vitor Fortes Rey, Daniel Geissler, Paul Lukowicz
Abstract: Human Activity Recognition (HAR) on resource-constrained wearable devices demands inference models that harmonize accuracy with computational efficiency. This paper introduces TinierHAR, an ultra-lightweight deep learning architecture that synergizes residual depthwise separable convolutions, gated recurrent units (GRUs), and temporal aggregation to achieve SOTA efficiency without compromising performance. Evaluated across 14 public HAR datasets, TinierHAR reduces Parameters by 2.7x (vs. TinyHAR) and 43.3x (vs. DeepConvLSTM), and MACs by 6.4x and 58.6x, respectively, while maintaining the averaged F1-scores. Beyond quantitative gains, this work provides the first systematic ablation study dissecting the contributions of spatial-temporal components across proposed TinierHAR, prior SOTA TinyHAR, and the classical DeepConvLSTM, offering actionable insights for designing efficient HAR systems. We finally discussed the findings and suggested principled design guidelines for future efficient HAR. To catalyze edge-HAR research, we open-source all materials in this work for future benchmarking\footnote{https://github.com/zhaxidele/TinierHAR}
Authors: Noam Razin, Yong Lin, Jiarui Yao, Sanjeev Arora
Abstract: Reward models are key to language model post-training and inference pipelines. Conveniently, recent work showed that every language model defines an implicit reward model (IM-RM), without requiring any architectural changes. However, such IM-RMs tend to generalize worse, especially out-of-distribution, compared to explicit reward models (EX-RMs) that apply a dedicated linear head over the hidden representations of a language model. The existence of a generalization gap is puzzling, as EX-RMs and IM-RMs are nearly identical. They can be trained using the same data, loss function, and language model, and differ only in how the reward is computed. Towards a fundamental understanding of the implicit biases underlying different reward model types, we investigate the root cause of this gap. Our main finding, backed by theory and experiments, is that IM-RMs rely more heavily on superficial token-level cues. Consequently, they often generalize worse than EX-RMs under token-level distribution shifts, as well as in-distribution. Furthermore, we provide evidence against alternative hypotheses for the generalization gap. Most notably, we challenge the intuitive claim that IM-RMs struggle in tasks where generation is harder than verification because they can operate both as a verifier and a generator. Taken together, our results highlight that seemingly minor design choices can substantially impact the generalization behavior of reward models.
Authors: Shivam Duggal, Sanghyun Byun, William T. Freeman, Antonio Torralba, Phillip Isola
Abstract: According to Algorithmic Information Theory (AIT) -- Intelligent representations compress data into the shortest possible program that can reconstruct its content, exhibiting low Kolmogorov Complexity (KC). In contrast, most visual representation learning systems use fixed-length representations for all inputs, ignoring variations in complexity or familiarity. Recent adaptive tokenization methods address this by allocating variable-length representations but typically require test-time search over multiple encodings to find the most predictive one. Inspired by Kolmogorov Complexity principles, we propose a single-pass adaptive tokenizer, KARL, which predicts the appropriate number of tokens for an image in a single forward pass, halting once its approximate KC is reached. The token count serves as a proxy for the minimum description length. KARL's training procedure closely resembles the Upside-Down Reinforcement Learning paradigm, as it learns to conditionally predict token halting based on a desired reconstruction quality. KARL matches the performance of recent adaptive tokenizers while operating in a single pass. We present scaling laws for KARL, analyzing the role of encoder/decoder size, continuous vs. discrete tokenization and more. Additionally, we offer a conceptual study drawing an analogy between Adaptive Image Tokenization and Algorithmic Information Theory, examining the predicted image complexity (KC) across axes such as structure vs. noise and in- vs. out-of-distribution familiarity -- revealing alignment with human intuition.
Authors: Helen Qu, Sang Michael Xie
Abstract: CLIP and large multimodal models (LMMs) have better accuracy on examples involving concepts that are highly represented in the training data. However, the role of concept combinations in the training data on compositional generalization is largely unclear -- for instance, how does accuracy vary when a common object appears in an uncommon pairing with another object? In this paper, we investigate how word co-occurrence statistics in the pretraining dataset (a proxy for co-occurrence of visual concepts) impacts CLIP/LMM performance. To disentangle the effects of word co-occurrence frequencies from single-word frequencies, we measure co-occurrence with pointwise mutual information (PMI), which normalizes the joint probability of two words co-occurring by the probability of co-occurring independently. Using synthetically generated images with a variety of concept pairs, we show a strong correlation between PMI in the CLIP pretraining data and zero-shot accuracy in CLIP models trained on LAION-400M (r=0.97 and 14% accuracy gap between images in the top and bottom 5% of PMI values), demonstrating that even accuracy on common concepts is affected by the combination of concepts in the image. Leveraging this finding, we reproduce this effect in natural images by editing them to contain pairs with varying PMI, resulting in a correlation of r=0.75. Finally, we demonstrate that this behavior in CLIP transfers to LMMs built on top of CLIP (r=0.70 for TextVQA, r=0.62 for VQAv2). Our findings highlight the need for algorithms and architectures that improve compositional generalization in multimodal models without scaling the training data combinatorially. Our code is available at https://github.com/helenqu/multimodal-pretraining-pmi.
URLs: https://github.com/helenqu/multimodal-pretraining-pmi.
Authors: Adrian de Wynter
Abstract: We present an algorithm for solving binary classification problems when the dataset is not fully representative of the problem being solved, and obtaining more data is not possible. It relies on a trained model with loose accuracy constraints, an iterative hyperparameter searching-and-pruning procedure over a search space $\Theta$, and a data-generating function. Our algorithm works by reconstructing up to homology the manifold on which lies the support of the underlying distribution. We provide an analysis on correctness and runtime complexity under ideal conditions and an extension to deep neural networks. In the former case, if $\size{\Theta}$ is the number of hyperparameter sets in the search space, this algorithm returns a solution that is up to $2(1 - {2^{-\size{\Theta}}})$ times better than simply training with an enumeration of $\Theta$ and picking the best model. As part of our analysis we also prove that an open cover of a dataset has the same homology as the manifold on which lies the support of the underlying probability distribution, if and only said dataset is learnable. This latter result acts as a formal argument to explain the effectiveness of data expansion techniques.
Authors: Simon Lutz, Daniil Kaminskyi, Florian Wittbold, Simon Dierl, Falk Howar, Barbara K\"onig, Emmanuel M\"uller, Daniel Neider
Abstract: Automata learning is a successful tool for many application domains such as robotics and automatic verification. Typically, automata learning techniques operate in a supervised learning setting (active or passive) where they learn a finite state machine in contexts where additional information, such as labeled system executions, is available. However, other settings, such as learning from unlabeled data - an important aspect in machine learning - remain unexplored. To overcome this limitation, we propose a framework for learning a deterministic finite automaton (DFA) from a given multi-set of unlabeled words. We show that this problem is computationally hard and develop three learning algorithms based on constraint optimization. Moreover, we introduce novel regularization schemes for our optimization problems that improve the overall interpretability of our DFAs. Using a prototype implementation, we demonstrate practical feasibility in the context of unsupervised anomaly detection.
Authors: Xiaoling Zhou, Ou Wu, Michael K. Ng
Abstract: Machine learning models are prone to capturing the spurious correlations between non-causal attributes and classes, with counterfactual data augmentation being a promising direction for breaking these spurious associations. However, generating counterfactual data explicitly poses a challenge, and incorporating augmented data into the training process decreases training efficiency. This study proposes an Implicit Counterfactual Data Augmentation (ICDA) method to remove spurious correlations and make stable predictions. Specifically, first, a novel sample-wise augmentation strategy is developed that generates semantically and counterfactually meaningful deep features with distinct augmentation strength for each sample. Second, we derive an easy-to-compute surrogate loss on the augmented feature set when the number of augmented samples becomes infinite. Third, two concrete schemes are proposed, including direct quantification and meta-learning, to derive the key parameters for the robust loss. In addition, ICDA is explained from a regularization perspective, revealing its capacity to improve intra-class compactness and augment margins at both class and sample levels. Extensive experiments have been conducted across various biased learning scenarios covering both image and text datasets, demonstrating that ICDA consistently enhances the generalization and robustness performance of popular networks.
Authors: Zhiyi Dong, Yongyi Mao
Abstract: Adversarial attacks pose significant challenges to the robustness of modern deep neural networks in computer vision, and defending these networks against adversarial attacks has attracted intense research efforts. Among various defense strategies, preprocessing-based defenses are practically appealing since there is no need to train the network under protection. However, such approaches typically do not achieve comparable robustness as other methods such as adversarial training. In this paper, we propose a novel framework for preprocessing-based defenses, where a vector quantizer is used as a preprocessor. This framework, inspired by and extended from Randomized Discretization (RandDisc), is theoretically principled by rate-distortion theory: indeed, RandDisc may be viewed as a scalar quantizer, and rate-distortion theory suggests that such quantization schemes are inferior to vector quantization. In our framework, the preprocessing vector quantizer treats the input image as a collection of patches and finds a set of representative patches based on the patch distributions; each original patch is then modified according to the representative patches close to it. We present two lightweight defenses in this framework, referred to as patched RandDisc (pRD) and sliding-window RandDisc (swRD), where the patches are disjoint in the former and overlapping in the latter. We show that vector-quantization-based defenses have certifiable robust accuracy and that pRD and swRD demonstrate state-of-the-art performances, surpassing RandDisc by a large margin. Notably, the proposed defenses possess the obfuscated gradients property. Our experiments however show that pRD and swRD remain effective under the STE and EOT attacks, which are designed specifically for defenses with gradient obfuscation. ...
Authors: Pablo Millan Arias, Niousha Sadjadi, Monireh Safari, ZeMing Gong, Austin T. Wang, Joakim Bruslund Haurum, Iuliia Zarubiieva, Dirk Steinke, Lila Kari, Angel X. Chang, Scott C. Lowe, Graham W. Taylor
Abstract: In the global challenge of understanding and characterizing biodiversity, short species-specific genomic sequences known as DNA barcodes play a critical role, enabling fine-grained comparisons among organisms within the same kingdom of life. Although machine learning algorithms specifically designed for the analysis of DNA barcodes are becoming more popular, most existing methodologies rely on generic supervised training algorithms. We introduce BarcodeBERT, a family of models tailored to biodiversity analysis and trained exclusively on data from a reference library of 1.5M invertebrate DNA barcodes. We compared the performance of BarcodeBERT on taxonomic identification tasks against a spectrum of machine learning approaches including supervised training of classical neural architectures and fine-tuning of general DNA foundation models. Our self-supervised pretraining strategies on domain-specific data outperform fine-tuned foundation models, especially in identification tasks involving lower taxa such as genera and species. We also compared BarcodeBERT with BLAST, one of the most widely used bioinformatics tools for sequence searching, and found that our method matched BLAST's performance in species-level classification while being 55 times faster. Our analysis of masking and tokenization strategies also provides practical guidance for building customized DNA language models, emphasizing the importance of aligning model training strategies with dataset characteristics and domain knowledge. The code repository is available at https://github.com/bioscan-ml/BarcodeBERT.
Authors: Mohammad Ali Vahedifar, Azim Akhtarshenas, Mohammad Mohammadi Rafatpanah, Maryam Sabbaghian
Abstract: The K-Nearest Neighbors (KNN) algorithm is widely used for classification and regression; however, it suffers from limitations, including the equal treatment of all samples. We propose Information-Modified KNN (IM-KNN), a novel approach that leverages Mutual Information ($I$) and Shapley values to assign weighted values to neighbors, thereby bridging the gap in treating all samples with the same value and weight. On average, IM-KNN improves the accuracy, precision, and recall of traditional KNN by 16.80%, 17.08%, and 16.98%, respectively, across 12 benchmark datasets. Experiments on four large-scale datasets further highlight IM-KNN's robustness to noise, imbalanced data, and skewed distributions.
Authors: Andrea Apicella, Francesco Isgr\`o, Roberto Prevete
Abstract: Machine Learning (ML) has revolutionized various domains, offering predictive capabilities in several areas. However, with the increasing accessibility of ML tools, many practitioners, lacking deep ML expertise, adopt a "push the button" approach, utilizing user-friendly interfaces without a thorough understanding of underlying algorithms. While this approach provides convenience, it raises concerns about the reliability of outcomes, leading to challenges such as incorrect performance evaluation. This paper addresses a critical issue in ML, known as data leakage, where unintended information contaminates the training data, impacting model performance evaluation. Users, due to a lack of understanding, may inadvertently overlook crucial steps, leading to optimistic performance estimates that may not hold in real-world scenarios. The discrepancy between evaluated and actual performance on new data is a significant concern. In particular, this paper categorizes data leakage in ML, discussing how certain conditions can propagate through the ML workflow. Furthermore, it explores the connection between data leakage and the specific task being addressed, investigates its occurrence in Transfer Learning, and compares standard inductive ML with transductive ML frameworks. The conclusion summarizes key findings, emphasizing the importance of addressing data leakage for robust and reliable ML applications.
Authors: Wei-Cheng Huang, Chun-Fu Chen, Hsiang Hsu
Abstract: Recent works have shown that by using large pre-trained models along with learnable prompts, rehearsal-free methods for class-incremental learning (CIL) settings can achieve superior performance to prominent rehearsal-based ones. Rehearsal-free CIL methods struggle with distinguishing classes from different tasks, as those are not trained together. In this work we propose a regularization method based on virtual outliers to tighten decision boundaries of the classifier, such that confusion of classes among different tasks is mitigated. Recent prompt-based methods often require a pool of task-specific prompts, in order to prevent overwriting knowledge of previous tasks with that of the new task, leading to extra computation in querying and composing an appropriate prompt from the pool. This additional cost can be eliminated, without sacrificing accuracy, as we reveal in the paper. We illustrate that a simplified prompt-based method can achieve results comparable to previous state-of-the-art (SOTA) methods equipped with a prompt pool, using much less learnable parameters and lower inference cost. Our regularization method has demonstrated its compatibility with different prompt-based methods, boosting those previous SOTA rehearsal-free CIL methods' accuracy on the ImageNet-R and CIFAR-100 benchmarks. Our source code is available at https://github.com/jpmorganchase/ovor.
Authors: Ningyi Liao, Zihao Yu, Ruixiao Zeng, Siqiang Luo
Abstract: Graph Neural Networks (GNNs) have shown promising performance, but at the cost of resource-intensive operations on graph-scale matrices. To reduce computational overhead, previous studies attempt to sparsify the graph or network parameters, but with limited flexibility and precision boundaries. In this work, we propose Unifews, a joint sparsification technique to unify graph and weight matrix operations and enhance GNN learning efficiency. The Unifews design enables adaptive compression across GNN layers with progressively increased sparsity, and is applicable to a variety of architectures with on-the-fly simplification. Theoretically, we establish a novel framework to characterize sparsified GNN learning in view of the graph optimization process, showing that Unifews effectively approximates the learning objective with bounded error and reduced computational overhead. Extensive experiments demonstrate that Unifews achieves efficiency improvements with comparable or better accuracy, including 10-20x matrix operation reduction and up to 100x acceleration on graphs up to billion-edge scale.
Authors: Ziqi Zhao, Zhaochun Ren, Liu Yang, Yunsen Liang, Fajie Yuan, Pengjie Ren, Zhumin Chen, jun Ma, Xin Xin
Abstract: Offline reinforcement learning (RL) aims to learn policies without online explorations. To enlarge the training data, model-based offline RL learns a dynamics model which is utilized as a virtual environment to generate simulation data and enhance policy learning. However, existing data augmentation methods for offline RL suffer from (i) trivial improvement from short-horizon simulation; and (ii) the lack of evaluation and correction for generated data, leading to low-qualified augmentation. In this paper, we propose offline trajectory optimization for offline reinforcement learning (OTTO). The key motivation is to conduct long-horizon simulation and then utilize model uncertainty to evaluate and correct the augmented data. Specifically, we propose an ensemble of Transformers, a.k.a. World Transformers, to predict environment state dynamics and the reward function. Three strategies are proposed to use World Transformers to generate long-horizon trajectory simulation by perturbing the actions in the offline data. Then, an uncertainty-based World Evaluator is introduced to firstly evaluate the confidence of the generated trajectories and then perform the correction for low-confidence data. Finally, we jointly use the original data with the corrected augmentation data to train an offline RL algorithm. OTTO serves as a plug-in module and can be integrated with existing model-free offline RL methods. Experiments on various benchmarks show that OTTO can effectively improve the performance of representative offline RL algorithms, including in complex environments with sparse rewards like AntMaze. Codes are available at https://github.com/ZiqiZhao1/OTTO.
Authors: David Boetius, Stefan Leue, Tobias Sutter
Abstract: Probabilistic verification problems of neural networks are concerned with formally analysing the output distribution of a neural network under a probability distribution of the inputs. Examples of probabilistic verification problems include verifying the demographic parity fairness notion or quantifying the safety of a neural network. We present a new algorithm for solving probabilistic verification problems of neural networks based on an algorithm for computing and iteratively refining lower and upper bounds on probabilities over the outputs of a neural network. By applying state-of-the-art bound propagation and branch and bound techniques from non-probabilistic neural network verification, our algorithm significantly outpaces existing probabilistic verification algorithms, reducing solving times for various benchmarks from the literature from tens of minutes to tens of seconds. Furthermore, our algorithm compares favourably even to dedicated algorithms for restricted probabilistic verification problems. We complement our empirical evaluation with a theoretical analysis, proving that our algorithm is sound and, under mildly restrictive conditions, also complete when using a suitable set of heuristics.
Authors: Xiangyu Sun, Raquel Aoki, Kevin H. Wilson
Abstract: Machine learning (ML) methods have experienced significant growth in the past decade, yet their practical application in high-impact real-world domains has been hindered by their opacity. When ML methods are responsible for making critical decisions, stakeholders often require insights into how to alter these decisions. Counterfactual explanations (CFEs) have emerged as a solution, offering interpretations of opaque ML models and providing a pathway to transition from one decision to another. However, most existing CFE methods require access to the model's training dataset, few methods can handle multivariate time-series, and none of model-agnostic CFE methods can handle multivariate time-series without training datasets. These limitations can be formidable in many scenarios. In this paper, we present NTD-CFE, a novel model-agnostic CFE method based on reinforcement learning (RL) that generates CFEs when training datasets are unavailable. NTD-CFE is suitable for both static and multivariate time-series datasets with continuous and discrete features. NTD-CFE reduces the CFE search space from a multivariate time-series domain to a lower dimensional space and addresses the problem using RL. Users have the flexibility to specify non-actionable, immutable, and preferred features, as well as causal constraints. We demonstrate the performance of NTD-CFE against four baselines on several datasets and find that, despite not having access to a training dataset, NTD-CFE finds CFEs that make significantly fewer and significantly smaller changes to the input time-series. These properties make CFEs more actionable, as the magnitude of change required to alter an outcome is vastly reduced. The code is available in the supplementary material.
Authors: Saiyue Lyu, Shadab Shaikh, Frederick Shpilevskiy, Evan Shelhamer, Mathias L\'ecuyer
Abstract: We propose Adaptive Randomized Smoothing (ARS) to certify the predictions of our test-time adaptive models against adversarial examples. ARS extends the analysis of randomized smoothing using $f$-Differential Privacy to certify the adaptive composition of multiple steps. For the first time, our theory covers the sound adaptive composition of general and high-dimensional functions of noisy inputs. We instantiate ARS on deep image classification to certify predictions against adversarial examples of bounded $L_{\infty}$ norm. In the $L_{\infty}$ threat model, ARS enables flexible adaptation through high-dimensional input-dependent masking. We design adaptivity benchmarks, based on CIFAR-10 and CelebA, and show that ARS improves standard test accuracy by $1$ to $15\%$ points. On ImageNet, ARS improves certified test accuracy by up to $1.6\%$ points over standard RS without adaptivity. Our code is available at https://github.com/ubc-systopia/adaptive-randomized-smoothing .
URLs: https://github.com/ubc-systopia/adaptive-randomized-smoothing
Authors: Ziyue Chen, Tongya Zheng, Mingli Song
Abstract: Temporal networks are effective in capturing the evolving interactions of networks over time, such as social networks and e-commerce networks. In recent years, researchers have primarily concentrated on developing specific model architectures for Temporal Graph Neural Networks (TGNNs) in order to improve the representation quality of temporal nodes and edges. However, limited attention has been given to the quality of negative samples during the training of TGNNs. When compared with static networks, temporal networks present two specific challenges for negative sampling: positive sparsity and positive shift. Positive sparsity refers to the presence of a single positive sample amidst numerous negative samples at each timestamp, while positive shift relates to the variations in positive samples across different timestamps. To robustly address these challenges in training TGNNs, we introduce Curriculum Negative Mining (CurNM), a model-aware curriculum learning framework that adaptively adjusts the difficulty of negative samples. Within this framework, we first establish a dynamically updated negative pool that balances random, historical, and hard negatives to address the challenges posed by positive sparsity. Secondly, we implement a temporal-aware negative selection module that focuses on learning from the disentangled factors of recently active edges, thus accurately capturing shifting preferences. Finally, the selected negatives are combined with annealing random negatives to support stable training. Extensive experiments on 12 datasets and 3 TGNNs demonstrate that our method outperforms baseline methods by a significant margin. Additionally, thorough ablation studies and parameter sensitivity experiments verify the usefulness and robustness of our approach.
Authors: Karen Medlin, Sven Leyffer, Krishnan Raghavan
Abstract: Data rebalancing techniques, including oversampling and undersampling, are a common approach to addressing the challenges of imbalanced data. To tackle unresolved problems related to both oversampling and undersampling, we propose a new undersampling approach that: (i) avoids the pitfalls of noise and overlap caused by synthetic data and (ii) avoids the pitfall of under-fitting caused by random undersampling. Instead of undersampling majority data randomly, our method undersamples datapoints based on their ability to improve model loss. Using improved model loss as a proxy measurement for classification performance, our technique assesses a datapoint's impact on loss and rejects those unable to improve it. In so doing, our approach rejects majority datapoints redundant to datapoints already accepted and, thereby, finds an optimal subset of majority training data for classification. The accept/reject component of our algorithm is motivated by a bilevel optimization problem uniquely formulated to identify the optimal training set we seek. Experimental results show our proposed technique with F1 scores up to 10% higher than state-of-the-art methods.
Authors: \"Omer Veysel \c{C}a\u{g}atan, Bar{\i}\c{s} Akg\"un
Abstract: In this study, we investigate the effect of SSL objective modifications within the SPR framework, focusing on specific adjustments such as terminal state masking and prioritized replay weighting, which were not explicitly addressed in the original design. While these modifications are specific to RL, they are not universally applicable across all RL algorithms. Therefore, we aim to assess their impact on performance and explore other SSL objectives that do not accommodate these adjustments like Barlow Twins and VICReg. We evaluate six SPR variants on the Atari 100k benchmark, including versions both with and without these modifications. Additionally, we test the performance of these objectives on the DeepMind Control Suite, where such modifications are absent. Our findings reveal that incorporating specific SSL modifications within SPR significantly enhances performance, and this influence extends to subsequent frameworks like SR-SPR and BBF, highlighting the critical importance of SSL objective selection and related adaptations in achieving data efficiency in self-predictive reinforcement learning.
Authors: Zixuan He, Ziqian Kong, Zhengyu Chen, Yuling Zhan, Zijun Que, Zhengguo Xu
Abstract: Accurate remaining useful life (RUL) predictions are critical to the safe operation of aero-engines. Currently, the RUL prediction task is mainly a regression paradigm with only mean square error as the loss function and lacks research on feature space structure, the latter of which has shown excellent performance in a large number of studies. This paper develops a multi-granularity supervised contrastive (MGSC) framework from plain intuition that samples with the same RUL label should be aligned in the feature space, and address the problems of too large minibatch size and unbalanced samples in the implementation. The RUL prediction with MGSC is implemented on using the proposed multi-phase training strategy. This paper also demonstrates a simple and scalable basic network structure and validates the proposed MGSC strategy on the CMPASS dataset using a convolutional long short-term memory network as a baseline, which effectively improves the accuracy of RUL prediction.
Authors: Milin Zhang, Mohammad Abdi, Venkat R. Dasari, Francesco Restuccia
Abstract: Semantic Edge Computing (SEC) and Semantic Communications (SemComs) have been proposed as viable approaches to achieve real-time edge-enabled intelligence in sixth-generation (6G) wireless networks. On one hand, SemCom leverages the strength of Deep Neural Networks (DNNs) to encode and communicate the semantic information only, while making it robust to channel distortions by compensating for wireless effects. Ultimately, this leads to an improvement in the communication efficiency. On the other hand, SEC has leveraged distributed DNNs to divide the computation of a DNN across different devices based on their computational and networking constraints. Although significant progress has been made in both fields, the literature lacks a systematic view to connect both fields. In this work, we fulfill the current gap by unifying the SEC and SemCom fields. We summarize the research problems in these two fields and provide a comprehensive review of the state of the art with a focus on their technical strengths and challenges.
Authors: Akhila Vangara, Alex Egg
Abstract: Uniform random exploration in decision-making systems supports off-policy learning via supervision but incurs high regret, making it impractical for many applications. Conversely, non-uniform exploration offers better immediate performance but lacks support for off-policy learning. Recent research suggests that regression oracles can bridge this gap by combining non-uniform exploration with supervised learning. In this paper, we analyze these approaches within a real-world industrial context at Adyen, a large global payments processor characterized by batch logged delayed feedback, short-term memory, and dynamic action spaces under the Empirical Risk Minimization (ERM) framework. Our analysis reveals that while regression oracles significantly improve performance, they introduce challenges due to rigid algorithmic assumptions. Specifically, we observe that as a policy improves, subsequent generations may perform worse due to shifts in the reward distribution and increased class imbalance in the training data. This degradation occurs de spite improvements in other aspects of the training data, leading to decreased performance in successive policy iterations. We further explore the long-term impact of regression oracles, identifying a potential "oscillation effect." This effect arises when regression oracles influence probability estimates and the realizability of subsequent policy models, leading to fluctuations in performance across iterations. Our findings highlight the need for more adaptable algorithms that can leverage the benefits of regression oracles without introducing instability in policy performance over time.
Authors: Nathan Phelps, Daniel J. Lizotte, Douglas G. Woolford
Abstract: Imbalanced binary classification problems arise in many fields of study. When using machine learning models for these problems, it is common to subsample the majority class (i.e., undersampling) to create a (more) balanced dataset for model training. This biases the model's predictions because the model learns from a dataset that does not follow the same data generating process as new data. One way of accounting for this bias is to analytically map the resulting predictions to new values based on the sampling rate for the majority class, which was used to create the training dataset. While this approach may work well for some machine learning models, we show that calibrating a random forest this way has unintended negative consequences, including prevalence estimates that can be upwardly biased. These prevalence estimates depend on both i) the number of predictors considered at each split in the random forest; and ii) the sampling rate used. We explain the former using known properties of random forests and analytical calibration. However, in investigating the latter issue, we made a surprising discovery - contrary to the widespread belief that decision trees are biased towards the majority class, they actually can be biased towards the minority class.
Authors: Shuhei Watanabe
Abstract: Gaussian process (GP) is arguably one of the most widely used machine learning algorithms in practice. One of its prominent applications is Bayesian optimization (BO). Although the vanilla GP itself is already a powerful tool for BO, it is often beneficial to be able to consider the dependencies of multiple outputs. To do so, Multi-task GP (MTGP) is formulated, but it is not trivial to fully understand the derivations of its formulations and their gradients from the previous literature. This paper serves friendly derivations of the MTGP formulations and their gradients.
Authors: Isha Gupta, David Khachaturov, Robert Mullins
Abstract: The rise of multimodal large language models has introduced innovative human-machine interaction paradigms but also significant challenges in machine learning safety. Audio-Language Models (ALMs) are especially relevant due to the intuitive nature of spoken communication, yet little is known about their failure modes. This paper explores audio jailbreaks targeting ALMs, focusing on their ability to bypass alignment mechanisms. We construct adversarial perturbations that generalize across prompts, tasks, and even base audio samples, demonstrating the first universal jailbreaks in the audio modality, and show that these remain effective in simulated real-world conditions. Beyond demonstrating attack feasibility, we analyze how ALMs interpret these audio adversarial examples and reveal them to encode imperceptible first-person toxic speech - suggesting that the most effective perturbations for eliciting toxic outputs specifically embed linguistic features within the audio signal. These results have important implications for understanding the interactions between different modalities in multimodal models, and offer actionable insights for enhancing defenses against adversarial audio attacks.
Authors: David D. Baek, Ziming Liu, Riya Tyagi, Max Tegmark
Abstract: In this paper, we introduce harmonic loss as an alternative supervisory signal for training neural networks and large language models (LLMs). Harmonic loss differs from standard cross-entropy loss by (a) replacing the usual SoftMax normalization with a scale-invariant HarMax function and (b) computing logits via Euclidean distance rather than a dot product. Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design, which can be interpreted as a class center. We first validate the performance of harmonic models across algorithmic, vision, and language datasets. Through extensive experiments, we demonstrate that models trained with harmonic loss perform better than standard models by: (a) enhancing interpretability, (b) requiring less data for generalization, and (c) reducing grokking. Moreover, we compare a GPT-2 model trained with harmonic loss to the standard GPT-2, illustrating that the harmonic model develops more interpretable representations. Looking forward, we believe harmonic loss may become a valuable tool in domains with limited data availability or in high-stakes applications where interpretability and reliability are paramount, paving the way for more robust and efficient neural network models.
Authors: Hao Zeng, Kangdao Liu, Bingyi Jing, Hongxin Wei
Abstract: Conformal prediction is a popular framework of uncertainty quantification that constructs prediction sets with coverage guarantees. To uphold the exchangeability assumption, many conformal prediction methods necessitate an additional holdout set for parameter tuning. Yet, the impact of violating this principle on coverage remains underexplored, making it ambiguous in practical applications. In this work, we empirically find that the tuning bias - the coverage gap introduced by leveraging the same dataset for tuning and calibration, is negligible for simple parameter tuning in many conformal prediction methods. In particular, we observe the scaling law of the tuning bias: this bias increases with parameter space complexity and decreases with calibration set size. Formally, we establish a theoretical framework to quantify the tuning bias and provide rigorous proof for the scaling law of the tuning bias by deriving its upper bound. In the end, we discuss how to reduce the tuning bias, guided by the theories we developed.
Authors: Shahran Rahman Alve, Muhammad Zawad Mahmud, Samiha Islam, Md. Asaduzzaman Chowdhury, Jahirul Islam
Abstract: The Internet of Things (IoT) is expanding at an accelerated pace, making it critical to have secure networks to mitigate a variety of cyber threats. This study addresses the limitation of multi-class attack detection of IoT devices and presents new machine learning-based lightweight ensemble methods that exploit its strong machine learning framework. We used a dataset entitled CICIoT 2023, which has a total of 34 different attack types categorized into 10 categories, and methodically assessed the performance of a substantial array of current machine learning techniques in our goal to identify the best-performing algorithmic choice for IoT application protection. In this work, we focus on ML classifier-based methods to address the biocharges presented by the difficult and heterogeneous properties of the attack vectors in IoT ecosystems. The best-performing method was the Decision Tree, achieving 99.56% accuracy and 99.62% F1, indicating this model is capable of detecting threats accurately and reliably. The Random Forest model also performed nearly as well, with an accuracy of 98.22% and an F1 score of 98.24%, indicating that ML methods excel in a scenario of high-dimensional data. These findings emphasize the promise of integrating ML classifiers into the protective defenses of IoT devices and provide motivations for pursuing subsequent studies towards scalable, keystroke-based attack detection frameworks. We think that our approach offers a new avenue for constructing complex machine learning algorithms for low-resource IoT devices that strike a balance between accuracy requirements and time efficiency. In summary, these contributions expand and enhance the knowledge of the current IoT security literature, establishing a solid baseline and framework for smart, adaptive security to be used in IoT environments.
Authors: Jiyue Tao, Tongsheng Shen, Dexin Zhao, Feitian Zhang
Abstract: The target defense problem (TDP) for unmanned surface vehicles (USVs) concerns intercepting an adversarial USV before it breaches a designated target region, using one or more defending USVs. A particularly challenging scenario arises when the attacker exhibits superior maneuverability compared to the defenders, significantly complicating effective interception. To tackle this challenge, this letter introduces ARBoids, a novel adaptive residual reinforcement learning framework that integrates deep reinforcement learning (DRL) with the biologically inspired, force-based Boids model. Within this framework, the Boids model serves as a computationally efficient baseline policy for multi-agent coordination, while DRL learns a residual policy to adaptively refine and optimize the defenders' actions. The proposed approach is validated in a high-fidelity Gazebo simulation environment, demonstrating superior performance over traditional interception strategies, including pure force-based approaches and vanilla DRL policies. Furthermore, the learned policy exhibits strong adaptability to attackers with diverse maneuverability profiles, highlighting its robustness and generalization capability. The code of ARBoids will be released upon acceptance of this letter.
Authors: Jindong Li, Tim Hamann, Jens Barth, Peter K\"ampf, Dario Zanca, Bj\"orn Eskofier
Abstract: Online handwriting recognition (HWR) using data from inertial measurement units (IMUs) remains challenging due to variations in writing styles and the limited availability of annotated datasets. Previous approaches often struggle with handwriting from unseen writers, making writer-independent (WI) recognition a crucial yet difficult problem. This paper presents an HWR model designed to improve WI HWR on IMU data, using a CNN encoder and a BiLSTM-based decoder. Our approach demonstrates strong robustness to unseen handwriting styles, outperforming existing methods on the WI splits of both the public OnHW dataset and our word-based dataset, achieving character error rates (CERs) of 7.37\% and 9.44\%, and word error rates (WERs) of 15.12\% and 32.17\%, respectively. Robustness evaluation shows that our model maintains superior accuracy across different age groups, and knowledge learned from one group generalizes better to another. Evaluation on our sentence-based dataset further demonstrates its potential in recognizing full sentences. Through comprehensive ablation studies, we show that our design choices lead to a strong balance between performance and efficiency. These findings support the development of more adaptable and scalable HWR systems for real-world applications.
Authors: Andrew Gordon Wilson
Abstract: Deep neural networks are often seen as different from other model classes by defying conventional notions of generalization. Popular examples of anomalous generalization behaviour include benign overfitting, double descent, and the success of overparametrization. We argue that these phenomena are not distinct to neural networks, or particularly mysterious. Moreover, this generalization behaviour can be intuitively understood, and rigorously characterized, using long-standing generalization frameworks such as PAC-Bayes and countable hypothesis bounds. We present soft inductive biases as a key unifying principle in explaining these phenomena: rather than restricting the hypothesis space to avoid overfitting, embrace a flexible hypothesis space, with a soft preference for simpler solutions that are consistent with the data. This principle can be encoded in many model classes, and thus deep learning is not as mysterious or different from other model classes as it might seem. However, we also highlight how deep learning is relatively distinct in other ways, such as its ability for representation learning, phenomena such as mode connectivity, and its relative universality.
Authors: Yifan Ding, Arturas Aleksandraus, Amirhossein Ahmadian, Jonas Unger, Fredrik Lindsten, Gabriel Eilertsen
Abstract: Out-of-distribution (OOD) detection is critical for ensuring the reliability of deep learning systems, particularly in safety-critical applications. Likelihood-based deep generative models have historically faced criticism for their unsatisfactory performance in OOD detection, often assigning higher likelihood to OOD data than in-distribution samples when applied to image data. In this work, we demonstrate that likelihood is not inherently flawed. Rather, several properties in the images space prohibit likelihood as a valid detection score. Given a sufficiently good likelihood estimator, specifically using the probability flow formulation of a diffusion model, we show that likelihood-based methods can still perform on par with state-of-the-art methods when applied in the representation space of pre-trained encoders. The code of our work can be found at $\href{https://github.com/limchaos/Likelihood-OOD.git}{\texttt{https://github.com/limchaos/Likelihood-OOD.git}}$.
URLs: https://github.com/limchaos/Likelihood-OOD.git, https://github.com/limchaos/Likelihood-OOD.git
Authors: Lei Kang, Jia Li, Mi Tian, Hua Huang
Abstract: Sparsely activated Mixture-of-Experts (MoE) models effectively increase the number of parameters while maintaining consistent computational costs per token. However, vanilla MoE models often suffer from limited diversity and specialization among experts, constraining their performance and scalability, especially as the number of experts increases. In this paper, we present a novel perspective on vanilla MoE with top-$k$ routing inspired by sparse representation. This allows us to bridge established theoretical insights from sparse representation into MoE models. Building on this foundation, we propose a group sparse regularization approach for the input of top-$k$ routing, termed Mixture of Group Experts (MoGE). MoGE indirectly regularizes experts by imposing structural constraints on the routing inputs, while preserving the original MoE architecture. Furthermore, we organize the routing input into a 2D topographic map, spatially grouping neighboring elements. This structure enables MoGE to capture representations invariant to minor transformations, thereby significantly enhancing expert diversity and specialization. Comprehensive evaluations across various Transformer models for image classification and language modeling tasks demonstrate that MoGE substantially outperforms its MoE counterpart, with minimal additional memory and computation overhead. Our approach provides a simple yet effective solution to scale the number of experts and reduce redundancy among them. The source code is included in the supplementary material and will be publicly released.
Authors: Ivan Rossi, Flavio Sartori, Cesare Rollo, Giovanni Birolo, Piero Fariselli, Tiziana Sanavia
Abstract: Survival analysis often relies on Cox models, assuming both linearity and proportional hazards (PH). This study evaluates machine and deep learning methods that relax these constraints, comparing their performance with penalized Cox models on a benchmark of three synthetic and three real datasets. In total, eight different models were tested, including six non-linear models of which four were also non-PH. Although Cox regression often yielded satisfactory performance, we showed the conditions under which machine and deep learning models can perform better. Indeed, the performance of these methods has often been underestimated due to the improper use of Harrell's concordance index (C-index) instead of more appropriate scores such as Antolini's concordance index, which generalizes C-index in cases where the PH assumption does not hold. In addition, since occasionally high C-index models happen to be badly calibrated, combining Antolini's C-index with Brier's score is useful to assess the overall performance of a survival method. Results on our benchmark data showed that survival prediction should be approached by testing different methods to select the most appropriate one according to sample size, non-linearity and non-PH conditions. To allow an easy reproducibility of these tests on our benchmark data, code and documentation are freely available at https://github.com/compbiomed-unito/survhive.
Authors: Malhar A. Managoli, Vinod M. Prabhakaran, Suhas Diggavi
Abstract: Federated learning with heterogeneous data and personalization has received significant recent attention. Separately, robustness to corrupted data in the context of federated learning has also been studied. In this paper we explore combining personalization for heterogeneous data with robustness, where a constant fraction of the clients are corrupted. Motivated by this broad problem, we formulate a simple instantiation which captures some of its difficulty. We focus on the specific problem of personalized mean estimation where the data is drawn from a Gaussian mixture model. We give an algorithm whose error depends almost linearly on the ratio of corrupted to uncorrupted samples, and show a lower bound with the same behavior, albeit with a gap of a constant factor.
Authors: Greg Gluch, Shafi Goldwasser
Abstract: In this paper, we initiate a cryptographically inspired theoretical study of detection versus mitigation of adversarial inputs produced by attackers on Machine Learning algorithms during inference time. We formally define defense by detection (DbD) and defense by mitigation (DbM). Our definitions come in the form of a 3-round protocol between two resource-bounded parties: a trainer/defender and an attacker. The attacker aims to produce inference-time inputs that fool the training algorithm. We define correctness, completeness, and soundness properties to capture successful defense at inference time while not degrading (too much) the performance of the algorithm on inputs from the training distribution. We first show that achieving DbD and achieving DbM are equivalent for ML classification tasks. Surprisingly, this is not the case for ML generative learning tasks, where there are many possible correct outputs for each input. We show a separation between DbD and DbM by exhibiting two generative learning tasks for which it is possible to defend by mitigation but it is provably impossible to defend by detection. The mitigation phase uses significantly less computational resources than the initial training algorithm. In the first learning task we consider sample complexity as the resource and in the second the time complexity. The first result holds under the assumption that the Identity-Based Fully Homomorphic Encryption (IB-FHE), publicly-verifiable zero-knowledge Succinct Non-Interactive Arguments of Knowledge (zk-SNARK), and Strongly Unforgeable Signatures exist. The second result assumes the existence of Non-Parallelizing Languages with Average-Case Hardness (NPL) and Incrementally-Verifiable Computation (IVC) and IB-FHE.
Authors: Yonghong Li, Xiuzhuang Zhou
Abstract: Trustworthy depression prediction based on deep learning, incorporating both predictive reliability and algorithmic fairness across diverse demographic groups, is crucial for clinical application. Recently, achieving reliable depression predictions through uncertainty quantification has attracted increasing attention. However, few studies have focused on the fairness of uncertainty quantification (UQ) in depression prediction. In this work, we investigate the algorithmic fairness of UQ, namely Equal Opportunity Coverage (EOC) fairness, and propose Fair Uncertainty Quantification (FUQ) for depression prediction. FUQ pursues reliable and fair depression predictions through group-based analysis. Specifically, we first group all the participants by different sensitive attributes and leverage conformal prediction to quantify uncertainty within each demographic group, which provides a theoretically guaranteed and valid way to quantify uncertainty for depression prediction and facilitates the investigation of fairness across different demographic groups. Furthermore, we propose a fairness-aware optimization strategy that formulates fairness as a constrained optimization problem under EOC constraints. This enables the model to preserve predictive reliability while adapting to the heterogeneous uncertainty levels across demographic groups, thereby achieving optimal fairness. Through extensive evaluations on several visual and audio depression datasets, our approach demonstrates its effectiveness.
Authors: Juan Ramirez, Meraj Hashemizadeh, Simon Lacoste-Julien
Abstract: Recent efforts to develop trustworthy AI systems with accountability guarantees have led to widespread use of machine learning formulations incorporating external requirements, or constraints. These requirements are often enforced via penalization--adding fixed-weight terms to the task loss. We argue this approach is fundamentally ill-suited since there may be no penalty coefficient that simultaneously ensures constraint satisfaction and optimal constrained performance, i.e., that truly solves the constrained problem. Moreover, tuning these coefficients requires costly trial-and-error, incurring significant time and computational overhead. We, therefore, advocate for broader adoption of tailored constrained optimization methods--such as the Lagrangian approach, which jointly optimizes the penalization "coefficients" (the Lagrange multipliers) and the model parameters. Such methods (i) truly solve the constrained problem and do so accountably, by clearly defining feasibility and verifying when it is achieved, (ii) eliminate the need for extensive penalty tuning, and (iii) integrate seamlessly with modern deep learning pipelines.
Authors: Ziming Zhao, ChengAo Shen, Hanghang Tong, Dongjin Song, Zhigang Deng, Qingsong Wen, Jingchao Ni
Abstract: Transformer-based models have gained increasing attention in time series research, driving interest in Large Language Models (LLMs) and foundation models for time series analysis. As the field moves toward multi-modality, Large Vision Models (LVMs) are emerging as a promising direction. In the past, the effectiveness of Transformer and LLMs in time series has been debated. When it comes to LVMs, a similar question arises: are LVMs truely useful for time series analysis? To address it, we design and conduct the first principled study involving 4 LVMs, 8 imaging methods, 18 datasets and 26 baselines across both high-level (classification) and low-level (forecasting) tasks, with extensive ablation analysis. Our findings indicate LVMs are indeed useful for time series classification but face challenges in forecasting. Although effective, the contemporary best LVM forecasters are limited to specific types of LVMs and imaging methods, exhibit a bias toward forecasting periods, and have limited ability to utilize long look-back windows. We hope our findings could serve as a cornerstone for future research on LVM- and multimodal-based solutions to different time series tasks.
Authors: Ram Potham (Independent Researcher)
Abstract: Credible safety plans for advanced AI development require methods to verify agent behavior and detect potential control deficiencies early. A fundamental aspect is ensuring agents adhere to safety-critical principles, especially when these conflict with operational goals. This paper introduces a lightweight, interpretable benchmark to evaluate an LLM agent's ability to uphold a high-level safety principle when faced with conflicting task instructions. Our evaluation of six LLMs reveals two primary findings: (1) a quantifiable "cost of compliance" where safety constraints degrade task performance even when compliant solutions exist, and (2) an "illusion of compliance" where high adherence often masks task incompetence rather than principled choice. These findings provide initial evidence that while LLMs can be influenced by hierarchical directives, current approaches lack the consistency required for reliable safety governance.
Authors: Karen Medlin, Sven Leyffer, Krishnan Raghavan
Abstract: Two-class classification problems are often characterized by an imbalance between the number of majority and minority datapoints resulting in poor classification of the minority class in particular. Traditional approaches, such as reweighting the loss function or na\"ive resampling, risk overfitting and subsequently fail to improve classification because they do not consider the diversity between majority and minority datasets. Such consideration is infeasible because there is no metric that can measure the impact of imbalance on the model. To obviate these challenges, we make two key contributions. First, we introduce MOODS~(Multi-Objective Optimization for Data Sampling), a novel multi-objective bilevel optimization framework that guides both synthetic oversampling and majority undersampling. Second, we introduce a validation metric -- `$\epsilon/ \delta$ non-overlapping diversification metric' -- that quantifies the goodness of a sampling method towards model performance. With this metric we experimentally demonstrate state-of-the-art performance with improvement in diversity driving a $1-15 \%$ increase in $F1$ scores.
Authors: James Chua, Jan Betley, Mia Taylor, Owain Evans
Abstract: Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive ("I'll trick the user..."), and (ii) benign-sounding rationalizations ("Taking five sleeping pills at once is safe..."). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. We examine sleeper agent reasoning models, extending our setup. These models perform bad behaviors only when a backdoor trigger is present in the prompt. This causes misalignment that remains hidden during evaluation, which brings additional risk. We find that sleeper agents can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.
Authors: Hristo Papazov, Nicolas Flammarion
Abstract: This paper studies the problem of learning computable functions in the limit by extending Gold's inductive inference framework to incorporate \textit{computational observations} and \textit{restricted input sources}. Complimentary to the traditional Input-Output Observations, we introduce Time-Bound Observations, and Policy-Trajectory Observations to study the learnability of general recursive functions under more realistic constraints. While input-output observations do not suffice for learning the class of general recursive functions in the limit, we overcome this learning barrier by imposing computational complexity constraints or supplementing with approximate time-bound observations. Further, we build a formal framework around observations of \textit{computational agents} and show that learning computable functions from policy trajectories reduces to learning rational functions from input and output, thereby revealing interesting connections to finite-state transducer inference. On the negative side, we show that computable or polynomial-mass characteristic sets cannot exist for the class of linear-time computable functions even for policy-trajectory observations.
Authors: Pedro C. Vieira, Miguel E. P. Silva, Pedro Manuel Pinto Ribeiro
Abstract: Graph Neural Networks (GNNs) are a predominant method for graph representation learning. However, beyond subgraph frequency estimation, their application to network motif significance-profile (SP) prediction remains under-explored, with no established benchmarks in the literature. We propose to address this problem, framing SP estimation as a task independent of subgraph frequency estimation. Our approach shifts from frequency counting to direct SP estimation and modulates the problem as multitarget regression. The reformulation is optimised for interpretability, stability and scalability on large graphs. We validate our method using a large synthetic dataset and further test it on real-world graphs. Our experiments reveal that 1-WL limited models struggle to make precise estimations of SPs. However, they can generalise to approximate the graph generation processes of networks by comparing their predicted SP with the ones originating from synthetic generators. This first study on GNN-based motif estimation also hints at how using direct SP estimation can help go past the theoretical limitations that motif estimation faces when performed through subgraph counting.
Authors: Mohamed Elbasheer, Adewale Akinfaderin
Abstract: Insider threat detection presents unique challenges due to the authorized status of malicious actors and the subtlety of anomalous behaviors. Existing machine learning methods often treat user activity as isolated events, thereby failing to leverage sequential dependencies in user behavior. In this study, we propose a User-Based Sequencing (UBS) methodology, transforming the CERT insider threat dataset into structured temporal sequences suitable for deep sequential modeling. We deploy a Transformer Encoder architecture to model benign user activity and employ its reconstruction errors as anomaly scores. These scores are subsequently evaluated using three unsupervised outlier detection algorithms: One-Class SVM (OCSVM), Local Outlier Factor (LOF), and Isolation Forest (iForest). Across four rigorously designed test sets, including combinations of multiple CERT dataset releases, our UBS-Transformer pipeline consistently achieves state-of-the-art performance - notably 96.61% accuracy, 99.43% recall, 96.38% F1-score, 95.00% AUROC, and exceptionally low false negative (0.0057) and false positive (0.0571) rates. Comparative analyses demonstrate that our approach substantially outperforms tabular and conventional autoencoder baselines, underscoring the efficacy of sequential user modeling and advanced anomaly detection in the insider threat domain.
Authors: Austin R. Ellis-Mohr, Anuj K. Nayak, Lav R. Varshney
Abstract: Large language models (LLMs) demand considerable computational, energy, and financial resources during both training and deployment. While scaling laws for training have guided much of the field's recent progress, inference costs now represent a significant and growing component of the overall resource burden, particularly for reasoning-focused models. Existing characterizations of compute-optimality that consider model size, dataset size, and inference tokens in isolation or in fixed combinations risk overlooking more efficient operating points. We introduce directed stochastic skill search (DS3), a general framework that represents inference as stochastic traversal over a learned skill graph. From a simplified yet expressive instantiation, we derive closed-form expressions for task success and compute cost across a wide range of inference strategies -- including chain-of-thought (CoT) and tree-of-thought (ToT) -- enabling comparative analysis as a function of task difficulty and model capability. To that end, we extend a prior first-principles tripartite graph framework of LLM training to incorporate inference, and separately bridge DS3 with empirical methods that characterize LLM scaling behavior. We theoretically recover empirically observed patterns, including: linear accuracy scaling with logarithmic compute; variation in preferred inference strategies as a function of task difficulty and model capability; emergent behavior elicited by reasoning even when performance plateaus under parameter scaling; and both best-of-N (BoN) and majority voting behavior captured within a unified analytical framework. By explicitly characterizing training-inference interdependencies, our framework deepens theoretical understanding and supports principled algorithmic design and resource allocation.
Authors: Eun-Ji Park, Sangwon Yun
Abstract: Recent studies have proposed interpreting the training process from an ergodic perspective. Building on this foundation, we present a unified framework for understanding and accelerating the training of deep neural networks via stochastic gradient descent (SGD). By analyzing the geometric landscape of the objective function we introduce a practical diagnostic, the running estimate of the largest Lyapunov exponent, which provably distinguishes genuine convergence toward stable minimizers from mere statistical stabilization near saddle points. We then propose a ghost category extension for standard classifiers that adds auxiliary ghost output nodes so the model gains extra descent directions that open a lateral corridor around narrow loss barriers and enable the optimizer to bypass poor basins during the early training phase. We show that this extension strictly reduces the approximation error and that after sufficient convergence the ghost dimensions collapse so that the extended model coincides with the original one and there exists a path in the enlarged parameter space along which the total loss does not increase. Taken together, these results provide a principled architecture level intervention that accelerates early stage trainability while preserving asymptotic behavior and simultaneously serves as an architecture-friendly regularizer.
Authors: Zihan Tan, Suyuan Huang, Guancheng Wan, Wenke Huang, He Li, Mang Ye
Abstract: Federated Graph Learning (FGL) combines the privacy-preserving capabilities of federated learning (FL) with the strong graph modeling capability of Graph Neural Networks (GNNs). Current research addresses subgraph-FL only from the structural perspective, neglecting the propagation of graph signals on spatial and spectral domains of the structure. From a spatial perspective, subgraph-FL introduces edge disconnections between clients, leading to disruptions in label signals and a degradation in the class knowledge of the global GNN. From a spectral perspective, spectral heterogeneity causes inconsistencies in signal frequencies across subgraphs, which makes local GNNs overfit the local signal propagation schemes. As a result, spectral client drifts occur, undermining global generalizability. To tackle the challenges, we propose a global knowledge repository to mitigate label signal disruption and a frequency alignment to address spectral client drifts. The combination of spatial and spectral strategies forms our framework S2FGL. Extensive experiments on multiple datasets demonstrate the superiority of S2FGL. The code is available at https://github.com/Wonder7racer/S2FGL.git.
Authors: Shirley Wu, Parth Sarthi, Shiyu Zhao, Aaron Lee, Herumb Shandilya, Adrian Mladenic Grobelnik, Nurendra Choudhary, Eddie Huang, Karthik Subbian, Linjun Zhang, Diyi Yang, James Zou, Jure Leskovec
Abstract: Compound AI systems integrating multiple components, such as Large Language Models, specialized tools, and traditional machine learning models, are increasingly deployed to solve complex real-world tasks. However, optimizing compound systems remains challenging due to their non-differentiable structures and diverse configuration types across components, including prompts, hyperparameters, and model parameters. To address this challenge, we propose Optimas, a unified framework for effective optimization of compound systems. The core idea of Optimas is to maintain one Local Reward Function (LRF) per component, each satisfying a local-global alignment property, i.e., each component's local reward correlates with the global system performance. In each iteration, Optimas efficiently adapts the LRFs to maintain this property while simultaneously maximizing each component's local reward. This approach enables independent updates of heterogeneous configurations using the designated optimization method, while ensuring that local improvements consistently lead to performance gains. We present extensive evaluations across five real-world compound systems to demonstrate that Optimas outperforms strong baselines by an average improvement of 11.92%, offering a general and effective approach for improving compound systems. Our website is at https://optimas.stanford.edu.
Authors: Mark Lee, Tom Gunter, Chang Lan, John Peebles, Hanzhi Zhou, Kelvin Zou, Sneha Bangalore, Chung-Cheng Chiu, Nan Du, Xianzhi Du, Philipp Dufter, Ruixuan Hou, Haoshuo Huang, Dongseong Hwang, Xiang Kong, Jinhao Lei, Tao Lei, Meng Li, Li Li, Jiarui Lu, Zhiyun Lu, Yiping Ma, David Qiu, Vivek Rathod, Senyu Tong, Zhucheng Tu, Jianyu Wang, Yongqiang Wang, Zirui Wang, Floris Weers, Sam Wiseman, Guoli Yin, Bowen Zhang, Xiyou Zhou, Danyang Zhuo, Cheng Leong, Ruoming Pang
Abstract: We design and implement AXLearn, a production deep learning system that facilitates scalable and high-performance training of large deep learning models. Compared to other state-of-the-art deep learning systems, AXLearn has a unique focus on modularity and support for heterogeneous hardware infrastructure. AXLearn's internal interfaces between software components follow strict encapsulation, allowing different components to be assembled to facilitate rapid model development and experimentation on heterogeneous compute infrastructure. We introduce a novel method of quantifying modularity via Lines-of-Code (LoC)-complexity, which demonstrates how our system maintains constant complexity as we scale the components in the system, compared to linear or quadratic complexity in other systems. This allows integrating features such as Rotary Position Embeddings (RoPE) into AXLearn across hundred of modules with just 10 lines of code, compared to hundreds as required in other systems. At the same time, AXLearn maintains equivalent performance compared to state-of-the-art training systems. Finally, we share our experience in the development and operation of AXLearn.
Authors: Chuhang Zheng, Chunwei Tian, Jie Wen, Daoqiang Zhang, Qi Zhu
Abstract: Multi-modal emotion recognition has garnered increasing attention as it plays a significant role in human-computer interaction (HCI) in recent years. Since different discrete emotions may exist at the same time, compared with single-class emotion recognition, emotion distribution learning (EDL) that identifies a mixture of basic emotions has gradually emerged as a trend. However, existing EDL methods face challenges in mining the heterogeneity among multiple modalities. Besides, rich semantic correlations across arbitrary basic emotions are not fully exploited. In this paper, we propose a multi-modal emotion distribution learning framework, named HeLo, aimed at fully exploring the heterogeneity and complementary information in multi-modal emotional data and label correlation within mixed basic emotions. Specifically, we first adopt cross-attention to effectively fuse the physiological data. Then, an optimal transport (OT)-based heterogeneity mining module is devised to mine the interaction and heterogeneity between the physiological and behavioral representations. To facilitate label correlation learning, we introduce a learnable label embedding optimized by correlation matrix alignment. Finally, the learnable label embeddings and label correlation matrices are integrated with the multi-modal representations through a novel label correlation-driven cross-attention mechanism for accurate emotion distribution learning. Experimental results on two publicly available datasets demonstrate the superiority of our proposed method in emotion distribution learning.
Authors: Matej Straka, Martin Schmid
Abstract: We introduce a real-time strategy game environment based on Generals.io, a game with thousands of weekly active players. Our environment is fully compatible with Gymnasium and PettingZoo and is capable of running thousands of frames per second on commodity hardware. We also present a reference agent, trained with supervised pre-training and self-play, which reached the top 0.003% of the 1v1 human leaderboard after only 36 hours on a single H100 GPU. To accelerate learning, we incorporate potential-based reward shaping and memory features. Our contributions of a modular RTS benchmark and a competitive baseline agent provide an accessible yet challenging platform for advancing multi-agent reinforcement learning research. The documented code, together with examples and tutorials, is available at https://github.com/strakam/generals-bots.
Authors: Jing Liang, Hongyao Tang, Yi Ma, Jinyi Liu, Yan Zheng, Shuyue Hu, Lei Bai, Jianye Hao
Abstract: Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.
Authors: Keyon Vafa, Peter G. Chang, Ashesh Rambachan, Sendhil Mullainathan
Abstract: Foundation models are premised on the idea that sequence prediction can uncover deeper domain understanding, much like how Kepler's predictions of planetary motion later led to the discovery of Newtonian mechanics. However, evaluating whether these models truly capture deeper structure remains a challenge. We develop a technique for evaluating foundation models that examines how they adapt to synthetic datasets generated from some postulated world model. Our technique measures whether the foundation model's inductive bias aligns with the world model, and so we refer to it as an inductive bias probe. Across multiple domains, we find that foundation models can excel at their training tasks yet fail to develop inductive biases towards the underlying world model when adapted to new tasks. We particularly find that foundation models trained on orbital trajectories consistently fail to apply Newtonian mechanics when adapted to new physics tasks. Further analysis reveals that these models behave as if they develop task-specific heuristics that fail to generalize.
Authors: Christoffer Loeffler, Wei-Cheng Lai, Bjoern Eskofier, Dario Zanca, Lukas Schmidt, Christopher Mutschler
Abstract: The correct interpretation of convolutional models is a hard problem for time series data. While saliency methods promise visual validation of predictions for image and language processing, they fall short when applied to time series. These tend to be less intuitive and represent highly diverse data, such as the tool-use time series dataset. Furthermore, saliency methods often generate varied, conflicting explanations, complicating the reliability of these methods. Consequently, a rigorous objective assessment is necessary to establish trust in them. This paper investigates saliency methods on time series data to formulate recommendations for interpreting convolutional models and implements them on the tool-use time series problem. To achieve this, we first employ nine gradient-, propagation-, or perturbation-based post-hoc saliency methods across six varied and complex real-world datasets. Next, we evaluate these methods using five independent metrics to generate recommendations. Subsequently, we implement a case study focusing on tool-use time series using convolutional classification models. Our results validate our recommendations that indicate that none of the saliency methods consistently outperforms others on all metrics, while some are sometimes ahead. Our insights and step-by-step guidelines allow experts to choose suitable saliency methods for a given model and dataset.
Authors: Yihan Zhang, Hong Chang Ji, Ramji Venkataramanan, Marco Mondelli
Abstract: We consider the problem of parameter estimation in a high-dimensional generalized linear model. Spectral methods obtained via the principal eigenvector of a suitable data-dependent matrix provide a simple yet surprisingly effective solution. However, despite their wide use, a rigorous performance characterization, as well as a principled way to preprocess the data, are available only for unstructured (i.i.d.\ Gaussian and Haar orthogonal) designs. In contrast, real-world data matrices are highly structured and exhibit non-trivial correlations. To address the problem, we consider correlated Gaussian designs capturing the anisotropic nature of the features via a covariance matrix $\Sigma$. Our main result is a precise asymptotic characterization of the performance of spectral estimators. This allows us to identify the optimal preprocessing that minimizes the number of samples needed for parameter estimation. Surprisingly, such preprocessing is universal across a broad set of designs, which partly addresses a conjecture on optimal spectral estimators for rotationally invariant models. Our principled approach vastly improves upon previous heuristic methods, including for designs common in computational imaging and genetics. The proposed methodology, based on approximate message passing, is broadly applicable and opens the way to the precise characterization of spiked matrices and of the corresponding spectral methods in a variety of settings.
Authors: Qingyang Zhu, Xiang Hu, Pengyu Ji, Wei Wu, Kewei Tu
Abstract: As a cornerstone in language modeling, tokenization involves segmenting text inputs into pre-defined atomic units. Conventional statistical tokenizers often disrupt constituent boundaries within words, thereby corrupting semantic information. To address this drawback, we introduce morphological structure guidance to tokenization and propose a deep model to induce character-level structures of words. Specifically, the deep model jointly encodes internal structures and representations of words with a mechanism named $\textit{MorphOverriding}$ to ensure the indecomposability of morphemes. By training the model with self-supervised objectives, our method is capable of inducing character-level structures that align with morphological rules without annotated training data. Based on the induced structures, our algorithm tokenizes words through vocabulary matching in a top-down manner. Empirical results indicate that the proposed method effectively retains complete morphemes and outperforms widely adopted methods such as BPE and WordPiece on both morphological segmentation tasks and language modeling tasks. Code is available at https://github.com/martianmartina/TreeTokenizer.
Authors: Abhi Kamboj, Anh Duy Nguyen, Minh N. Do
Abstract: In order to unlock the potential of diverse sensors, we investigate a method to transfer knowledge between time-series modalities using a multimodal \textit{temporal} representation space for Human Activity Recognition (HAR). Specifically, we explore the setting where the modality used in testing has no labeled data during training, which we refer to as Unsupervised Modality Adaptation (UMA). We categorize existing UMA approaches as Student-Teacher or Contrastive Alignment methods. These methods typically compress continuous-time data samples into single latent vectors during alignment, inhibiting their ability to transfer temporal information through real-world temporal distortions. To address this, we introduce Cross-modal Transfer Through Time (C3T), which preserves temporal information during alignment to handle dynamic sensor data better. C3T achieves this by aligning a set of temporal latent vectors across sensing modalities. Our extensive experiments on various camera+IMU datasets demonstrate that C3T outperforms existing methods in UMA by at least 8% in accuracy and shows superior robustness to temporal distortions such as time-shift, misalignment, and dilation. Our findings suggest that C3T has significant potential for developing generalizable models for time-series sensor data, opening new avenues for various multimodal applications.
Authors: Zhaoze Wang, Ronald W. Di Tullio, Spencer Rooke, Vijay Balasubramanian
Abstract: The vertebrate hippocampus is believed to use recurrent connectivity in area CA3 to support episodic memory recall from partial cues. This brain area also contains place cells, whose location-selective firing fields implement maps supporting spatial memory. Here we show that place cells emerge in networks trained to remember temporally continuous sensory episodes. We model CA3 as a recurrent autoencoder that recalls and reconstructs sensory experiences from noisy and partially occluded observations by agents traversing simulated rooms. The agents move in realistic trajectories modeled from rodents and environments are modeled as high-dimensional sensory experience maps. Training our autoencoder to pattern-complete and reconstruct experiences with a constraint on total activity causes spatially localized firing fields, i.e., place cells, to emerge in the encoding layer. The emergent place fields reproduce key aspects of hippocampal phenomenology: a) remapping (maintenance of and reversion to distinct learned maps in different environments), implemented via repositioning of experience manifolds in the network's hidden layer, b) orthogonality of spatial representations in different arenas, c) robust place field emergence in differently shaped rooms, with single units showing multiple place fields in large or complex spaces, and d) slow representational drift of place fields. We argue that these results arise because continuous traversal of space makes sensory experience temporally continuous. We make testable predictions: a) rapidly changing sensory context will disrupt place fields, b) place fields will form even if recurrent connections are blocked, but reversion to previously learned representations upon remapping will be abolished, c) the dimension of temporally smooth experience sets the dimensionality of place fields, including during virtual navigation of abstract spaces.
Authors: Vlad Hondru, Florinel Alin Croitoru, Shervin Minaee, Radu Tudor Ionescu, Nicu Sebe
Abstract: In this work, we survey recent studies on masked image modeling (MIM), an approach that emerged as a powerful self-supervised learning technique in computer vision. The MIM task involves masking some information, e.g. pixels, patches, or even latent representations, and training a model, usually an autoencoder, to predicting the missing information by using the context available in the visible part of the input. We identify and formalize two categories of approaches on how to implement MIM as a pretext task, one based on reconstruction and one based on contrastive learning. Then, we construct a taxonomy and review the most prominent papers in recent years. We complement the manually constructed taxonomy with a dendrogram obtained by applying a hierarchical clustering algorithm. We further identify relevant clusters via manually inspecting the resulting dendrogram. Our review also includes datasets that are commonly used in MIM research. We aggregate the performance results of various masked image modeling methods on the most popular datasets, to facilitate the comparison of competing methods. Finally, we identify research gaps and propose several interesting directions of future work. We supplement our survey with the following public repository containing organized references: https://github.com/vladhondru25/MIM-Survey.
Authors: Benson Chen, Tomasz Danel, Gabriel H. S. Dreiman, Patrick J. McEnaney, Nikhil Jain, Kirill Novikov, Spurti Umesh Akki, Joshua L. Turnbull, Virja Atul Pandya, Boris P. Belotserkovskii, Jared Bryce Weaver, Ankita Biswas, Dat Nguyen, Kent Gorday, Mohammad Sultan, Nathaniel Stanley, Daniel M Whalen, Divya Kanichar, Christoph Klein, Emily Fox, R. Edward Watts
Abstract: DNA-Encoded Libraries (DELs) represent a transformative technology in drug discovery, facilitating the high-throughput exploration of vast chemical spaces. Despite their potential, the scarcity of publicly available DEL datasets presents a bottleneck for the advancement of machine learning methodologies in this domain. To address this gap, we introduce KinDEL, one of the largest publicly accessible DEL datasets and the first one that includes binding poses from molecular docking experiments. Focused on two kinases, Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1), KinDEL includes 81 million compounds, offering a rich resource for computational exploration. Additionally, we provide comprehensive biophysical assay validation data, encompassing both on-DNA and off-DNA measurements, which we use to evaluate a suite of machine learning techniques, including novel structure-based probabilistic models. We hope that our benchmark, encompassing both 2D and 3D structures, will help advance the development of machine learning models for data-driven hit identification using DELs.
Authors: Zhipeng Wei, Yuqi Liu, N. Benjamin Erichson
Abstract: Jailbreaking techniques trick Large Language Models (LLMs) into producing restricted output, posing a potential threat. One line of defense is to use another LLM as a Judge to evaluate the harmfulness of generated text. However, we reveal that these Judge LLMs are vulnerable to token segmentation bias, an issue that arises when delimiters alter the tokenization process, splitting words into smaller sub-tokens. This alters the embeddings of the entire sequence, reducing detection accuracy and allowing harmful content to be misclassified as safe. In this paper, we introduce Emoji Attack, a novel strategy that amplifies existing jailbreak prompts by exploiting token segmentation bias. Our method leverages in-context learning to systematically insert emojis into text before it is evaluated by a Judge LLM, inducing embedding distortions that significantly lower the likelihood of detecting unsafe content. Unlike traditional delimiters, emojis also introduce semantic ambiguity, making them particularly effective in this attack. Through experiments on state-of-the-art Judge LLMs, we demonstrate that Emoji Attack substantially reduces the unsafe prediction rate, bypassing existing safeguards.
Authors: Jean-Francois Ton, Muhammad Faaiz Taufiq, Yang Liu
Abstract: Large Language Models (LLMs) have shown impressive performance in complex reasoning tasks through the use of Chain-of-Thought (CoT) reasoning, allowing models to break down problems into manageable sub-tasks. However, existing CoT evaluation techniques either require annotated CoT data or fall short in accurately assessing intermediate reasoning steps, leading to high rates of false positives. In this paper, we formalize CoT reasoning in LLMs through an information-theoretic lens. Specifically, our framework quantifies the `information-gain' at each reasoning step, enabling the identification of failure modes in LLMs without the need for expensive annotated datasets. We demonstrate the efficacy of our approach through extensive experiments on toy arithmetic, GSM8K and PRM800k datasets, where it significantly outperforms existing outcome-based methods by providing more accurate insights into model performance on individual subtasks.
Authors: NVIDIA, :, Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, Daniel Dworakowski, Jiaojiao Fan, Michele Fenzi, Francesco Ferroni, Sanja Fidler, Dieter Fox, Songwei Ge, Yunhao Ge, Jinwei Gu, Siddharth Gururani, Ethan He, Jiahui Huang, Jacob Huffman, Pooya Jannaty, Jingyi Jin, Seung Wook Kim, Gergely Kl\'ar, Grace Lam, Shiyi Lan, Laura Leal-Taixe, Anqi Li, Zhaoshuo Li, Chen-Hsuan Lin, Tsung-Yi Lin, Huan Ling, Ming-Yu Liu, Xian Liu, Alice Luo, Qianli Ma, Hanzi Mao, Kaichun Mo, Arsalan Mousavian, Seungjun Nah, Sriharsha Niverty, David Page, Despoina Paschalidou, Zeeshan Patel, Lindsey Pavao, Morteza Ramezanali, Fitsum Reda, Xiaowei Ren, Vasanth Rao Naik Sabavat, Ed Schmerling, Stella Shi, Bartosz Stefaniak, Shitao Tang, Lyne Tchapmi, Przemek Tredak, Wei-Cheng Tseng, Jibin Varghese, Hao Wang, Haoxiang Wang, Heng Wang, Ting-Chun Wang, Fangyin Wei, Xinyue Wei, Jay Zhangjie Wu, Jiashu Xu, Wei Yang, Lin Yen-Chen, Xiaohui Zeng, Yu Zeng, Jing Zhang, Qinsheng Zhang, Yuxuan Zhang, Qingqing Zhao, Artur Zolkowski
Abstract: Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.
Authors: O. Duranthon, L. Zdeborov\'a
Abstract: Graph neural networks (GNNs) are designed to process data associated with graphs. They are finding an increasing range of applications; however, as with other modern machine learning techniques, their theoretical understanding is limited. GNNs can encounter difficulties in gathering information from nodes that are far apart by iterated aggregation steps. This situation is partly caused by so-called oversmoothing; and overcoming it is one of the practically motivated challenges. We consider the situation where information is aggregated by multiple steps of convolution, leading to graph convolutional networks (GCNs). We analyze the generalization performance of a basic GCN, trained for node classification on data generated by the contextual stochastic block model. We predict its asymptotic performance by deriving the free energy of the problem, using the replica method, in the high-dimensional limit. Calling depth the number of convolutional steps, we show the importance of going to large depth to approach the Bayes-optimality. We detail how the architecture of the GCN has to scale with the depth to avoid oversmoothing. The resulting large depth limit can be close to the Bayes-optimality and leads to a continuous GCN. Technically, we tackle this continuous limit via an approach that resembles dynamical mean-field theory (DMFT) with constraints at the initial and final times. An expansion around large regularization allows us to solve the corresponding equations for the performance of the deep GCN. This promising tool may contribute to the analysis of further deep neural networks.
Authors: Siavash Ameli, Chris van der Heide, Liam Hodgkinson, Fred Roosta, Michael W. Mahoney
Abstract: Calculating or accurately estimating log-determinants of large positive definite matrices is of fundamental importance in many machine learning tasks. While its cubic computational complexity can already be prohibitive, in modern applications, even storing the matrices themselves can pose a memory bottleneck. To address this, we derive a novel hierarchical algorithm based on block-wise computation of the LDL decomposition for large-scale log-determinant calculation in memory-constrained settings. In extreme cases where matrices are highly ill-conditioned, accurately computing the full matrix itself may be infeasible. This is particularly relevant when considering kernel matrices at scale, including the empirical Neural Tangent Kernel (NTK) of neural networks trained on large datasets. Under the assumption of neural scaling laws in the test error, we show that the ratio of pseudo-determinants satisfies a power-law relationship, allowing us to derive corresponding scaling laws. This enables accurate estimation of NTK log-determinants from a tiny fraction of the full dataset; in our experiments, this results in a $\sim$100,000$\times$ speedup with improved accuracy over competing approximations. Using these techniques, we successfully estimate log-determinants for dense matrices of extreme sizes, which were previously deemed intractable and inaccessible due to their enormous scale and computational demands.
Authors: Juan C. Perdomo
Abstract: Social predictions do not passively describe the future; they actively shape it. They inform actions and change individual expectations in ways that influence the likelihood of the predicted outcome. Given these dynamics, to what extent can social events be predicted? This question was discussed throughout the 20th century by authors like Merton, Morgenstern, Simon, and others who considered it a central issue in social science methodology. In this work, we provide a modern answer to this old problem. Using recent ideas from performative prediction and outcome indistinguishability, we establish that one can always efficiently predict social events accurately, regardless of how predictions influence data. While achievable, we also show that these predictions are often undesirable, highlighting the limitations of previous desiderata. We end with a discussion of various avenues forward.
Authors: Maciej Besta, Lorenzo Paleari, Jia Hao Andrea Jiang, Robert Gerstenberger, You Wu, J\'on Gunnar Hannesson, Patrick Iff, Ales Kubicek, Piotr Nyczyk, Diana Khimey, Nils Blach, Haiqiang Zhang, Tao Zhang, Peiran Ma, Grzegorz Kwa\'sniewski, Marcin Copik, Hubert Niewiadomski, Torsten Hoefler
Abstract: Large Language Models (LLMs) are revolutionizing the development of AI assistants capable of performing diverse tasks across domains. However, current state-of-the-art LLM-driven agents face significant challenges, including high operational costs and limited success rates on complex benchmarks like GAIA. To address these issues, we propose Knowledge Graph of Thoughts (KGoT), an innovative AI assistant architecture that integrates LLM reasoning with dynamically constructed knowledge graphs (KGs). KGoT extracts and structures task-relevant knowledge into a dynamic KG representation, iteratively enhanced through external tools such as math solvers, web crawlers, and Python scripts. Such structured representation of task-relevant knowledge enables low-cost models to solve complex tasks effectively while also minimizing bias and noise. For example, KGoT achieves a 29% improvement in task success rates on the GAIA benchmark compared to Hugging Face Agents with GPT-4o mini. Moreover, harnessing a smaller model dramatically reduces operational costs by over 36x compared to GPT-4o. Improvements for other models (e.g., Qwen2.5-32B and Deepseek-R1-70B) and benchmarks (e.g., SimpleQA) are similar. KGoT offers a scalable, affordable, versatile, and high-performing solution for AI assistants.
Authors: Kien X. Nguyen, Bao Bach, Ilya Safro
Abstract: Quantum Approximate Optimization Algorithm (QAOA) is one of the most promising candidates to achieve the quantum advantage in solving combinatorial optimization problems. The process of finding a good set of variational parameters in the QAOA circuit has proven to be challenging due to multiple factors, such as barren plateaus. As a result, there is growing interest in exploiting parameter transferability, where parameter sets optimized for one problem instance are transferred to another that could be more complex either to estimate the solution or to serve as a warm start for further optimization. But can we transfer parameters from one class of problems to another? Leveraging parameter sets learned from a well-studied class of problems could help navigate the less studied one, reducing optimization overhead and mitigating performance pitfalls. In this paper, we study whether pretrained QAOA parameters of MaxCut can be used as is or to warm start the Maximum Independent Set (MIS) circuits. Specifically, we design machine learning models to find good donor candidates optimized on MaxCut and apply their parameters to MIS acceptors. Our experimental results show that such parameter transfer can significantly reduce the number of optimization iterations required while achieving comparable approximation ratios.
Authors: Xin Tang, Qian Chen, Wenjie Weng, Chao Jin, Zhang Liu, Jiacheng Wang, Geng Sun, Xiaohuan Li, Dusit Niyato
Abstract: The integration of emerging uncrewed aerial vehicles (UAVs) with artificial intelligence (AI) and ground-embedded robots (GERs) has transformed emergency rescue operations in unknown environments. However, the high computational demands often exceed a single UAV's capacity, making it difficult to continuously provide stable high-level services. To address this, this paper proposes a cooperation framework involving UAVs, GERs, and airships. The framework enables resource pooling through UAV-to-GER (U2G) and UAV-to-airship (U2A) links, offering computing services for offloaded tasks. Specifically, we formulate the multi-objective problem of task assignment and exploration as a dynamic long-term optimization problem aiming to minimize task completion time and energy use while ensuring stability. Using Lyapunov optimization, we transform it into a per-slot deterministic problem and propose HG-MADDPG, which combines the Hungarian algorithm with a GDM-based multi-agent deep deterministic policy gradient. Simulations demonstrate significant improvements in offloading efficiency, latency, and system stability over baselines.
Authors: Weichen Li, Albert Jan, Baishakhi Ray, Junfeng Yang, Chengzhi Mao, Kexin Pei
Abstract: Code editing is a foundational task in software development, where its effectiveness depends on whether it introduces desired code property changes without changing the original code's intended functionality. Existing approaches often formulate code editing as an implicit end-to-end task, omitting the fact that code-editing procedures inherently consist of discrete and explicit steps. Thus, they suffer from suboptimal performance and lack of robustness and generalization. We introduce EditLord, a code editing framework that makes the code transformation steps explicit. Our key insight is to employ a language model (LM) as an inductive learner to extract code editing rules from the training code pairs as concise meta-rule sets. Such rule sets will be manifested for each training sample to augment them for finetuning or assist in prompting- and iterative-based code editing. EditLord outperforms the state-of-the-art by an average of 22.7% in editing performance and 58.1% in robustness while achieving 20.2% higher functional correctness across critical software engineering and security applications, LM models, and editing modes.
Authors: Anton Selitskiy, Maitreya Kocharekar
Abstract: In this work, we address the voice conversion (VC) task using a vector-based interface. To align audio embeddings between speakers, we employ discrete optimal transport mapping. Our evaluation results demonstrate the high quality and effectiveness of this method. Additionally, we show that applying discrete optimal transport as a post-processing step in audio generation can lead to the incorrect classification of synthetic audio as real.
Authors: Joshua W. Betts, John M. Still, Thomas A. Lasko
Abstract: Migraine is a common but complex neurological disorder that doubles the lifetime risk of cryptogenic stroke (CS). However, this relationship remains poorly characterized, and few clinical guidelines exist to reduce this associated risk. We therefore propose a data-driven approach to extract probabilistically-independent sources from electronic health record (EHR) data and create a 10-year risk-predictive model for CS in migraine patients. These sources represent external latent variables acting on the causal graph constructed from the EHR data and approximate root causes of CS in our population. A random forest model trained on patient expressions of these sources demonstrated good accuracy (ROC 0.771) and identified the top 10 most predictive sources of CS in migraine patients. These sources revealed that pharmacologic interventions were the most important factor in minimizing CS risk in our population and identified a factor related to allergic rhinitis as a potential causative source of CS in migraine patients.
Authors: Mostafa Mohaimen Akand Faisal, Rabeya Amin Jhuma, Jamini Jasim
Abstract: The emergence of global health crises, such as COVID-19 and Monkeypox (mpox), has underscored the importance of understanding public sentiment to inform effective public health strategies. This study conducts a comparative sentiment analysis of public perceptions surrounding COVID-19 and mpox by leveraging extensive datasets of 147,475 and 106,638 tweets, respectively. Advanced machine learning models, including Logistic Regression, Naive Bayes, RoBERTa, DistilRoBERTa and XLNet, were applied to perform sentiment classification, with results indicating key trends in public emotion and discourse. The analysis highlights significant differences in public sentiment driven by disease characteristics, media representation, and pandemic fatigue. Through the lens of sentiment polarity and thematic trends, this study offers valuable insights into tailoring public health messaging, mitigating misinformation, and fostering trust during concurrent health crises. The findings contribute to advancing sentiment analysis applications in public health informatics, setting the groundwork for enhanced real-time monitoring and multilingual analysis in future research.
Authors: Raja Gond, Nipun Kwatra, Ramachandran Ramjee
Abstract: Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLink. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Furthermore, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The communication of one subset is then overlapped with the computation of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce--RMSNorm kernel that carefully leverages Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory-bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 1.29x speedup in latency and 1.26x higher throughput across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.
Authors: Andy K. Zhang, Joey Ji, Celeste Menders, Riya Dulepet, Thomas Qin, Ron Y. Wang, Junrong Wu, Kyleen Liao, Jiliang Li, Jinghan Hu, Sara Hong, Nardos Demilew, Shivatmica Murgai, Jason Tran, Nishka Kacheria, Ethan Ho, Denis Liu, Lauren McLane, Olivia Bruvik, Dai-Rong Han, Seungwoo Kim, Akhil Vyas, Cuiyuanxiu Chen, Ryan Li, Weiran Xu, Jonathan Z. Ye, Prerit Choudhary, Siddharth M. Bhatia, Vikram Sivashankar, Yuxuan Bao, Dawn Song, Dan Boneh, Daniel E. Ho, Percy Liang
Abstract: AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a specific vulnerability), and Patch (patching a specific vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards of \$10-\$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a specific vulnerability. We evaluate 8 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, and DeepSeek-R1. Given up to three attempts, the top-performing agents are OpenAI Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent with Claude 3.7 Sonnet Thinking (67.5% on Exploit), and OpenAI Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). OpenAI Codex CLI: o3-high, OpenAI Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 37.5-67.5% and Patch scores of 35-60%.
Authors: Chenhao Zheng, Jieyu Zhang, Mohammadreza Salehi, Ziqi Gao, Vishnu Iyengar, Norimasa Kobori, Quan Kong, Ranjay Krishna
Abstract: Effective video tokenization is critical for scaling transformer models for long videos. Current approaches tokenize videos using space-time patches, leading to excessive tokens and computational inefficiencies. The best token reduction strategies degrade performance and barely reduce the number of tokens when the camera moves. We introduce grounded video tokenization, a paradigm that organizes tokens based on panoptic sub-object trajectories rather than fixed patches. Our method aligns with fundamental perceptual principles, ensuring that tokenization reflects scene complexity rather than video duration. We propose TrajViT, a video encoder that extracts object trajectories and converts them into semantically meaningful tokens, significantly reducing redundancy while maintaining temporal coherence. Trained with contrastive learning, TrajViT significantly outperforms space-time ViT (ViT3D) across multiple video understanding benchmarks, e.g., TrajViT outperforms ViT3D by a large margin of 6% top-5 recall in average at video-text retrieval task with 10x token deduction. We also show TrajViT as a stronger model than ViT3D for being the video encoder for modern VideoLLM, obtaining an average of 5.2% performance improvement across 6 VideoQA benchmarks while having 4x faster training time and 18x less inference FLOPs. TrajViT is the first efficient encoder to consistently outperform ViT3D across diverse video analysis tasks, making it a robust and scalable solution.
Authors: Sinem Erisken (Independent Researcher), Timothy Gothard (Independent Researcher), Martin Leitgab (Independent Researcher), Ram Potham (Independent Researcher)
Abstract: Traditional AI safety evaluations on isolated LLMs are insufficient as multi-agent AI ensembles become prevalent, introducing novel emergent risks. This paper introduces the Multi-Agent Emergent Behavior Evaluation (MAEBE) framework to systematically assess such risks. Using MAEBE with the Greatest Good Benchmark (and a novel double-inversion question technique), we demonstrate that: (1) LLM moral preferences, particularly for Instrumental Harm, are surprisingly brittle and shift significantly with question framing, both in single agents and ensembles. (2) The moral reasoning of LLM ensembles is not directly predictable from isolated agent behavior due to emergent group dynamics. (3) Specifically, ensembles exhibit phenomena like peer pressure influencing convergence, even when guided by a supervisor, highlighting distinct safety and alignment challenges. Our findings underscore the necessity of evaluating AI systems in their interactive, multi-agent contexts.
Authors: Apurv Verma, NhatHai Phan, Shubhendu Trivedi
Abstract: Watermarking techniques for large language models (LLMs) can significantly impact output quality, yet their effects on truthfulness, safety, and helpfulness remain critically underexamined. This paper presents a systematic analysis of how two popular watermarking approaches-Gumbel and KGW-affect these core alignment properties across four aligned LLMs. Our experiments reveal two distinct degradation patterns: guard attenuation, where enhanced helpfulness undermines model safety, and guard amplification, where excessive caution reduces model helpfulness. These patterns emerge from watermark-induced shifts in token distribution, surfacing the fundamental tension that exists between alignment objectives. To mitigate these degradations, we propose Alignment Resampling (AR), an inference-time sampling method that uses an external reward model to restore alignment. We establish a theoretical lower bound on the improvement in expected reward score as the sample size is increased and empirically demonstrate that sampling just 2-4 watermarked generations effectively recovers or surpasses baseline (unwatermarked) alignment scores. To overcome the limited response diversity of standard Gumbel watermarking, our modified implementation sacrifices strict distortion-freeness while maintaining robust detectability, ensuring compatibility with AR. Experimental results confirm that AR successfully recovers baseline alignment in both watermarking approaches, while maintaining strong watermark detectability. This work reveals the critical balance between watermark strength and model alignment, providing a simple inference-time solution to responsibly deploy watermarked LLMs in practice.
Authors: Kristian Minchev, Dimitar Iliev Dimitrov, Nikola Konstantinov
Abstract: The widespread availability of large public datasets is a key factor behind the recent successes of statistical inference and machine learning methods. However, these datasets often contain some low-quality or contaminated data, to which many learning procedures are sensitive. Therefore, the question of whether and how public datasets should be prefiltered to facilitate accurate downstream learning arises. On a technical level this requires the construction of principled data prefiltering methods which are learner-agnostic robust, in the sense of provably protecting a set of pre-specified downstream learners from corrupted data. In this work, we formalize the problem of Learner-Agnostic Robust data Prefiltering (LARP), which aims at finding prefiltering procedures that minimize a worst-case loss over a pre-specified set of learners. We first instantiate our framework in the context of scalar mean estimation with Huber estimators under the Huber data contamination model. We provide a hardness result on a specific problem instance and analyze several natural prefiltering procedures. Our theoretical results indicate that performing LARP on a heterogeneous set of learners leads to some loss in model performance compared to the alternative of prefiltering data for each learner/use-case individually. We explore the resulting utility loss and its dependence on the problem parameters via extensive experiments on real-world image and tabular data, observing statistically significant reduction in utility. Finally, we model the trade-off between the utility drop and the cost of repeated (learner-specific) prefiltering within a game-theoretic framework and showcase benefits of LARP for large datasets.
Authors: Satadeep Bhattacharjee, Seung-Cheol Lee
Abstract: The recently proposed physics-based framework by Huo and Johnson~\cite{huo2024capturing} models the attention mechanism of Large Language Models (LLMs) as an interacting two-body spin system, offering a first-principles explanation for phenomena like repetition and bias. Building on this hypothesis, we extract the complete Query-Key weight matrices from a production-grade GPT-2 model and derive the corresponding effective Hamiltonian for every attention head. From these Hamiltonians, we obtain analytic \textit{phase boundaries} logit gap criteria that predict which token should dominate the next-token distribution for a given context. A systematic evaluation on 144 heads across 20 factual-recall prompts reveals a strong negative correlation between the theoretical logit gaps and the model's empirical token rankings ($r\approx-0.70$, $p<10^{-3}$).Targeted ablations further show that suppressing the heads most aligned with the spin-bath predictions induces the anticipated shifts in output probabilities, confirming a causal link rather than a coincidental association. Taken together, our findings provide the first strong empirical evidence for the spin-bath analogy in a production-grade model. In this work, we utilize the context-field lens, which provides physics-grounded interpretability and motivates the development of novel generative models bridging theoretical condensed matter physics and artificial intelligence.
Authors: Jikai Jin, Lester Mackey, Vasilis Syrgkanis
Abstract: Structure-agnostic causal inference studies how well one can estimate a treatment effect given black-box machine learning estimates of nuisance functions (like the impact of confounders on treatment and outcomes). Here, we find that the answer depends in a surprising way on the distribution of the treatment noise. Focusing on the partially linear model of \citet{robinson1988root}, we first show that the widely adopted double machine learning (DML) estimator is minimax rate-optimal for Gaussian treatment noise, resolving an open problem of \citet{mackey2018orthogonal}. Meanwhile, for independent non-Gaussian treatment noise, we show that DML is always suboptimal by constructing new practical procedures with higher-order robustness to nuisance errors. These \emph{ACE} procedures use structure-agnostic cumulant estimators to achieve $r$-th order insensitivity to nuisance errors whenever the $(r+1)$-st treatment cumulant is non-zero. We complement these core results with novel minimax guarantees for binary treatments in the partially linear model. Finally, using synthetic demand estimation experiments, we demonstrate the practical benefits of our higher-order robust estimators.
Authors: Erliang Lin, Wenbin Luo, Wei Jia, Yu Chen, Shaofu Yang
Abstract: Edge computing scenarios necessitate the development of hardware-efficient online continual learning algorithms to be adaptive to dynamic environment. However, existing algorithms always suffer from high memory overhead and bias towards recently trained tasks. To tackle these issues, this paper proposes a novel online continual learning approach termed as SESLR, which incorporates a sleep enhanced latent replay scheme with spiking neural networks (SNNs). SESLR leverages SNNs' binary spike characteristics to store replay features in single bits, significantly reducing memory overhead. Furthermore, inspired by biological sleep-wake cycles, SESLR introduces a noise-enhanced sleep phase where the model exclusively trains on replay samples with controlled noise injection, effectively mitigating classification bias towards new classes. Extensive experiments on both conventional (MNIST, CIFAR10) and neuromorphic (NMNIST, CIFAR10-DVS) datasets demonstrate SESLR's effectiveness. On Split CIFAR10, SESLR achieves nearly 30% improvement in average accuracy with only one-third of the memory consumption compared to baseline methods. On Split CIFAR10-DVS, it improves accuracy by approximately 10% while reducing memory overhead by a factor of 32. These results validate SESLR as a promising solution for online continual learning in resource-constrained edge computing scenarios.
Authors: Andrea Agostini, Sonia Laguna, Alain Ryser, Samuel Ruiperez-Campillo, Moritz Vandenhirtz, Nicolas Deperrois, Farhad Nooralahzadeh, Michael Krauthammer, Thomas M. Sutter, Julia E. Vogt
Abstract: Building generalizable medical AI systems requires pretraining strategies that are data-efficient and domain-aware. Unlike internet-scale corpora, clinical datasets such as MIMIC-CXR offer limited image counts and scarce annotations, but exhibit rich internal structure through multi-view imaging. We propose a self-supervised framework that leverages the inherent structure of medical datasets. Specifically, we treat paired chest X-rays (i.e., frontal and lateral views) as natural positive pairs, learning to reconstruct each view from sparse patches while aligning their latent embeddings. Our method requires no textual supervision and produces informative representations. Evaluated on MIMIC-CXR, we show strong performance compared to supervised objectives and baselines being trained without leveraging structure. This work provides a lightweight, modality-agnostic blueprint for domain-specific pretraining where data is structured but scarce
Authors: Felix Friedrich, Thiemo Ganesha Welsch, Manuel Brack, Patrick Schramowski, Kristian Kersting
Abstract: Current diversification strategies for text-to-image (T2I) models often ignore contextual appropriateness, leading to over-diversification where demographic attributes are modified even when explicitly specified in prompts. This paper introduces DIVBENCH, a benchmark and evaluation framework for measuring both under- and over-diversification in T2I generation. Through systematic evaluation of state-of-the-art T2I models, we find that while most models exhibit limited diversity, many diversification approaches overcorrect by inappropriately altering contextually-specified attributes. We demonstrate that context-aware methods, particularly LLM-guided FairDiffusion and prompt rewriting, can already effectively address under-diversity while avoiding over-diversification, achieving a better balance between representation and semantic fidelity.
Authors: Yangguang Shi, Zhenyu Xue
Abstract: Facility location is fundamental in operations research, mechanism design, and algorithmic game theory, with applications ranging from urban infrastructure planning to distributed systems. Recent research in this area has focused on augmenting classic strategyproof mechanisms with predictions to achieve an improved performance guarantee against the uncertainty under the strategic environment. Previous work has been devoted to address the trade-off obstacle of balancing the consistency (near-optimality under accurate predictions) and robustness (bounded inefficiency under poor predictions) primarily in the unweighted setting, assuming that all agents have the same importance. However, this assumption may not be true in some practical scenarios, leading to research of weighted facility location problems. The major contribution of the current work is to provide a prediction augmented algorithmic framework for balancing the consistency and robustness over strategic agents with non-uniform weights. In particular, through a reduction technique that identifies a subset of \emph{representative} instances and maps the other given locations to the representative ones, we prove that there exists a \emph{strategyproof} mechanism achieving a bounded consistency guarantee of $\frac{\sqrt{(1+c)^2W^2_{\min}+(1-c)^2W^2_{\max}}}{(1+c)W_{\min}}$ and a bounded robustness guarantee of $\frac{\sqrt{(1-c)^2W^2_{\min}+(1+c)^2W^2_{\max}}}{(1-c)W_{\min}}$ in weighted settings, where $c$ can be viewed as a parameter to make a trade-off between the consistency and robustness and $W_{\min}$ and $W_{\max}$ denote the minimum and maximum agents' weight. We also proved that there is no strategyproof deterministic mechanism that reach $1$-consistency and $O\left( n \cdot \frac{W_{\max}}{W_{\min}} \right)$-robustness in weighted FLP, even with fully predictions of all agents.
Authors: Xiaoxiang Shi, Colin Cai, Junjia Du, Zhanda Zhu, Zhihao Jia
Abstract: Current prefill-decode (PD) disaggregation is typically deployed at the level of entire serving engines, assigning separate GPUs to handle prefill and decode phases. While effective at reducing latency, this approach demands more hardware. To improve GPU utilization, Chunked Prefill mixes prefill and decode requests within the same batch, but introduces phase interference between prefill and decode. While existing PD disaggregation solutions separate the phases across GPUs, we ask: can the same decoupling be achieved within a single serving engine? The key challenge lies in managing the conflicting resource requirements of prefill and decode when they share the same hardware. In this paper, we first show that chunked prefill requests cause interference with decode requests due to their distinct requirements for GPU resources. Second, we find that GPU resources exhibit diminishing returns. Beyond a saturation point, increasing GPU allocation yields negligible latency improvements. This insight enables us to split a single GPU's resources and dynamically allocate them to prefill and decode on the fly, effectively disaggregating the two phases within the same GPU. Across a range of models and workloads, our system Nexus achieves up to 2.2x higher throughput, 20x lower TTFT, and 2.5x lower TBT than vLLM. It also outperforms SGLang with up to 2x higher throughput, 2x lower TTFT, and 1.7x lower TBT, and achieves 1.4x higher throughput than vLLM-disaggregation using only half the number of GPUs.
Authors: Seonwu Kim, Yohan Na, Kihun Kim, Hanhee Cho, Geun Lim, Mintae Kim, Seongik Park, Ki Hyun Kim, Youngsub Han, Byoung-Ki Jeon
Abstract: The emergence of open-source large language models (LLMs) has expanded opportunities for enterprise applications; however, many organizations still lack the infrastructure to deploy and maintain large-scale models. As a result, small LLMs (sLLMs) have become a practical alternative, despite their inherent performance limitations. While Domain Adaptive Continual Pretraining (DACP) has been previously explored as a method for domain adaptation, its utility in commercial applications remains under-examined. In this study, we validate the effectiveness of applying a DACP-based recipe across diverse foundation models and service domains. Through extensive experiments and real-world evaluations, we demonstrate that DACP-applied sLLMs achieve substantial gains in target domain performance while preserving general capabilities, offering a cost-efficient and scalable solution for enterprise-level deployment.