Peer Revenue Returns subject to peer review

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

Computer Science Machine Learning

cs.LG Watch Machine Learning (sign in to use)

Tuesday 20 January

Machine Learning

KAGE-Bench: Fast Known-Axis Visual Generalization Evaluation for Reinforcement Learning

Pixel-based reinforcement learning agents often fail under purely visual distribution shift even when latent dynamics and rewards are unchanged, but existing benchmarks entangle multiple sources of shift and hinder systematic analysis. We introduce KAGE-Env, a JAX-native 2D platformer that factorizes the observation process into independently controllable visual axes while keeping the underlying control problem fixed. By construction, varying a visual axis affects performance only through the induced state-conditional action distribution of a pixel policy, providing a clean abstraction for visual generalization. Building on this environment, we define KAGE-Bench, a benchmark of six known-axis suites comprising 34 train-evaluation configuration pairs that isolate individual visual shifts. Using a standard PPO-CNN baseline, we observe strong axis-dependent failures, with background and photometric shifts often collapsing success, while agent-appearance shifts are comparatively benign. Several shifts preserve forward motion while breaking task completion, showing that return alone can obscure generalization failures. Finally, the fully vectorized JAX implementation enables up to 33M environment steps per second on a single GPU, enabling fast and reproducible sweeps over visual factors. Code: https://avanturist322.github.io/KAGEBench/.
Read the full abstract

All authors Egor Cherepanov, Daniil Zelezetsky, Alexey K. Kovalev, Aleksandr I. Panov

Machine Learning

Attention-Based Offline Reinforcement Learning and Clustering for Interpretable Sepsis Treatment

Sepsis remains one of the leading causes of mortality in intensive care units, where timely and accurate treatment decisions can significantly impact patient outcomes. In this work, we propose an interpretable decision support framework. Our system integrates four core components: (1) a clustering-based stratification module that categorizes patients into low, intermediate, and high-risk groups upon ICU admission, using clustering with statistical validation; (2) a synthetic data augmentation pipeline leveraging variational autoencoders (VAE) and diffusion models to enrich underrepresented trajectories such as fluid or vasopressor administration; (3) an offline reinforcement learning (RL) agent trained using Advantage Weighted Regression (AWR) with a lightweight attention encoder and supported by an ensemble models for conservative, safety-aware treatment recommendations; and (4) a rationale generation module powered by a multi-modal large language model (LLM), which produces natural-language justifications grounded in clinical context and retrieved expert knowledge. Evaluated on the MIMIC-III and eICU datasets, our approach achieves high treatment accuracy while providing clinicians with interpretable and robust policy recommendations.
Read the full abstract

All authors Punit Kumar, Vaibhav Saran, Divyesh Patel, Nitin Kulkarni, Alina Vereshchaka

Machine Learning · Cross-listed from quant-ph

Deep Learning Approaches to Quantum Error Mitigation

We present a systematic investigation of deep learning methods applied to quantum error mitigation of noisy output probability distributions from measured quantum circuits. We compare different architectures, from fully connected neural networks to transformers, and we test different design/training modalities, identifying sequence-to-sequence, attention-based models as the most effective on our datasets. These models consistently produce mitigated distributions that are closer to the ideal outputs when tested on both simulated and real device data obtained from IBM superconducting quantum processing units (QPU) up to five qubits. Across several different circuit depths, our approach outperforms other baseline error mitigation techniques. We perform a series of ablation studies to examine: how different input features (circuit, device properties, noisy output statistics) affect performance; cross-dataset generalization across circuit families; and transfer learning to a different IBM QPU. We observe that generalization performance across similar devices with the same architecture works effectively, without needing to fully retrain models.
Read the full abstract

All authors Leonardo Placidi, Ifan Williams, Enrico Rinaldi, Daniel Mills, Cristina Cîrstoiu, Vanya Eccles, Ross Duncan

Machine Learning

InT: Self-Proposed Interventions Enable Credit Assignment in LLM Reasoning

Outcome-reward reinforcement learning (RL) has proven effective at improving the reasoning capabilities of large language models (LLMs). However, standard RL assigns credit only at the level of the final answer, penalizing entire reasoning traces when the outcome is incorrect and uniformly reinforcing all steps when it is correct. As a result, correct intermediate steps may be discouraged in failed traces, while spurious steps may be reinforced in successful ones. We refer to this failure mode as the problem of credit assignment. While a natural remedy is to train a process reward model, accurately optimizing such models to identify corrective reasoning steps remains challenging. We introduce Intervention Training (InT), a training paradigm in which the model performs fine-grained credit assignment on its own reasoning traces by proposing short, targeted corrections that steer trajectories toward higher reward. Using reference solutions commonly available in mathematical reasoning datasets and exploiting the fact that verifying a model-generated solution is easier than generating a correct one from scratch, the model identifies the first error in its reasoning and proposes a single-step intervention to redirect the trajectory toward the correct solution. We then apply supervised fine-tuning (SFT) to the on-policy rollout up to the point of error concatenated with the intervention, localizing error to the specific step that caused failure. We show that the resulting model serves as a far better initialization for RL training. After running InT and subsequent fine-tuning with RL, we improve accuracy by nearly 14% over a 4B-parameter base model on IMO-AnswerBench, outperforming larger open-source models such as gpt-oss-20b.
Read the full abstract

All authors Matthew Y. R. Yang, Hao Bai, Ian Wu, Gene Yang, Amrith Setlur, Aviral Kumar

Machine Learning · Cross-listed from cs.CV

Rig-Aware 3D Reconstruction of Vehicle Undercarriages using Gaussian Splatting

Inspecting the undercarriage of used vehicles is a labor-intensive task that requires inspectors to crouch or crawl underneath each vehicle to thoroughly examine it. Additionally, online buyers rarely see undercarriage photos. We present an end-to-end pipeline that utilizes a three-camera rig to capture videos of the undercarriage as the vehicle drives over it, and produces an interactive 3D model of the undercarriage. The 3D model enables inspectors and customers to rotate, zoom, and slice through the undercarriage, allowing them to detect rust, leaks, or impact damage in seconds, thereby improving both workplace safety and buyer confidence. Our primary contribution is a rig-aware Structure-from-Motion (SfM) pipeline specifically designed to overcome the challenges of wide-angle lens distortion and low-parallax scenes. Our method overcomes the challenges of wide-angle lens distortion and low-parallax scenes by integrating precise camera calibration, synchronized video streams, and strong geometric priors from the camera rig. We use a constrained matching strategy with learned components, the DISK feature extractor, and the attention-based LightGlue matcher to generate high-quality sparse point clouds that are often unattainable with standard SfM pipelines. These point clouds seed the Gaussian splatting process to generate photorealistic undercarriage models that render in real-time. Our experiments and ablation studies demonstrate that our design choices are essential to achieve state-of-the-art quality.
Read the full abstract

All authors Nitin Kulkarni, Akhil Devarashetti, Charlie Cluss, Livio Forte, Dan Buckmaster, Philip Schneider, Chunming Qiao, Alina Vereshchaka

Machine Learning

Differentiated Pickup Point Offering for Emission Reduction in Last-Mile Delivery

Pickup points are widely recognized as a sustainable alternative to home delivery, as consolidating orders at pickup locations can shorten delivery routes and improve first-attempt success rates. However, these benefits may be negated when customers drive to pick up their orders. This study proposes a Differentiated Pickup Point Offering (DPO) policy that aims to jointly reduce emissions from delivery truck routes and customer travel. Under DPO, each arriving customer is offered a single recommended pickup point, rather than an unrestricted choice among all locations, while retaining the option of home delivery. We study this problem in a dynamic and stochastic setting, where the pickup point offered to each customer depends on previously realized customer locations and delivery choices. To design effective DPO policies, we adopt a reinforcement learning-based approach that accounts for spatial relationships between customers and pickup points and their implications for future route consolidation. Computational experiments show that differentiated pickup point offerings can substantially reduce total carbon emissions. The proposed policies reduce total emissions by up to 9% relative to home-only delivery and by 2% on average compared with alternative policies, including unrestricted pickup point choice and nearest pickup point assignment. Differentiated offerings are particularly effective in dense urban settings with many pickup points and short inter-location distances. Moreover, explicitly accounting for the dynamic nature of customer arrivals and choices is especially important when customers are less inclined to choose pickup point delivery over home delivery.
Read the full abstract

Machine Learning

A model of errors in transformers

We study the error rate of LLMs on tasks like arithmetic that require a deterministic output, and repetitive processing of tokens drawn from a small set of alternatives. We argue that incorrect predictions arise when small errors in the attention mechanism accumulate to cross a threshold, and use this insight to derive a quantitative two-parameter relationship between the accuracy and the complexity of the task. The two parameters vary with the prompt and the model; they can be interpreted in terms of an elementary noise rate, and the number of plausible erroneous tokens that can be predicted. Our analysis is inspired by an “effective field theory” perspective: the LLM's many raw parameters can be reorganized into just two parameters that govern the error rate. We perform extensive empirical tests, using Gemini 2.5 Flash, Gemini 2.5 Pro and DeepSeek R1, and find excellent agreement between the predicted and observed accuracy for a variety of tasks, although we also identify deviations in some cases. Our model provides an alternative to suggestions that errors made by LLMs on long repetitive tasks indicate the “collapse of reasoning”, or an inability to express “compositional” functions. Finally, we show how to construct prompts to reduce the error rate.
Read the full abstract

Machine Learning

Penalizing Localized Dirichlet Energies in Low Rank Tensor Products

We study low-rank tensor-product B-spline (TPBS) models for regression tasks and investigate Dirichlet energy as a measure of smoothness. We show that TPBS models admit a closed-form expression for the Dirichlet energy, and reveal scenarios where perfect interpolation is possible with exponentially small Dirichlet energy. This renders global Dirichlet energy-based regularization ineffective. To address this limitation, we propose a novel regularization strategy based on local Dirichlet energies defined on small hypercubes centered at the training points. Leveraging pretrained TPBS models, we also introduce two estimators for inference from incomplete samples. Comparative experiments with neural networks demonstrate that TPBS models outperform neural networks in the overfitting regime for most datasets, and maintain competitive performance otherwise. Overall, TPBS models exhibit greater robustness to overfitting and consistently benefit from regularization, while neural networks are more sensitive to overfitting and less effective in leveraging regularization.
Read the full abstract

Machine Learning · Cross-listed from cs.SD

ConceptCaps: a Distilled Concept Dataset for Interpretability in Music Models

Concept-based interpretability methods like TCAV require clean, well-separated positive and negative examples for each concept. Existing music datasets lack this structure: tags are sparse, noisy, or ill-defined. We introduce ConceptCaps, a dataset of 21k music-caption-tags triplets with explicit labels from a 200-attribute taxonomy. Our pipeline separates semantic modeling from text generation: a VAE learns plausible attribute co-occurrence patterns, a fine-tuned LLM converts attribute lists into professional descriptions, and MusicGen synthesizes corresponding audio. This separation improves coherence and controllability over end-to-end approaches. We validate the dataset through audio-text alignment (CLAP), linguistic quality metrics (BERTScore, MAUVE), and TCAV analysis confirming that concept probes recover musically meaningful patterns. Dataset and code are available online.
Read the full abstract

Machine Learning · Cross-listed from cs.CL

Lost in the Prompt Order: Revealing the Limitations of Causal Attention in Language Models

Large language models exhibit surprising sensitivity to the structure of the prompt, but the mechanisms underlying this sensitivity remain poorly understood. In this work, we conduct an in-depth investigation on a striking case: in multiple-choice question answering, placing context before the questions and options (CQO) outperforms the reverse order (QOC) by over 14%p, consistently over a wide range of models and datasets. Through systematic architectural analysis, we identify causal attention as the core mechanism: in QOC prompts, the causal mask prevents option tokens from attending to context, creating an information bottleneck where context becomes invisible to options.
Read the full abstract

Machine Learning

Riemannian Liquid Spatio-Temporal Graph Network

Liquid Time-Constant networks (LTCs), a type of continuous-time graph neural network, excel at modeling irregularly-sampled dynamics but are fundamentally confined to Euclidean space. This limitation introduces significant geometric distortion when representing real-world graphs with inherent non-Euclidean structures (e.g., hierarchies and cycles), degrading representation quality. To overcome this limitation, we introduce the Riemannian Liquid Spatio-Temporal Graph Network (RLSTG), a framework that unifies continuous-time liquid dynamics with the geometric inductive biases of Riemannian manifolds. RLSTG models graph evolution through an Ordinary Differential Equation (ODE) formulated directly on a curved manifold, enabling it to faithfully capture the intrinsic geometry of both structurally static and dynamic spatio-temporal graphs. Moreover, we provide rigorous theoretical guarantees for RLSTG, extending stability theorems of LTCs to the Riemannian domain and quantifying its expressive power via state trajectory analysis. Extensive experiments on real-world benchmarks demonstrate that, by combining advanced temporal dynamics with a Riemannian spatial representation, RLSTG achieves superior performance on graphs with complex structures. Project Page: https://rlstg.github.io
Read the full abstract

All authors Liangsi Lu, Jingchao Wang, Zhaorong Dai, Hanqian Liu, Yang Shi

Machine Learning · Cross-listed from cs.CL

Learning to Explain: Supervised Token Attribution from Transformer Attention Patterns

Explainable AI (XAI) has become critical as transformer-based models are deployed in high-stakes applications including healthcare, legal systems, and financial services, where opacity hinders trust and accountability. Transformers self-attention mechanisms have proven valuable for model interpretability, with attention weights successfully used to understand model focus and behavior (Xu et al., 2015); (Wiegreffe and Pinter, 2019). However, existing attention-based explanation methods rely on manually defined aggregation strategies and fixed attribution rules (Abnar and Zuidema, 2020a); (Chefer et al., 2021), while model-agnostic approaches (LIME, SHAP) treat the model as a black box and incur significant computational costs through input perturbation. We introduce Explanation Network (ExpNet), a lightweight neural network that learns an explicit mapping from transformer attention patterns to token-level importance scores. Unlike prior methods, ExpNet discovers optimal attention feature combinations automatically rather than relying on predetermined rules. We evaluate ExpNet in a challenging cross-task setting and benchmark it against a broad spectrum of model-agnostic methods and attention-based techniques spanning four methodological families.
Read the full abstract

Machine Learning

Causal feature selection framework for stable soft sensor modeling based on time-delayed cross mapping

Soft sensor modeling plays a crucial role in process monitoring. Causal feature selection can enhance the performance of soft sensor models in industrial applications. However, existing methods ignore two critical characteristics of industrial processes. Firstly, causal relationships between variables always involve time delays, whereas most causal feature selection methods investigate causal relationships in the same time dimension. Secondly, variables in industrial processes are often interdependent, which contradicts the decorrelation assumption of traditional causal inference methods. Consequently, soft sensor models based on existing causal feature selection approaches often lack sufficient accuracy and stability. To overcome these challenges, this paper proposes a causal feature selection framework based on time-delayed cross mapping. Time-delayed cross mapping employs state space reconstruction to effectively handle interdependent variables in causality analysis, and considers varying causal strength across time delay. Time-delayed convergent cross mapping (TDCCM) is introduced for total causal inference, and time-delayed partial cross mapping (TDPCM) is developed for direct causal inference. Then, in order to achieve automatic feature selection, an objective feature selection strategy is presented. The causal threshold is automatically determined based on the model performance on the validation set, and the causal features are then selected. Two real-world case studies show that TDCCM achieves the highest average performance, while TDPCM improves soft sensor stability and performance in the worst scenario. The code is publicly available at https://github.com/dirge1/TDPCM.
Read the full abstract

Machine Learning

Optimizing Energy and Data Collection in UAV-aided IoT Networks using Attention-based Multi-Objective Reinforcement Learning

Due to their adaptability and mobility, Unmanned Aerial Vehicles (UAVs) are becoming increasingly essential for wireless network services, particularly for data harvesting tasks. In this context, Artificial Intelligence (AI)-based approaches have gained significant attention for addressing UAV path planning tasks in large and complex environments, bridging the gap with real-world deployments. However, many existing algorithms suffer from limited training data, which hampers their performance in highly dynamic environments. Moreover, they often overlook the inherently multi-objective nature of the task, treating it in an overly simplistic manner. To address these limitations, we propose an attention-based Multi-Objective Reinforcement Learning (MORL) architecture that explicitly handles the trade-off between data collection and energy consumption in urban environments, even without prior knowledge of wireless channel conditions. Our method develops a single model capable of adapting to varying trade-off preferences and dynamic scenario parameters without the need for fine-tuning or retraining. Extensive simulations show that our approach achieves substantial improvements in performance, model compactness, sample efficiency, and most importantly, generalization to previously unseen scenarios, outperforming existing RL solutions.
Read the full abstract

All authors Babacar Toure, Dimitrios Tsilimantos, Omid Esrafilian, Marios Kountouris

Machine Learning · Cross-listed from cs.AR

'1'-bit Count-based Sorting Unit to Reduce Link Power in DNN Accelerators

Interconnect power consumption remains a bottleneck in Deep Neural Network (DNN) accelerators. While ordering data based on '1'-bit counts can mitigate this via reduced switching activity, practical hardware sorting implementations remain underexplored. This work proposes the hardware implementation of a comparison-free sorting unit optimized for Convolutional Neural Networks (CNN). By leveraging approximate computing to group population counts into coarse-grained buckets, our design achieves hardware area reductions while preserving the link power benefits of data reordering. Our approximate sorting unit achieves up to 35.4% area reduction while maintaining 19.50% BT reduction compared to 20.42% of precise implementation.
Read the full abstract

All authors Ruichi Han, Yizhi Chen, Tong Lei, Jordi Altayo Gonzalez, Ahmed Hemani

Machine Learning · Cross-listed from cs.CV

Two-Stream temporal transformer for video action classification

Motion representation plays an important role in video understanding and has many applications including action recognition, robot and autonomous guidance or others. Lately, transformer networks, through their self-attention mechanism capabilities, have proved their efficiency in many applications. In this study, we introduce a new two-stream transformer video classifier, which extracts spatio-temporal information from content and optical flow representing movement information. The proposed model identifies self-attention features across the joint optical flow and temporal frame domain and represents their relationships within the transformer encoder mechanism. The experimental results show that our proposed methodology provides excellent classification results on three well-known video datasets of human activities.
Read the full abstract

Machine Learning · Cross-listed from cs.CV

Unsupervised Video Class-Incremental Learning via Deep Embedded Clustering Management

Unsupervised video class incremental learning (uVCIL) represents an important learning paradigm for learning video information without forgetting, and without considering any data labels. Prior approaches have focused on supervised class-incremental learning, relying on using the knowledge of labels and task boundaries, which is costly, requires human annotation, or is simply not a realistic option. In this paper, we propose a simple yet effective approach to address the uVCIL. We first consider a deep feature extractor network, providing a set of representative video features during each task without assuming any class or task information. We then progressively build a series of deep clusters from the extracted features. During the successive task learning, the model updated from the previous task is used as an initial state in order to transfer knowledge to the current learning task. We perform in-depth evaluations on three standard video action recognition datasets, including UCF101, HMDB51, and Something-to-Something V2, by ignoring the labels from the supervised setting. Our approach significantly outperforms other baselines on all datasets.
Read the full abstract

Machine Learning · Cross-listed from cs.CR

SecureSplit: Mitigating Backdoor Attacks in Split Learning

Split Learning (SL) offers a framework for collaborative model training that respects data privacy by allowing participants to share the same dataset while maintaining distinct feature sets. However, SL is susceptible to backdoor attacks, in which malicious clients subtly alter their embeddings to insert hidden triggers that compromise the final trained model. To address this vulnerability, we introduce SecureSplit, a defense mechanism tailored to SL. SecureSplit applies a dimensionality transformation strategy to accentuate subtle differences between benign and poisoned embeddings, facilitating their separation. With this enhanced distinction, we develop an adaptive filtering approach that uses a majority-based voting scheme to remove contaminated embeddings while preserving clean ones. Rigorous experiments across four datasets (CIFAR-10, MNIST, CINIC-10, and ImageNette), five backdoor attack scenarios, and seven alternative defenses confirm the effectiveness of SecureSplit under various challenging conditions.
Read the full abstract

All authors Zhihao Dou, Dongfei Cui, Weida Wang, Anjun Gao, Yueyang Quan, Mengyao Ma, Viet Vo, Guangdong Bai, Zhuqing Liu, Minghong Fang

Machine Learning

LLMOrbit: A Circular Taxonomy of Large Language Models -From Scaling Walls to Agentic AI Systems

The field of artificial intelligence has undergone a revolution from foundational Transformer architectures to reasoning-capable systems approaching human-level performance. We present LLMOrbit, a comprehensive circular taxonomy navigating the landscape of large language models spanning 2019-2025. This survey examines over 50 models across 15 organizations through eight interconnected orbital dimensions, documenting architectural innovations, training methodologies, and efficiency patterns defining modern LLMs, generative AI, and agentic systems. We identify three critical crises: (1) data scarcity (9-27T tokens depleted by 2026-2028), (2) exponential cost growth ($3M to $300M+ in 5 years), and (3) unsustainable energy consumption (22x increase), establishing the scaling wall limiting brute-force approaches. Our analysis reveals six paradigms breaking this wall: (1) test-time compute (o1, DeepSeek-R1 achieve GPT-4 performance with 10x inference compute), (2) quantization (4-8x compression), (3) distributed edge computing (10x cost reduction), (4) model merging, (5) efficient training (ORPO reduces memory 50%), and (6) small specialized models (Phi-4 14B matches larger models). Three paradigm shifts emerge: (1) post-training gains (RLHF, GRPO, pure RL contribute substantially, DeepSeek-R1 achieving 79.8% MATH), (2) efficiency revolution (MoE routing 18x efficiency, Multi-head Latent Attention 8x KV cache compression enables GPT-4-level performance at $<$$0.30/M tokens), and (3) democratization (open-source Llama 3 88.6% MMLU surpasses GPT-4 86.4%). We provide insights into techniques (RLHF, PPO, DPO, GRPO, ORPO), trace evolution from passive generation to tool-using agents (ReAct, RAG, multi-agent systems), and analyze post-training innovations.
Read the full abstract

Machine Learning · Cross-listed from cs.CL

Kakugo: Distillation of Low-Resource Languages into Small Language Models

We present Kakugo, a novel and cost-effective pipeline designed to train general-purpose Small Language Models (SLMs) for low-resource languages using only the language name as input. By using a large teacher model to generate synthetic prompts and translate instruction datasets, we produced training data and SLMs for 54 low-resource languages. Evaluations across a diverse set of general natural language processing tasks, including translation, classification, and question answering, demonstrate that our pipeline consistently improves performance over base models. With a total generation and training cost of under $50 per language, Kakugo offers an accessible method for communities to develop language-specific AI.
Read the full abstract

All authors Peter Devine, Mardhiyah Sanni, Farid Adilazuarda, Julieta Gil Loizaga, Barry Haddow

Machine Learning · Cross-listed from cs.CV

Federated Balanced Learning

Federated learning is a paradigm of joint learning in which clients collaborate by sharing model parameters instead of data. However, in the non-iid setting, the global model experiences client drift, which can seriously affect the final performance of the model. Previous methods tend to correct the global model that has already deviated based on the loss function or gradient, overlooking the impact of the client samples. In this paper, we rethink the role of the client side and propose Federated Balanced Learning, i.e., FBL, to prevent this issue from the beginning through sample balance on the client side. Technically, FBL allows unbalanced data on the client side to achieve sample balance through knowledge filling and knowledge sampling using edge-side generation models, under the limitation of a fixed number of data samples on clients. Furthermore, we design a Knowledge Alignment Strategy to bridge the gap between synthetic and real data, and a Knowledge Drop Strategy to regularize our method. Meanwhile, we scale our method to real and complex scenarios, allowing different clients to adopt various methods, and extend our framework to further improve performance. Numerous experiments show that our method outperforms state-of-the-art baselines. The code is released upon acceptance.
Read the full abstract

All authors Jiaze Li, Haoran Xu, Wanyi Wu, Changwei Wang, Shuaiguang Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, Youyang Qu, Longxiang Gao, Xudong Yang, Lumin Xing

Machine Learning

Private Prediction via PAC Privacy

Machine learning models are increasingly served behind APIs. This renders private prediction, i.e., privatizing a model's outputs rather than its parameters, a natural privacy target: model outputs are lower-dimensional and far more stable to training-data changes than weights. While differential privacy (DP) cannot effectively exploit this as it calibrates noise to worst-case sensitivity that is intractable to bound for non-convex models, we argue that PAC privacy is a natural fit for private prediction. It is instance-based, and calibrates noise to a black-box function's empirical stability to control mutual-information (MI) leakage. The missing ingredient is efficient, adaptive composition. Serving predictions means answering a long stream of adaptively chosen queries from untrusted users; existing composition either fails under adaptivity, grows quadratically, or reverts to input-independent, DP-like noise. We close this gap with a new adversarial composition result via adaptive noise calibration and prove that MI accumulates only linearly under adaptive and adversarial querying. Experiments across modalities show that prediction stability enables high utility even at a tiny per-query budget: on CIFAR-10, we achieve 87.79% accuracy with a per-query MI budget of $2^{-32}$. This enables serving one million queries while provably bounding membership-inference success to 51.08% – the same guarantee as $(0.04, 10^{-5})$-DP. Further, in the presence of auxiliary public data, the large volume of PAC-private predictions enables us to distill a publishable model that can be queried without limit. Concretely, 210,000 private labels on an ImageNet subset distill into a student reaching 91.86% accuracy on CIFAR-10 with membership inference success bounded by 50.49%, comparable to $(0.02, 10^{-5})$-DP.
Read the full abstract

Machine Learning · Cross-listed from stat.ML

Intermittent time series forecasting: local vs global models

Forecasting intermittent time series, which contain zeros, is a crucial challenge in supply chains as inventory policies require probabilistic forecasts to establish safety levels. Intermittent time series are commonly forecast using local models, trained individually on each time series. In the last years global models, trained on a large collection of time series, have become popular for time series forecasting. Global models are often based on neural networks or gradient boosted trees. We carry out the first study comparing state-of-the-art probabilistic local and global models on intermittent time series. For global models we consider three different distribution heads suitable for intermittent time series: negative binomial, hurdle-shifted negative binomial and Tweedie. To the best of our knowledge, this is the first use of the latter two with neural networks. We perform experiments on five datasets comprising overall more than 40'000 real-world time series. Among global models, TiDE, a simple neural network architecture, achieves the best accuracy; it also consistently outperforms local models and has lower computational requirements. Large global models are instead much more computationally demanding and less accurate. Among the distribution heads, the Tweedie provides the best estimates of the highest quantiles.
Read the full abstract

All authors Stefano Damato, Nicolò Rubattu, Dario Azzimonti, Giorgio Corani

Machine Learning

Universal Approximation Theorem for Input-Connected Multilayer Perceptrons

We present the Input-Connected Multilayer Perceptron (IC-MLP), a feedforward neural network architecture in which each hidden neuron receives, in addition to the outputs of the preceding layer, a direct affine connection from the raw input. We first study this architecture in the univariate setting and give an explicit and systematic description of IC-MLPs with an arbitrary finite number of hidden layers, including iterated formulas for the network functions. In this setting, we prove a universal approximation theorem showing that deep IC-MLPs can approximate any continuous function on a closed interval of the real line if and only if the activation function is nonlinear. We then extend the analysis to vector-valued inputs and establish a corresponding universal approximation theorem for continuous functions on compact subsets of $\mathbb{R}^n$.
Read the full abstract

Machine Learning

Credible CO2 Comparisons: A Machine Learning Approach to Vehicle Powertrain Assessment

Decarbonizing road transport requires consistent and transparent methods for comparing CO2 emissions across vehicle technologies. This paper proposes a machine learning-based framework for like-for-like operational assessment of internal combustion engine vehicles (ICEVs) and electric vehicles (EVs) under identical, real-world driving conditions. The approach isolates technology-specific effects by holding the observed speed profile and environmental context fixed, enabling direct comparison of powertrain performance. Recurrent neural network models are trained independently for each domain to learn the mapping from contextual driving variables (speed, acceleration, temperature) to internal actuation variables (torque, throttle) and instantaneous CO2-equivalent emission rates. This structure allows the construction of counterfactual scenarios that answer: What emissions would an EV have generated if it had followed the same driving profile as an ICEV? By aligning both vehicle types on a unified instantaneous emissions metric, the framework enables fair and reproducible evaluation of powertrain technologies. It offers a scalable foundation for credible, data-driven assessments of vehicle carbon performance under real-world operating conditions.
Read the full abstract

All authors Rodrigo Pereira David, Luciano Araujo Dourado Filho, Daniel Marques da Silva, João Alfredo Cal-Braz

Machine Learning · Cross-listed from cs.IR

Cross-Sensory Brain Passage Retrieval: Scaling Beyond Visual to Audio

Query formulation from internal information needs remains fundamentally challenging across all Information Retrieval paradigms due to cognitive complexity and physical impairments. Brain Passage Retrieval (BPR) addresses this by directly mapping EEG signals to passage representations without intermediate text translation. However, existing BPR research exclusively uses visual stimuli, leaving critical questions unanswered: Can auditory EEG enable effective retrieval for voice-based interfaces and visually impaired users? Can training on combined EEG datasets from different sensory modalities improve performance despite severe data scarcity? We present the first systematic investigation of auditory EEG for BPR and evaluate cross-sensory training benefits. Using dual encoder architectures with four pooling strategies (CLS, mean, max, multi-vector), we conduct controlled experiments comparing auditory-only, visual-only, and combined training on the Alice (auditory) and Nieuwland (visual) datasets. Results demonstrate that auditory EEG consistently outperforms visual EEG, and cross-sensory training with CLS pooling achieves substantial improvements over individual training: 31% in MRR (0.474), 43% in Hit@1 (0.314), and 28% in Hit@10 (0.858). Critically, combined auditory EEG models surpass BM25 text baselines (MRR: 0.474 vs 0.428), establishing neural queries as competitive with traditional retrieval whilst enabling accessible interfaces. These findings validate auditory neural interfaces for IR tasks and demonstrate that cross-sensory training addresses data scarcity whilst outperforming single-modality approaches Code: https://github.com/NiallMcguire/Audio_BPR
Read the full abstract

Machine Learning · Cross-listed from cs.RO

Group-Invariant Unsupervised Skill Discovery: Symmetry-aware Skill Representations for Generalizable Behavior

Unsupervised skill discovery aims to acquire behavior primitives that improve exploration and accelerate downstream task learning. However, existing approaches often ignore the geometric symmetries of physical environments, leading to redundant behaviors and sample inefficiency. To address this, we introduce Group-Invariant Skill Discovery (GISD), a framework that explicitly embeds group structure into the skill discovery objective. Our approach is grounded in a theoretical guarantee: we prove that in group-symmetric environments, the standard Wasserstein dependency measure admits a globally optimal solution comprised of an equivariant policy and a group-invariant scoring function. Motivated by this, we formulate the Group-Invariant Wasserstein dependency measure, which restricts the optimization to this symmetry-aware subspace without loss of optimality. Practically, we parameterize the scoring function using a group Fourier representation and define the intrinsic reward via the alignment of equivariant latent features, ensuring that the discovered skills generalize systematically under group transformations. Experiments on state-based and pixel-based locomotion benchmarks demonstrate that GISD achieves broader state-space coverage and improved efficiency in downstream task learning compared to a strong baseline.
Read the full abstract

All authors Junwoo Chang, Joseph Park, Roberto Horowitz, Jongmin Lee, Jongeun Choi

Machine Learning

Linearized subspace refinement framework to expose hidden accuracy in trained neural networks

Neural networks trained by gradient-based methods often exhibit optimization-induced accuracy plateaus in scientific machine learning tasks. We present Linearized Subspace Refinement (LSR), an architecture-agnostic post-training framework that exploits the local linearized model at a fixed trained state. By solving a reduced direct least-squares problem in a Jacobian-defined low-dimensional space, LSR computes a subspace-optimal linearized correction and yields a refined predictor with markedly improved accuracy. Across function approximation, data-driven operator learning, physics-informed operator fine-tuning, and noisy inverse problems, LSR shows that standard nonlinear training can remain far above this subspace-attainable error level. Similar accuracy plateaus persist even for the convex quadratic problem from local linearization when solved with standard iterative optimizers, identifying numerical ill-conditioning as a primary bottleneck. LSR frequently delivers order-of-magnitude error reductions, while the subspace rank provides an explicit capacity-control mechanism that balances correction strength, numerical stability, and noise sensitivity. Together, LSR exposes conditioning-limited attainable accuracy in trained-state linearized models and provides direct access to it.
Read the full abstract

Machine Learning · Cross-listed from cs.CV

Harmonizing the Deep: A Unified Information Pipeline for Robust Marine Biodiversity Assessment Across Heterogeneous Domains

Marine biodiversity monitoring requires scalability and reliability across complex underwater environments to support conservation and invasive-species management. Yet existing detection solutions often exhibit a pronounced deployment gap, with performance degrading sharply when transferred to new sites. This work establishes the foundational detection layer for a multi-year invasive species monitoring initiative targeting Arctic and Atlantic marine ecosystems. We address this challenge by developing a Unified Information Pipeline that standardises heterogeneous datasets into a comparable information flow and evaluates a fixed, deployment-relevant detector under controlled cross-domain protocols. Across multiple domains, we find that structural factors, such as scene composition, object density, and contextual redundancy, explain cross-domain performance loss more strongly than visual degradation such as turbidity, with sparse scenes inducing a characteristic "Context Collapse" failure mode. We further validate operational feasibility by benchmarking inference on low-cost edge hardware, showing that runtime optimisation enables practical sampling rates for remote monitoring. The results shift emphasis from image enhancement toward structure-aware reliability, providing a democratised tool for consistent marine ecosystem assessment.
Read the full abstract

All authors Marco Piccolo, Qiwei Han, Astrid van Toor, Joachim Vanneste

Machine Learning · Cross-listed from cs.AI

Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval

Retrieving evidence for language model queries from knowledge graphs requires balancing broad search across the graph with multi-hop traversal to follow relational links. Similarity-based retrievers provide coverage but remain shallow, whereas traversal-based methods rely on selecting seed nodes to start exploration, which can fail when queries span multiple entities and relations. We introduce ARK: Adaptive Retriever of Knowledge, a tool-using KG retriever that gives a language model control over this breadth-depth tradeoff using a two-operation toolset: global lexical search over node descriptors and one-hop neighborhood exploration that composes into multi-hop traversal. ARK alternates between breadth-oriented discovery and depth-oriented expansion without depending on a fragile seed selection, a pre-set hop depth, or requiring retrieval training. ARK adapts tool use to queries, using global search for language-heavy queries and neighborhood exploration for relation-heavy queries. On STaRK, ARK reaches 59.1% average Hit@1 and 67.4 average MRR, improving average Hit@1 by up to 31.4% and average MRR by up to 28.0% over retrieval-based and agent-based training-free methods. Finally, we distill ARK's tool-use trajectories from a large teacher into an 8B model via label-free imitation, improving Hit@1 by +7.0, +26.6, and +13.5 absolute points over the base 8B model on AMAZON, MAG, and PRIME datasets, respectively, while retaining up to 98.5% of the teacher's Hit@1 rate.
Read the full abstract

All authors Joaquín Polonuer, Lucas Vittor, Iñaki Arango, Ayush Noori, David A. Clifton, Luciano Del Corro, Marinka Zitnik