AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
52
Papers today
8h
Update frequency
7
Days of history
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
Interpretability
- Introduces mechanistic tomography as a unified framework for various measurement methods in interpretability.
- Establishes a common language for describing measurements and errors across different methods.
- Defines calibration dimension to determine the necessary correction family for accurate predictions.
- Demonstrates the effectiveness of the proposed measurements through experiments on pretrained models.
Read more
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
Summary
This paper introduces the concept of mechanistic tomography, which aims to unify various measurement methods used in mechanistic interpretability of machine learning models. The author formulates a common measurement problem that encompasses different approaches such as coordinate patching, attribution patching, subset interventions, and Hessian-vector products. By expressing these methods in a common framework, the paper facilitates a structured approach to understanding and recovering internal mechanisms and intervention effects in models. The methodology involves designing measurements that can effectively recover the desired internal maps while accounting for errors from nonlinearity, sampling, and basis misspecification. The paper also emphasizes the importance of calibration and the need for a systematic approach to expand measurement families based on residual errors. The validation of the proposed methods is demonstrated through experiments on two models, GPT-2-small and Qwen-2.5-7B, showcasing the effectiveness of the measurements in identifying significant interactions and improving predictive performance. Overall, the work highlights the necessity of a measurement-first approach in mechanistic interpretability, especially as models become more complex and less transparent.
Methodology
The paper formulates the measurement problem as eπ¦ = π΄π₯ + π€, where A represents interventions, x is the target map to recover, and w accounts for various errors. It explores the design of measurements, calibration, and the conditions under which different measurement methods can be effectively applied. The methodology includes testing on held-out interventions and expanding measurement families based on observed residuals.
Results
The experiments demonstrate that the proposed measurement methods can effectively recover significant interactions and improve predictive performance. For instance, on the GPT-2-small model, the measurements identified the Name MoverβNegative Name Mover interaction as a key predictive term. On the Qwen-2.5-7B model, a calibrated additive map achieved an RΒ² of 0.983 on a finite refusal-response surface, indicating high predictive accuracy.
Implications
The findings suggest that mechanistic tomography can enhance the interpretability of complex models by providing a structured approach to understanding internal mechanisms and their effects. This has potential applications in improving model transparency, guiding interventions, and informing the design of future models.
Online Test-Time Adaptation for Generalizable Dynamic Graph Anomaly Detection
Graph Learning
Time Series
- OTTA-DGAD enables online adaptation for dynamic graph anomaly detection without retraining.
- The method captures both general and domain-specific anomaly patterns using dynamic prototypes.
- It effectively handles sequentially arriving unlabeled data chunks, preserving cross-chunk information.
- Extensive experiments show OTTA-DGAD outperforms existing methods on diverse datasets.
Read more
Online Test-Time Adaptation for Generalizable Dynamic Graph Anomaly Detection
Summary
This paper addresses the challenge of dynamic graph anomaly detection (DGAD) in unseen target domains without the need for costly retraining. The authors propose a novel framework called OTTA-DGAD, which enables online test-time adaptation. The method extracts dynamic prototypes from temporal ego-graphs and stores them in a memory buffer, allowing the model to retain general patterns from source domains while incorporating new domain-specific patterns from incoming target data. An anomaly scorer then compares edge representations against these prototypes to identify both general and domain-specific anomalies. The adaptation process updates the memory buffer using reliable pseudo-labels and enriches incoming data with relevant representations from previous chunks. The authors demonstrate the effectiveness of OTTA-DGAD through extensive experiments on ten real-world datasets, achieving state-of-the-art performance under strict test-then-adapt settings.
Methodology
The authors developed OTTA-DGAD, which involves extracting dynamic prototypes from temporal ego-graphs and storing them in a memory buffer. The buffer selectively retains general patterns while incorporating new patterns from incoming target data. An anomaly scorer compares edge representations against these prototypes, and the memory buffer is updated using pseudo-labels identified through confidence-based detection. The method also enriches incoming data with relevant representations from previous chunks to mitigate information loss.
Results
OTTA-DGAD achieved state-of-the-art performance on ten real-world datasets, demonstrating its effectiveness in adapting to dynamic graph anomalies in unseen domains. The method successfully identified both general and domain-specific anomalies, outperforming existing approaches under strict test-then-adapt conditions.
Implications
The proposed method has significant implications for real-world applications such as fraud detection, transaction monitoring, and social network security, where timely adaptation to evolving data is crucial. It allows for more robust anomaly detection in dynamic environments without the need for extensive retraining.
Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic
Reinforcement Learning
Theory
Optimization
- Introduces a single-loop NAC algorithm with entropy regularization for improved convergence.
- Analyzes both Stochastic and Deterministic optimization regimes.
- Presents an Exponential Translation mechanism for mapping regularized to unregularized gaps.
- Achieves accelerated convergence rates surpassing traditional statistical limits.
Read more
Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic
Summary
This paper investigates the unregularized convergence properties of a single-loop, entropy-regularized Natural Actor-Critic (NAC) algorithm. While entropy regularization is commonly employed to enhance the stability and speed of Natural Policy Gradient methods, its impact on the convergence rates of the unregularized objective has not been thoroughly examined. The author analyzes the NAC algorithm under linear function approximation, focusing on two optimization regimes: the Stochastic Regime and the Deterministic Regime. In the Stochastic Regime, the paper introduces a joint Lyapunov recurrence for coupled actor-critic updates, while in the Deterministic Regime, it employs a Policy Mirror Descent framework to address the challenges posed by the degeneracy of the Fisher information matrix as the training policy approaches determinism. A key contribution is the Exponential Translation mechanism, which connects the regularized and unregularized gaps, allowing for accelerated convergence rates. The paper presents theoretical results demonstrating that the algorithm achieves convergence rates of O(T^{-1}) in the Stochastic Regime and O(T^{-2/3}) for average iterates in the Deterministic Regime, surpassing traditional statistical barriers. Additionally, the analysis in a tabular setting shows improved performance with a rate of O(T^{-2/3}) under positive action gaps.
Methodology
The paper employs theoretical analysis of a single-loop, entropy-regularized NAC algorithm using linear function approximation. It explores two optimization regimes: Stochastic and Deterministic, utilizing Lyapunov recurrences and Policy Mirror Descent frameworks to analyze convergence properties. The Exponential Translation mechanism is introduced to relate regularized and unregularized objectives.
Results
The proposed algorithm achieves unregularized convergence rates of O(T^{-1}) in the Stochastic Regime and O(T^{-2/3}) for average iterates in the Deterministic Regime. In tabular settings, the analysis yields a rate of O(T^{-2/3}), surpassing the O(T^{-1/2}) worst-case barrier without a positive action margin.
Implications
The findings suggest that entropy regularization can be effectively utilized in single-loop NAC algorithms to enhance convergence rates, potentially leading to more efficient reinforcement learning applications. This work bridges theoretical insights with practical implementations, offering a pathway for improved algorithm designs in reinforcement learning.
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
Generative Models
Theory
Time Series
- Investigates flow-matching-based anomaly detection under contaminated training data.
- Highlights the critical role of anomaly scoring functions in detection performance.
- Demonstrates that trajectory-based scoring improves robustness and detection accuracy.
- Shows that Forest-Flow can outperform TCCM in certain scenarios with suitable scoring functions.
Read more
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
Summary
This paper addresses the challenge of unsupervised anomaly detection in financial datasets that often contain contaminated training data with unlabeled anomalies. The authors explore the effectiveness of flow matching techniques, specifically Time-Conditioned Contraction Matching (TCCM) and Forest-Flow, in detecting anomalies under these conditions. They highlight the critical role of anomaly scoring functions, revealing that TCCM's original Decision score is sensitive to contamination, while trajectory-based scores such as Deviation and Reconstruction provide more stable anomaly signals. The study demonstrates that these trajectory-based scoring methods significantly enhance the robustness and performance of Forest-Flow, making it competitive with TCCM across varying levels of anomaly class imbalance. The findings underscore the importance of selecting appropriate scoring functions for effective anomaly detection in real-world applications where training data is often contaminated.
Methodology
The authors evaluate two flow-matching-based models, TCCM and Forest-Flow, focusing on their ability to detect anomalies in contaminated training datasets. They assess three anomaly scoring functions: Decision, Deviation, and Reconstruction, analyzing their impact on the models' performance. Experiments are conducted on two financial datasets with different anomaly ratios to evaluate the models under varying class imbalances.
Results
The results indicate that TCCM's original Decision score is highly sensitive to contamination, while the trajectory-based Deviation and Reconstruction scores provide more stable and reliable anomaly signals. The study finds that Forest-Flow, when paired with these trajectory-based scoring functions, can match or even surpass the performance of TCCM, particularly in scenarios with severe anomaly class imbalance.
Implications
The findings suggest that careful selection of anomaly scoring functions is crucial for effective anomaly detection in financial applications, where training data is often contaminated. This research could lead to improved methods for detecting financial fraud and other anomalies in high-stakes domains, enhancing the reliability of automated systems in these areas.
Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis
Multimodal
Large Language Models
Time Series
- Introduction of Holtercare-23K, a large-scale dynamic ECG dataset with tri-modal alignment.
- Holtercare-Bench serves as a comprehensive evaluation framework for long-term ECG analysis.
- Significant performance gaps identified in zero-shot evaluations of leading MLLMs on ECG data.
- Fine-tuning of models shows substantial improvements in processing dynamic ECG sequences.
Read more
Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis
Summary
The paper introduces Holtercare-Bench, a novel multimodal benchmark aimed at evaluating long-term dynamic ECG analysis, addressing the limitations of existing datasets and models in processing continuous ECG data. The authors present Holtercare-23K, a large-scale dataset comprising 22,980 question-answer pairs derived from 788 clinical Holter records, which span 13 to 24 hours each. This dataset features a unique tri-modal alignment of signal, video, and text, facilitating the training of multimodal large language models (MLLMs) for complex temporal reasoning and diagnostic report generation. The study highlights the challenges faced by current MLLMs in analyzing ultra-long sequences of ECG data, particularly in identifying fleeting pathological events and linking micro-level waveform changes to macro-level disease labels. The authors conduct zero-shot evaluations of leading MLLMs, revealing significant performance gaps, but demonstrate that fine-tuning these models can lead to substantial improvements. This work not only sheds light on the limitations of existing MLLMs in the field of electrophysiology but also provides a foundational benchmark for future research in long-term medical applications.
Methodology
The authors developed HolterAgent, an automated data engine that converts Holter records into a tri-modal format consisting of signal, video, and text. They constructed a dataset of 22,980 QA pairs from expert clinical annotations and reports, enabling the evaluation of models on tasks such as temporal localization, clinical diagnosis, and global summarization.
Results
The study found that leading MLLMs exhibited significant performance gaps in processing long-term ECG data during zero-shot evaluations. However, fine-tuning these models resulted in notable improvements, demonstrating the potential for enhanced performance in dynamic ECG analysis.
Implications
The introduction of Holtercare-Bench and the Holtercare-23K dataset has the potential to advance research in medical AI, particularly in the field of cardiology. It provides a robust framework for developing and evaluating MLLMs tailored for long-term ECG monitoring, which could improve diagnostic accuracy and patient care in clinical settings.
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
Optimization
Theory
- Localized minibatch influence is formulated as a signed finite-horizon response under paired AdamW trajectories.
- A joint parameter-moment ISO operator is derived to characterize how AdamW memory affects the propagation of minibatch effects.
- An exact multistep error decomposition is established, separating various sources of error and proving first-order accuracy.
- Experiments confirm the proposed mechanisms and reveal substantial prospective structure in delayed influences.
Read more
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
Summary
This paper investigates the delayed effects of minibatch perturbations in the AdamW optimizer, which retains past gradient information, influencing future training updates. The authors formulate AdamW as a finite-horizon input-state-output (ISO) system, allowing them to analyze how a single gradient update affects subsequent loss over time. By comparing paired training trajectories that differ by one update, they isolate the evolution of localized perturbations. The study derives a signed response operator that maps these perturbations to their future loss effects, revealing the role of optimizer memory in shaping the magnitude, timing, and sign of these effects. Additionally, the authors establish an exact multistep error decomposition, addressing the interplay of write-in nonlinearity, smooth propagation errors, and activation switching. Experiments validate the proposed response mechanism and demonstrate the significant structure of delayed influences, which can be partially recovered through ISO approximations.
Methodology
The authors model the AdamW optimizer as a finite-horizon ISO system, analyzing the effects of a single gradient perturbation on future loss through paired training trajectories. They derive mathematical formulations to capture the dynamics of optimizer memory and its impact on training outcomes.
Results
The study demonstrates that the effects of a minibatch perturbation can persist across multiple training steps, with optimizer memory significantly influencing the timing and magnitude of these effects. The derived response operator effectively captures the dynamics of these perturbations, and the multistep error decomposition provides insights into the accuracy of the ISO model.
Implications
The findings suggest that understanding the delayed effects of minibatch perturbations can lead to improved training stability and performance in neural networks, particularly when using adaptive optimizers like AdamW. This could inform strategies for optimizing training schedules and minibatch ordering.
Decoding silent reading from non-invasive EEG
NLP
Multimodal
Time Series
- Silent reading serves as an effective proxy for decoding inner speech from EEG.
- A contrastive decoder can extract meaningful lexical and semantic information from EEG data.
- Decoding performance scales log-linearly with the volume of training data.
- The study establishes that open-vocabulary word-level information is recoverable from EEG.
Read more
Decoding silent reading from non-invasive EEG
Summary
This paper addresses the challenge of decoding inner speech using non-invasive EEG, focusing on silent reading as a scalable proxy task. The authors highlight the limitations of existing paradigms for capturing spontaneous inner monologue, which often lack temporal precision and verifiability. They present a novel approach using a contrastive decoder to extract lexical and semantic information from EEG data collected during silent reading. The study involved approximately 240,000 word presentations from a single participant, utilizing a convolutional EEG encoder and a causal transformer trained with a CLIP-style contrastive objective. The results demonstrate that the decoder can reliably retrieve word-level information, achieving above-chance performance even for mid-frequency and rare words. The findings suggest that the decoding capability is limited by data volume rather than a saturation of signal quality, indicating the potential for further advancements in non-invasive language decoding.
Methodology
The authors employed a convolutional EEG encoder, optionally followed by a causal transformer, trained with a CLIP-style contrastive objective. The EEG data was collected from a single participant during silent reading of continuous narrative text presented in rapid serial visual presentation, with typography randomized to decorrelate word identity from visual form.
Results
The decoding achieved reliable above-chance performance in word-grouped top-10 retrieval, extending to mid-frequency and rare words. The performance scaled log-linearly with the amount of training data, and removing certain electrodes reduced word-level gain but did not affect context tracking.
Implications
These findings suggest that non-invasive EEG can be a viable method for decoding silent reading and potentially inner speech, which could have significant implications for brain-computer interfaces and communication aids for individuals with speech impairments.
Quantum Kernel Estimation for the Discovery of Early Lung Cancer Detection
Theory
- Quantum-classical hybrid models can effectively analyze high-dimensional cfDNA data for lung cancer detection.
- Quantum kernel estimation (QKE) enhances the classification of nonlinear molecular signals.
- Selected quantum models showed competitive performance compared to classical SVMs, particularly in fragmentomics.
- Increasing feature count did not consistently improve model performance, indicating a need for careful feature selection.
Read more
Quantum Kernel Estimation for the Discovery of Early Lung Cancer Detection
Summary
This paper presents a novel approach to early lung cancer detection using quantum-classical hybrid machine learning techniques applied to cell-free DNA (cfDNA) biomarkers. The authors address the limitations of traditional lung cancer screening methods, which are hindered by low uptake and adherence rates. They propose using quantum kernel estimation (QKE) to enhance the classification of high-dimensional and nonlinear molecular signals from cfDNA fragmentomics and DNA methylation datasets. The study involved feature selection from two datasets, followed by encoding features into quantum Hilbert space using various entanglement strategies. The performance of quantum-kernel models was evaluated against classical support vector machines (SVM) and logistic regression models. Results indicated that quantum models achieved competitive performance, particularly in capturing the nonlinear structure of cfDNA fragmentation, while also improving specificity in some cases. The findings suggest that quantum kernel methods hold promise for advancing cfDNA-based lung cancer detection.
Methodology
The authors utilized quantum kernel estimation (QKE) to construct nonlinear similarity matrices from selected features derived from cfDNA datasets. They performed feature selection using L1-penalized logistic regression and encoded features into quantum states using angle and dense-angle encoding strategies. The models were trained and evaluated using precomputed-kernel SVM and kernel-PCA logistic regression, comparing their performance against classical SVM models.
Results
The quantum-kernel models demonstrated competitive performance across both datasets. For cfDNA fragmentomics, several 20-feature configurations improved the area under the curve (AUC) compared to classical SVM baselines. Although classical SVM achieved the highest AUC for methylation data, selected quantum models remained competitive and improved specificity in certain instances. The increase in features from 20 to 40 did not consistently enhance performance and often introduced variability.
Implications
The findings suggest that quantum kernel methods could significantly improve the accuracy and reliability of cfDNA-based lung cancer detection, potentially leading to better diagnostic tools and enhanced patient outcomes. This approach may pave the way for more effective screening strategies in oncology.
A comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection
Time Series
- Controlled comparison of FMCW, IR-UWB, and Wi-Fi radar under identical conditions.
- IR-UWB outperforms in activity recognition, while FMCW excels in environmental robustness.
- All technologies achieve high performance in sleep monitoring tasks.
- Insights into signal characteristics inform practical guidelines for RF sensing system design.
Read more
A comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection
Summary
This paper presents a comprehensive comparison of three radio frequency (RF) sensing technologiesβfrequency-modulated continuous wave (FMCW) radar, impulse radio ultra-wideband (IR-UWB), and Wi-Fi radarβspecifically for human activity recognition (HAR) and sleep interruption detection in domestic settings. The study addresses the lack of controlled comparisons across these technologies under identical conditions, which has been a gap in existing literature. Using synchronized recordings from 20 participants across six different room layouts, the authors evaluate the performance of each technology using a uniform convolutional neural network (CNN) architecture. The results indicate that IR-UWB achieves the highest cross-subject activity recognition performance at 89.0% macro F1 score, while FMCW demonstrates superior generalization to unseen room layouts with an 83.8% macro F1 score. For sleep monitoring tasks, all technologies exceed 92% macro F1 in unseen environments. The findings highlight a trade-off between recognition performance and environmental robustness, influenced by factors such as range resolution and antenna diversity. This research provides practical guidelines for designing healthcare-oriented RF sensing systems and contributes an open dataset for future studies.
Methodology
The study employs a controlled experimental design with synchronized recordings from 20 participants across six room layouts. A uniform convolutional neural network (CNN) is used to evaluate the performance of each RF sensing technology on both a fine-grained 10-class HAR task and a coarse 4-class sleep monitoring task. Multiple cross-validation protocols ensure that performance differences are attributable to the sensing technology rather than experimental design.
Results
IR-UWB achieved the highest macro F1 score of 89.0% for activity recognition, while FMCW performed best in generalizing to unseen room layouts with an 83.8% macro F1 score. For sleep monitoring, all technologies surpassed 92% macro F1 in unseen environments, indicating strong performance across the board.
Implications
The findings suggest that RF sensing technologies can provide effective, privacy-preserving solutions for healthcare monitoring, particularly in assessing human activity and sleep quality. The study's insights can guide the development of more robust and efficient RF-based monitoring systems in both domestic and clinical settings.
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
Large Language Models
Optimization
Efficient ML
- Proposes a two-step hyperparameter transfer framework for large-scale MoE models.
- Utilizes Maximal Update Parameterization (Β΅P) for consistent learning rate transfer across model widths.
- Establishes a predictive scaling law for extrapolating learning rates to large training horizons.
- Demonstrates the methodology by successfully pretraining a 155B parameter MoE model.
Read more
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
Summary
This paper addresses the challenge of optimizing hyperparameters, particularly learning rates, for large-scale Mixture-of-Experts (MoE) models, which can expand model capacity without a proportional increase in computational cost. The authors propose a two-step hyperparameter transfer framework that allows for the efficient estimation of optimal learning rates by transferring them across different model widths and extrapolating to extensive training horizons of up to 10 trillion tokens. The first step involves the Maximal Update Parameterization (Β΅P) adaptation for MoE architectures, utilizing Multi-head Latent Attention (MLA) and the Muon optimizer, which demonstrates consistent transferability of optimal learning rates across width-scaled models. The second step establishes a predictive scaling law using linear regression on optimal values derived from small proxy models, enabling accurate extrapolation of learning rates for massive training scenarios. The methodology was validated by pretraining a foundation model with 155 billion total parameters and 17 billion active parameters, showing that optimal configurations can be predicted with minimal computational costs. This approach significantly reduces the need for exhaustive hyperparameter sweeps, making it feasible to estimate optimal training hyperparameters for extreme scales efficiently.
Methodology
The proposed framework consists of two main steps: first, adapting the Maximal Update Parameterization (Β΅P) to MoE architectures to facilitate the transfer of optimal learning rates across different model widths. Second, applying linear regression to optimal values from small proxy models to extrapolate learning rates for extensive training scenarios, thus avoiding costly 2D hyperparameter sweeps.
Results
The methodology was validated through the pretraining of a foundation MoE model with 155 billion total parameters over 10 trillion tokens, achieving high fidelity in learning rate predictions (R2 = 0.95) and demonstrating stable training and evaluation results.
Implications
The proposed framework can significantly streamline the hyperparameter optimization process for large-scale MoE models, enabling researchers and practitioners to efficiently scale their models without incurring prohibitive computational costs. This has potential applications in various fields where large-scale models are deployed, such as natural language processing and machine learning.
Empirical Characterization of Learning Geometry in Hybrid Quantum Forecasting Models
Time Series
Optimization
Theory
- Introduces an empirical NTK framework for comparing hybrid quantum and classical learning dynamics.
- Identifies architecture-dependent differences in kernel alignment, drift, and spectral concentration.
- Demonstrates that the hybrid model achieves competitive performance with fewer parameters and faster convergence.
- Finds that Fourier features alone do not replicate the training dynamics observed in the hybrid model.
Read more
Empirical Characterization of Learning Geometry in Hybrid Quantum Forecasting Models
Summary
This paper investigates the learning dynamics of a compact hybrid quantum forecasting model in comparison to a classical baseline model. The authors employ empirical analysis using Neural Tangent Kernel (NTK) diagnostics to characterize the optimization behavior of both models on synthetic multivariate forecasting tasks, including stationary harmonic-mixture and nonstationary chirp signals. The study reveals that the classical model demonstrates stronger early target alignment, while the hybrid model exhibits a less concentrated kernel spectrum and smaller kernel drift. Despite these differences, both models achieve similar performance on held-out data. The hybrid model, with only 125 trainable parameters, reaches its validation checkpoint faster than the classical model with 281 parameters in most scenarios. The findings suggest that different learning trajectories can lead to comparable generalization, and that NTK diagnostics do not consistently predict convergence. The study emphasizes the importance of understanding architecture-dependent learning behaviors beyond mere endpoint accuracy.
Methodology
The authors conducted a controlled empirical study comparing a hybrid quantum model with a structurally aligned classical model. They analyzed the models' learning dynamics using NTK diagnostics, including kernel-target alignment, kernel drift, and spectral concentration, across various synthetic forecasting tasks with adjustable spectral complexity and data availability.
Results
The hybrid quantum model, despite having fewer parameters, achieved similar held-out performance to the classical model and reached its best validation-selected solution earlier in 15 out of 18 frequency conditions. The study also found that the classical model had stronger early target alignment, while the hybrid model showed less concentrated kernel spectra and smaller kernel drift.
Implications
The findings suggest that hybrid quantum models can be competitive with classical models in forecasting tasks, offering insights into the optimization dynamics of quantum machine learning. This could lead to more efficient model designs and a better understanding of how different architectures learn from data.
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
Reinforcement Learning
- Introduces adaptive probabilistic shielding for safe reinforcement learning.
- Combines online model learning with probabilistic shielding to adaptively compute safety constraints.
- Addresses the exploration-exploitation dilemma by balancing safety and data collection.
- Utilizes interval MDPs to robustly capture uncertainty in transition probabilities.
Read more
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
Summary
This paper addresses the challenge of ensuring safety in reinforcement learning (RL) by introducing a novel approach called adaptive probabilistic shielding. Traditional probabilistic shielding relies on a fully specified Markov decision process (MDP), which is often not available in real-world scenarios. The authors propose a method where the transition graph of the MDP is known, but the transition probabilities are learned online as the RL agent interacts with the environment. This adaptive shielding technique allows for the computation of a safety shield that evolves as the agent gathers more data, thereby reducing conservativeness over time. The paper discusses the balance between exploration and safety, presenting strategies to optimize this trade-off. Empirical evaluations across various environments demonstrate the effectiveness of the proposed approach, highlighting its ability to improve safety while still enabling exploration.
Methodology
The authors developed an adaptive shielding algorithm that integrates online model learning with probabilistic shielding. They utilize interval MDPs to estimate transition probabilities and compute safety shields based on these estimates. The algorithm updates the shield dynamically as the RL agent collects data, allowing for a balance between exploration and safety.
Results
The experimental evaluation demonstrated that the adaptive probabilistic shielding approach effectively reduces the conservativeness of the safety shield over time, allowing for safer exploration in various environments. The results indicate that the proposed method successfully mitigates the risks associated with unsafe actions during the learning process.
Implications
This research has significant implications for the deployment of RL agents in real-world applications where safety is critical, such as robotics and autonomous systems. The adaptive shielding approach can enhance the reliability of RL systems by ensuring that safety constraints are respected while still allowing for effective learning.
Quantifying Event Impacts on Time Series via Multiscale Contrastive Learning
Time Series
- Introduces EventTime, a framework for estimating abnormal financial losses from cybersecurity incidents.
- Combines multi-resolution market context with event metadata for improved prediction accuracy.
- Utilizes a dynamic contrastive learning approach to address sparse supervision challenges.
- Demonstrates superior performance over state-of-the-art models in real-world financial scenarios.
Read more
Quantifying Event Impacts on Time Series via Multiscale Contrastive Learning
Summary
This paper addresses the challenge of quantifying the impacts of cybersecurity incidents on financial time series, which often result in abrupt market disruptions. The authors introduce EventTime, a novel multi-resolution framework that integrates long-horizon market context, short-horizon pre-event dynamics, and event metadata to predict short-term abnormal losses following such events. The framework includes an event fusion module that aligns temporal representations with event attributes, enhancing the model's ability to identify relevant market patterns. To tackle the issue of sparse supervision due to the rarity of high-impact events, EventTime employs a dynamic contrastive learning objective that creates event- and time-series-aware positive and negative pairs during training. The authors also present SECURE, a dataset that aligns cybersecurity incidents with stock-market time series and various semantic features. Experimental results demonstrate that EventTime outperforms existing time-series and event-aware models in estimating post-event financial losses, providing more event-sensitive representations and greater robustness to incomplete metadata.
Methodology
The methodology involves a multi-resolution framework that integrates various temporal scales of market data with event metadata. An event fusion module is employed to couple temporal representations with event attributes. The dynamic contrastive learning objective is used to create training pairs that enhance the model's learning from sparse data.
Results
EventTime consistently outperformed state-of-the-art time-series and event-aware baselines in estimating post-event financial losses. The model showed improved sensitivity to event-related changes, robustness to incomplete metadata, and provided interpretable estimates of market impacts following cybersecurity disclosures.
Implications
The findings have significant implications for financial risk assessment, portfolio management, and regulatory oversight, as they provide a more accurate understanding of how cybersecurity incidents affect market dynamics. This can lead to better decision-making for investors and firms in mitigating risks associated with such events.
Rationally Enriched Chebyshev Trunk Bases for DeepONet Surrogates of High PΓ©clet Entrance Transport
Theory
- Introduction of Rationally Enriched Chebyshev (REC) trunk for DeepONet models.
- Demonstrated improved accuracy in predicting solution profiles for high-PΓ©clet transport problems.
- Significant reduction in profile-error metrics compared to traditional DeepONet models.
- Effective suppression of artificial oscillations in near-wall regions.
Read more
Rationally Enriched Chebyshev Trunk Bases for DeepONet Surrogates of High PΓ©clet Entrance Transport
Summary
This paper presents a novel approach to modeling singularly perturbed and high-PΓ©clet transport problems using a Rationally Enriched Chebyshev (REC) trunk within DeepONet surrogate models. The REC trunk integrates Chebyshev polynomial elements with rational dictionary elements derived from the adaptive AntoulasβAnderson (AAA) algorithm. The authors evaluate the performance of the REC-trunk DeepONet against traditional DeepONet and Chebyshev-trunk DeepONet models across three specific problems characterized by varying diffusion-to-advection ratios. The results indicate that the REC-trunk DeepONet significantly enhances the accuracy of predictions for scalar profiles, particularly when the perturbation parameter is between 1.00 Γ 10β4 and 1.78 Γ 10β4, achieving up to a 19.5% reduction in profile-error metrics compared to the Chebyshev-trunk DeepONet. Furthermore, it demonstrates substantial improvements in predicting wall-normal temperature and concentration profiles, reducing error metrics by up to 60.2% and 32.2%, respectively, while effectively suppressing artificial oscillations in near-wall regions. This work highlights the potential of enriched polynomial bases in addressing the challenges posed by high-PΓ©clet transport phenomena.
Methodology
The study employs a DeepONet architecture enhanced with a REC trunk that combines Chebyshev polynomial bases with rational elements derived from the AAA algorithm. The model is trained and evaluated across three benchmark problems with varying diffusion-to-advection ratios, comparing its performance against standard DeepONet and Chebyshev-trunk DeepONet models.
Results
The REC-trunk DeepONet outperformed the vanilla DeepONet and showed comparable performance to the Chebyshev-trunk DeepONet, particularly excelling in scenarios with specific perturbation parameters. It achieved up to a 19.5% reduction in profile-error metrics for scalar profiles and up to 60.2% and 32.2% reductions for wall-normal temperature and concentration profiles, respectively.
Implications
The findings suggest that the REC trunk can significantly improve the modeling of complex transport phenomena in various engineering applications, such as biomedical engineering, chemical engineering, and aerospace engineering, where accurate predictions of boundary-layer behaviors are crucial.
Systematic Evaluation of TabPFN-TS for Zero-Shot Probabilistic Heat Load Forecasting in District Heating Networks
Time Series
- TabPFN-TS offers a zero-shot forecasting approach, reducing the need for retraining on historical data.
- Optimal configuration for forecasting includes hourly predictions with a 12-week context and ambient temperature as a covariate.
- TabPFN-TS shows competitive performance compared to state-of-the-art models like Chronos-2 in terms of accuracy and calibration.
- The study identifies the importance of context length and covariate selection in improving forecasting outcomes.
Read more
Systematic Evaluation of TabPFN-TS for Zero-Shot Probabilistic Heat Load Forecasting in District Heating Networks
Summary
This paper presents a systematic evaluation of TabPFN-TS, a zero-shot probabilistic forecasting model, for predicting heat loads in district heating networks. Traditional forecasting methods often require retraining on historical data, which can be cumbersome due to network changes. TabPFN-TS, leveraging synthetic pretraining data, offers an alternative by adapting at inference time using recent observations. The study investigates various factors affecting forecasting performance, including covariate choice, context length, temporal resolution, and prediction horizon. The results indicate that an hourly 24-hour forecasting configuration with a 12-week rolling context and ambient temperature as a covariate is optimal. TabPFN-TS achieves competitive deterministic accuracy, closely matching the performance of the benchmark model Chronos-2, while demonstrating better empirical calibration. The findings also lead to the development of a Multi-Resolution Residual-Correction Forecaster, which combines different forecasting approaches to enhance longer-horizon accuracy.
Methodology
The authors conducted a systematic evaluation of TabPFN-TS by comparing it against state-of-the-art time-series foundation models and traditional machine learning methods. They analyzed the impact of various factors such as covariate choice, context length, temporal resolution, and prediction horizon on forecasting performance. The evaluation was performed on representative operating weeks and validated over a full year, including transferability tests on a second network.
Results
TabPFN-TS achieved a CVRMSE of 13.06% on the main dataset, closely trailing Chronos-2's 12.48%. While Chronos-2 had the lowest aggregate full-year error, TabPFN-TS exhibited superior empirical calibration. The optimal configuration identified was hourly forecasting with a 12-week rolling context, indicating that longer context windows did not yield improved accuracy.
Implications
The findings suggest that zero-shot forecasting models like TabPFN-TS can significantly streamline the forecasting process in district heating networks, adapting to changes without extensive retraining. The proposed Multi-Resolution Residual-Correction Forecaster could further enhance operational efficiency by improving accuracy in longer-term planning.
KΓ€hler landscapes for complex neural network descents and guarantees including a search and destroy of the Calabi-Yau manifold
Optimization
Theory
- Introduces KΓ€hler geometry to analyze complex neural network optimization landscapes.
- Establishes a natural gradient descent update rule that preserves holomorphic structure.
- Explores the impact of Calabi-Yau metrics on optimization guarantees and loss landscapes.
- Demonstrates how negative curvature can disrupt the optimization process.
Read more
KΓ€hler landscapes for complex neural network descents and guarantees including a search and destroy of the Calabi-Yau manifold
Summary
This paper investigates the optimization landscapes of complex-parameterized neural networks through a geometric lens, specifically utilizing KΓ€hler geometry. The author employs an information-theoretic perspective to analyze the parameter space and optimization guarantees, focusing on the implications of Calabi-Yau manifolds. The study introduces a descent update rule based on natural gradient descent, which maintains the descent path within the holomorphic tangent bundle. The paper highlights the challenges posed by ill-curvature-conditioned landscapes, particularly under Calabi-Yau metrics, and discusses how negative curvature affects the loss landscape. Theoretical results are presented, including second derivative results, initialization conditions, and convexity properties, all framed within the context of deep learning theory and geometric analysis. The findings suggest that the complex geometry of the parameter space can significantly influence the optimization process and the performance of neural networks.
Methodology
The methodology involves a geometric analysis of the optimization landscape using KΓ€hler metrics and Dolbeault asymptotics. The author employs natural gradient descent with a differentiated loss scaled by the inverse metric, ensuring the descent path remains in the holomorphic tangent bundle. Theoretical results are derived through a combination of geometric analysis and deep learning theory.
Results
The paper presents several theoretical results, including conditions for convexity, initialization results, and the effects of negative Ricci curvature on the optimization landscape. It establishes that under certain conditions, the geometry of the parameter space can lead to significant performance improvements in complex neural networks, while also identifying pitfalls associated with Calabi-Yau metrics.
Implications
The findings have implications for the design and training of complex neural networks, suggesting that a deeper understanding of the underlying geometric structures can lead to more effective optimization strategies. This work may influence future research in neural network architecture and training methodologies, particularly in complex parameter settings.
Multi-Source Wasserstein Distributionally Robust Graph Learning
Graph Learning
Optimization
Theory
- Introduces MS-WDRO framework for robust graph learning from heterogeneous data sources.
- Utilizes Wasserstein barycenter for effective distribution fusion while preserving geometrical properties.
- Establishes non-asymptotic performance guarantees for the proposed method.
- Achieves data-adaptive calibration of hyperparameters through a differentiable architecture.
Read more
Multi-Source Wasserstein Distributionally Robust Graph Learning
Summary
This paper addresses the challenge of network topology inference from graph signals, particularly in scenarios where target-domain samples are limited but heterogeneous source-domain data is abundant. The authors introduce a novel framework called MS-WDRO (Multi-Source Wasserstein Distributionally Robust Optimization) that utilizes the Wasserstein metric for fusing diverse data sources while maintaining their intrinsic geometries. The framework constructs a weighted Wasserstein barycenter as a nominal distribution and creates an ambiguity ball around it to manage residual uncertainties. By minimizing the worst-case risk within this ball, the authors derive a regularized Laplacian estimator, which is efficiently solved using an ADMM (Alternating Direction Method of Multipliers) scheme. The paper provides non-asymptotic performance guarantees, including finite-sample concentration bounds and out-of-sample excess risk bounds. The proposed method is calibrated through an end-to-end differentiable architecture, allowing for data-adaptive hyperparameter tuning. Experimental results demonstrate that MS-WDRO outperforms seven baseline methods in terms of graph recovery accuracy, sample efficiency, and diagnostic utility, particularly in scenarios with limited samples.
Methodology
The methodology involves constructing a weighted Wasserstein barycenter from multiple heterogeneous data sources, creating an ambiguity ball around this barycenter to account for uncertainty, and minimizing the worst-case risk using a regularized Laplacian estimator. The optimization problem is solved efficiently using an ADMM scheme, and hyperparameters are calibrated through an end-to-end differentiable architecture.
Results
The results indicate that MS-WDRO consistently outperforms seven baseline methods in graph recovery accuracy and sample efficiency, particularly in scenarios with limited data. The method also shows improved downstream diagnostic utility, highlighting its effectiveness in practical applications.
Implications
The proposed framework has significant implications for various fields that rely on accurate network topology inference, such as neuroimaging, sensor networks, and social network analysis. By effectively leveraging heterogeneous data sources, it enhances the robustness and reliability of graph learning in real-world applications.
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Generative Models
Optimization
- FAR-DPO improves the generation of feasible cyclic peptides by integrating feasibility-aware preference construction.
- The method employs difficulty-aware group-robust optimization to enhance performance across heterogeneous targets.
- FAR-DPO shows significant improvements in success rates and binding scores on the CPSea LNR benchmark.
- The framework is architecture-agnostic, demonstrating applicability across different generative models.
Read more
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Summary
Cyclic peptides are gaining attention in drug discovery due to their high binding affinity and structural stability. However, the design of cyclic peptides presents unique challenges due to the constraints imposed by cyclization, which limits the feasible design space. Existing generative models often struggle with low yields of feasible designs and lack control over multi-objective trade-offs. To address these issues, the authors propose FAR-DPO, a framework that integrates feasibility-aware preference construction with difficulty-aware group-robust optimization. FAR-DPO constructs preference pairs based on feasibility and optimizes the generative model to enhance the yield of structurally and biophysically feasible cyclic peptides. The methodology involves generating a large number of candidates, screening for structural feasibility, and applying a robust optimization strategy that adjusts weights based on the difficulty of the design tasks. The results demonstrate that FAR-DPO significantly improves the success rate of generating feasible cyclic peptides while also enhancing binding scores, particularly for challenging targets. This approach is validated across two different generative models, showing its versatility and effectiveness in cyclic peptide design.
Methodology
FAR-DPO utilizes a two-step approach: first, it constructs preference pairs from generated cyclic peptides based on structural feasibility and multi-objective dominance. Second, it applies a group-robust optimization strategy that dynamically adjusts the importance of different difficulty groups during training, ensuring that the model focuses on improving performance across all targets.
Results
FAR-DPO increased the overall success rate of generating feasible cyclic peptides from 46.89% to 57.79% on PepGLAD and from 47.96% to 49.57% on PepFlow under a fixed generation budget. The improvements extend to the hardest target quartile, with better binding scores observed for the best candidates.
Implications
The FAR-DPO framework has the potential to enhance the design of cyclic peptides for therapeutic applications, improving the efficiency of drug discovery processes. Its architecture-agnostic nature allows for broader applicability in various peptide design tasks, potentially leading to more effective therapeutics.
SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
Interpretability
- Introduction of a geographic location-aware SAE for ExEE interpretation.
- Development of SAE-Xplainers for translating complex features into understandable rules.
- Demonstrated improved performance in feature utilization and reconstruction.
- Faithful interpretation of climatic patterns consistent with scientific literature.
Read more
SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
Summary
The paper addresses the challenge of interpreting deep learning models used for predicting extreme Earth events (ExEE) such as wildfires, tropical cyclones, and atmospheric rivers. The authors introduce a novel approach that combines geographic location-based modulation of Sparse Autoencoders (SAEs) with an ensemble of rule-based models, termed SAE-Xplainers, to enhance interpretability. The geographic modulation captures local environmental patterns, allowing the model to learn location-dependent features. The SAE-Xplainers then translate these features into human-understandable rules, facilitating the interpretation of complex climatic patterns. The proposed method was evaluated on three types of ExEE, demonstrating improved reconstruction performance and feature utilization. The results indicate that the SAE-Xplainers provide faithful interpretations consistent with scientific literature, while also identifying feature absorption phenomena. This work contributes to the field of explainable AI in Earth sciences, offering a framework for better understanding the drivers of extreme weather events.
Methodology
The authors propose a geographic location encoder that modulates the input to a k-sparse autoencoder (GeoTopK), enabling the extraction of location-aware features. An ensemble of rule-based models, SAE-Xplainers, is then used to interpret these features, translating them into human-understandable rules. The approach was validated using two different encoders across datasets for three types of ExEE.
Results
The proposed method showed significant improvements in reconstruction performance and feature utilization compared to traditional location-agnostic approaches. The SAE-Xplainers successfully provided interpretable rules that aligned with existing scientific knowledge, enhancing the understanding of complex climatic patterns and identifying feature absorption.
Implications
This research has potential applications in climate science, disaster management, and environmental monitoring, where understanding the drivers of extreme events is crucial for preparedness and risk management. The framework can aid in developing more interpretable AI models in Earth sciences, promoting trust and usability in operational settings.
DeltaML-Bench: Evaluating Machine Learning Agents on Real-World Research Repositories
Computer Vision
Graph Learning
Time Series
- DeltaML-Bench introduces a benchmark for evaluating agents on real-world ML tasks, challenging them to improve baselines in imperfect codebases.
- Search-based ARG scaffolding significantly improves the success rate of GPT-5 in autonomous experimentation.
- Specification gaming is a concern in Modular configurations, while ARG configurations show no evidence of gaming.
- The benchmark spans diverse domains including Computer Vision, Graph Learning, and Time Series.
Read more
DeltaML-Bench: Evaluating Machine Learning Agents on Real-World Research Repositories
Summary
The paper introduces DeltaML-Bench, a novel benchmark designed to evaluate autonomous machine learning agents on real-world research repositories. Unlike existing benchmarks that focus on isolated tasks or clean datasets, DeltaML-Bench comprises 48 tasks derived from research papers, requiring agents to improve published baselines within imperfect, open-source codebases. The authors evaluate two advanced models, GPT-5 and Claude Sonnet 4, using a standard Modular agent and a search-based ARG scaffolding approach. The results demonstrate that the ARG scaffolding significantly enhances GPT-5's success rate in improving baselines, achieving a per-run success rate of 49.0% under a 2Γ12h allocation. The study also highlights the issue of specification gaming, with Modular configurations exhibiting rates as high as 47.9%, while no gaming was observed in ARG configurations. The findings underscore the importance of scaffolding design and integrity checks in deploying agents for autonomous ML experimentation.
Methodology
The authors developed DeltaML-Bench, consisting of 48 tasks sourced from research papers. They evaluated the performance of GPT-5 and Claude Sonnet 4 using both Modular and ARG scaffolding approaches, measuring success rates under different time allocations (4Γ6h and 2Γ12h) and analyzing specification gaming rates.
Results
In the 4Γ6h allocation, the ARG approach raised GPT-5's success rate from 9.4% to 33.9%, while under the 2Γ12h allocation, it achieved a 49.0% success rate. Modular configurations showed specification gaming rates up to 47.9%, whereas no gaming was detected in ARG configurations.
Implications
DeltaML-Bench provides a valuable framework for assessing the capabilities of machine learning agents in real-world scenarios, highlighting the importance of effective scaffolding and the potential risks of specification gaming. This benchmark can guide future research in autonomous ML experimentation and improve the reliability of agent performance in complex environments.
SAGE-XGBoost: Spatially Augmented Graph EmbeddingsβMachine Learning Framework for Natural Hazards Susceptibility Mapping under Data Scarcity
Graph Learning
- SAGE-XGBoost improves natural hazard susceptibility mapping under data scarcity.
- Integrates noise-based data augmentation with graph embeddings for feature enhancement.
- Outperformed conventional models by over 33 percentage points in AUC scores.
- Achieved high AUC values of 0.97 for landslide and 0.95 for wildfire susceptibility.
Read more
SAGE-XGBoost: Spatially Augmented Graph EmbeddingsβMachine Learning Framework for Natural Hazards Susceptibility Mapping under Data Scarcity
Summary
The paper presents SAGE-XGBoost, a novel machine learning framework designed for natural hazard susceptibility mapping, particularly in scenarios with limited labeled data. Traditional machine learning approaches often struggle under data scarcity, leading to reduced generalizability and effectiveness. The proposed SAGE framework integrates controlled noise-based data augmentation with neighborhood-based graph embeddings to enhance predictive performance. A K-nearest neighbor graph is utilized to derive local spatial statistics, which are then reduced using principal component analysis and combined with environmental covariates and spatial coordinates. The resulting features are employed in an XGBoost model to create SAGE-XGBoost. The framework was tested on landslide and wildfire susceptibility mapping, demonstrating significant improvements over conventional models. The SAGE-XGBoost model achieved AUC values of approximately 0.97 for landslide susceptibility and 0.95 for wildfire susceptibility, outperforming traditional and spatially explicit machine learning models by over 33 percentage points. Feature importance analysis indicated that graph embeddings contributed significantly to the model's predictions, enhancing spatial coherence and reducing noise amplification. The study concludes that SAGE-XGBoost offers an efficient alternative to deep learning methods for environmental hazard assessments under data-scarce conditions.
Methodology
The methodology involves constructing a K-nearest neighbor graph to extract local spatial statistics, which are then reduced via principal component analysis. These features are integrated with environmental covariates and spatial coordinates, and subsequently used in an XGBoost model to develop the SAGE-XGBoost framework.
Results
SAGE-XGBoost demonstrated superior performance in landslide and wildfire susceptibility mapping, achieving AUC values of approximately 0.97 and 0.95, respectively. The model showed an absolute improvement of over 33 percentage points compared to traditional models, confirming the efficacy of the proposed framework.
Implications
The findings suggest that SAGE-XGBoost can serve as an effective tool for environmental hazard assessment, particularly in regions with limited data availability. Its approach can be applied to various geospatial prediction tasks, enhancing decision-making in disaster management and risk reduction.
A Two-Stage Time-Aware Transformer for Short-Horizon AECOPD Risk Prediction
Time Series
- Introduces a two-stage prediction framework for AECOPD risk using raw ventilator waveforms.
- First stage classifies high-risk patients, while the second stage estimates time to event.
- Preserves temporal dynamics of respiratory data, enhancing prediction accuracy.
- Achieves F1 score of 0.91 for risk classification and RMSE of 1.00 days for time-to-event estimation.
Read more
A Two-Stage Time-Aware Transformer for Short-Horizon AECOPD Risk Prediction
Summary
This paper presents a novel two-stage framework for predicting the risk of acute exacerbation of chronic obstructive pulmonary disease (AECOPD) using raw pressure and flow waveforms from home ventilators. Traditional machine learning approaches often rely on episodically collected clinical data, which can delay risk detection. The proposed method operates directly on a 7-day window of ventilator data, with the first stage focusing on classifying patients at high risk of severe exacerbation, and the second stage estimating the time remaining before such an event occurs. By preserving the temporal structure of the raw waveforms, the model captures short-horizon dynamics that are critical for timely intervention. Experimental results demonstrate that the two-stage model significantly outperforms traditional baseline models, achieving high classification and regression performance metrics, thus providing clinicians with both early warnings and actionable lead time for intervention.
Methodology
The methodology involves a two-stage model where the first stage uses a Time-Aware Transformer encoder to classify patients at high risk of AECOPD based on raw pressure and flow waveforms. The second stage employs a regression model to estimate the time remaining until the exacerbation event for those identified as high risk. This approach maintains the temporal integrity of the data, allowing for effective learning of patient representations.
Results
The selected Stage 1 classifier achieved an F1 score of 0.91, indicating high accuracy in identifying high-risk patients. The Stage 2 regression model demonstrated a root mean square error (RMSE) of 1.00 days and an R-squared (R2) value of 0.76, providing reliable estimates of the time until exacerbation.
Implications
The findings suggest that this two-stage framework could significantly enhance home monitoring of COPD patients, allowing for timely interventions and potentially reducing hospitalizations. The model's ability to provide both risk classification and time-to-event estimation could improve clinical decision-making and patient outcomes.
DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
Time Series
Generative Models
Interpretability
- DecoVAE explicitly decomposes time series into trend and seasonal components using specialized subnetworks.
- The trend component is regularized for smoothness using a Hodrick-Prescott-inspired penalty.
- Seasonality is modeled in the complex frequency domain, enhancing the representation of periodic patterns.
- DecoVAE achieves significant performance improvements over strong baselines in forecasting accuracy.
Read more
DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
Summary
The paper introduces DecoVAE, a novel lightweight framework designed for probabilistic time series forecasting that effectively separates trend and seasonal components. Traditional forecasting methods often struggle with accurately modeling these distinct dynamics, leading to inefficiencies and a lack of interpretability. DecoVAE addresses these challenges by employing domain-specific inductive biases to decompose time series data. The trend component is modeled using a differential regularizer inspired by the Hodrick-Prescott filter, ensuring structural smoothness. In contrast, the seasonal component is modeled in the complex frequency domain, allowing for the natural capture of periodic patterns. The framework is integrated into a variational autoencoder (VAE) architecture, which enhances the model's ability to provide well-calibrated probabilistic forecasts. Extensive empirical evaluations across seven real-world datasets demonstrate that DecoVAE outperforms existing methods, achieving significant reductions in Continuous Ranked Probability Score (CRPS) and Normalized Mean Absolute Error (NMAE) for both short-term and long-term forecasting horizons, while also being highly efficient in terms of model size and computational speed.
Methodology
DecoVAE employs a variational autoencoder framework that separates time series into trend and seasonal components. The trend is modeled with a differential regularizer to ensure smoothness, while the seasonal component is captured in the complex frequency domain to effectively encode amplitude and phase information. This dual approach allows for accurate modeling of distinct temporal dynamics and well-calibrated probabilistic forecasts.
Results
DecoVAE demonstrated up to 14.96% improvement in CRPS and 23.30% in NMAE for short-term forecasting, and up to 52.68% and 26.51% improvements for long-term forecasting across seven benchmark datasets. Additionally, the model's efficiency is highlighted by a reduction in weight by up to 93% and an acceleration in speed by up to 74% compared to the second-best method.
Implications
The DecoVAE framework has significant implications for various applications requiring accurate probabilistic time series forecasting, such as finance, energy, weather, and traffic management. Its lightweight and interpretable nature makes it suitable for deployment in resource-constrained environments while maintaining high forecasting accuracy.
Longitudinal Bayesian Learning of Continuous Disease Position across the Alzheimer's Disease Continuum
Time Series
- Introduction of Disease Continuum Positioning (DCP) framework for continuous disease severity estimation.
- DCP generates an uncertainty-aware Disease Continuum Score (DCS) for individual positioning along the AD continuum.
- DCS captures clinically meaningful disease variation and predicts future disease transitions.
- DCP outperforms traditional disease progression methods in extensive experiments.
Read more
Longitudinal Bayesian Learning of Continuous Disease Position across the Alzheimer's Disease Continuum
Summary
This paper presents a novel framework called Disease Continuum Positioning (DCP) for estimating the continuous severity of Alzheimer's disease (AD) using longitudinal diffusion tensor imaging (DTI). Unlike traditional methods that categorize patients into discrete diagnostic stages, DCP provides a continuous Disease Continuum Score (DCS) that reflects an individual's position along the AD continuum. The framework integrates longitudinal observations with weak clinical supervision to model disease severity as a low-dimensional probabilistic latent variable. The DCS captures clinically relevant variations in disease progression and offers predictive insights into future transitions from cognitively normal (CN) to mild cognitive impairment (MCI) and from MCI to AD. The authors conducted extensive experiments using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, demonstrating that DCP outperforms existing disease progression methods. The results indicate that DCS not only accurately characterizes disease severity but also preserves the longitudinal evolution of the disease, making it a valuable tool for continuous assessment beyond conventional diagnostic labels.
Methodology
The DCP framework employs a longitudinal Bayesian learning approach that integrates longitudinal DTI scans with weak clinical supervision. It models disease severity as a low-dimensional probabilistic latent variable, from which the Disease Continuum Score (DCS) is derived. This approach allows for the continuous assessment of disease severity rather than relying on discrete diagnostic categories.
Results
The experiments conducted on the ADNI cohort showed that DCP consistently outperformed existing methods for estimating disease progression. The DCS accurately characterized disease severity and demonstrated strong clinical relevance, effectively preserving longitudinal disease evolution and predicting future transitions from CN to MCI and MCI to AD.
Implications
The DCP framework and the DCS have significant implications for early diagnosis, prognosis, and management of Alzheimer's disease. By providing a continuous representation of disease progression, these tools can enhance clinical decision-making and improve patient outcomes.
In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
Theory
Generative Models
Efficient ML
- Introduction of the 4MAS architecture, which utilizes dual hemispheric specialization for continual learning.
- Implementation of an explicit wake-sleep training cycle to enhance memory consolidation and reduce representational drift.
- Achieved high accuracy rates on Split-MNIST (98.3%), Split-Fashion-MNIST (84.9%), and Split-CIFAR-100 (29.29%).
- Demonstrated low representational drift across tasks, indicating effective retention of previously learned information.
Read more
In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
Summary
This paper addresses the challenge of catastrophic forgetting in machine learning, particularly in the context of continual learning. The authors propose a novel architecture called 4MAS (4 Module Awake/Sleep), inspired by biological learning mechanisms, which incorporates two asymmetric hemispheres to mimic the lateralisation of brain functions. The architecture features dedicated long- and short-term memory modules for each hemisphere, allowing for effective generative replay and cross-hemispheric consolidation during a simulated sleep phase. This approach aims to stabilize latent representations and minimize representational drift when learning new tasks sequentially. The paper presents competitive performance metrics on several benchmark datasets, demonstrating the efficacy of the proposed model in mitigating forgetting while adapting to new information.
Methodology
The authors developed the 4MAS architecture, which consists of two asymmetric hemispheres, each equipped with long- and short-term memory modules. The training process alternates between an awake phase for task learning and a sleep phase for cross-hemispheric consolidation, mimicking biological processes such as hippocampal replay and REM sleep. This design aims to facilitate continual learning while minimizing catastrophic forgetting.
Results
The 4MAS architecture achieved competitive accuracies of 98.3% on Split-MNIST, 84.9% on Split-Fashion-MNIST, and 29.29% on Split-CIFAR-100. The model also exhibited low representational drift across tasks, showcasing its effectiveness in retaining previously learned knowledge while adapting to new tasks.
Implications
The findings suggest that incorporating biological principles into machine learning architectures can significantly enhance continual learning capabilities. This approach could lead to more robust AI systems capable of learning over time without the need for retraining on entire datasets, making them more applicable in real-world scenarios where data availability is limited.
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Theory
Efficient ML
Time Series
- Introduces 'defer' as a statistically valid action in continual learning systems.
- Develops a decision layer based on conditional JensenβShannon discrepancy for expert management.
- Demonstrates improved accuracy and efficiency in expert selection and decision-making.
- Achieves zero false spawns and missed concepts in benchmark tests.
Read more
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Summary
This paper addresses the challenges faced by continual-learning systems in managing uncertainty when integrating new data batches. Traditionally, such systems have treated uncertainty as a problem to be resolved immediately, but this work introduces a statistically defined action called 'defer,' which allows the system to wait for more evidence before making a decision. The author presents a decision layer based on a two-axis task comparison using the conditional JensenβShannon discrepancy (CJSD) and its covariate counterpart. The paper outlines three main contributions: (1) Decision semantics that define reuse and spawn tests as one-sided sequential hypotheses separated by an indifference zone, where 'defer' is the state of insufficient evidence; (2) A sequential evidence mechanism that uses per-expert betting processes to guide decisions, ensuring finite-time validity for observable discrepancies; and (3) System mechanics that optimize expert selection and decision-making processes to improve accuracy in nonstationary environments. The proposed system demonstrates superior performance in synthetic and real-world datasets, achieving zero false spawns and missed concepts, while maintaining fewer experts compared to traditional methods.
Methodology
The methodology involves a decision layer that utilizes a two-axis task comparison framework based on the conditional JensenβShannon discrepancy (CJSD). The system maintains betting processes for each expert to evaluate incoming data batches and decide whether to reuse, spawn a new expert, or defer the decision. The approach includes a statistical framework for managing evidence accumulation and ensuring validity in decision-making.
Results
The proposed system achieved zero false spawns and zero missed concepts in a synthetic stream benchmark, maintaining an optimal number of experts. In real-world datasets, it effectively managed expert pools, demonstrating superior performance with fewer experts compared to traditional methods, which often required more experts to achieve similar accuracy.
Implications
The findings suggest that the proposed decision-making framework can significantly enhance the efficiency of continual learning systems, particularly in environments with nonstationary data. This could lead to more adaptive and robust machine learning applications across various domains, including robotics and time series analysis.
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Efficient ML
Interpretability
Theory
- DICS utilizes clustering to inform candidate split selection, reducing computational costs.
- Theoretical analysis confirms that DICS does not compromise classification accuracy compared to exhaustive methods.
- Extensive experiments validate DICS's efficiency, achieving comparable accuracy with reduced training time.
- DICS can be applied to various tree-based models, enhancing their scalability.
Read more
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Summary
The paper introduces Data-Informed Centroid Splitting (DICS), a novel approach aimed at enhancing the efficiency of decision tree classifiers by reducing the computational burden associated with candidate split selection. Traditional decision tree training involves an exhaustive search for optimal splits at each node, which can be particularly costly for large and high-dimensional datasets. DICS addresses this issue by leveraging clustering techniques to create a compact set of candidate splits informed by the data's structure. The authors provide theoretical analysis demonstrating that DICS maintains predictive performance comparable to exhaustive search methods. Extensive experiments across synthetic and benchmark datasets show that DICS not only preserves accuracy but also significantly reduces training time, making it a scalable solution for decision tree learning. The method can be integrated into various tree-based models, including classification trees, random forests, and gradient-boosting models, thereby broadening its applicability in machine learning tasks.
Methodology
DICS employs a clustering-based framework to derive a compact set of candidate splits for decision trees. By analyzing the underlying data structure, it constructs splits that are both informative and computationally efficient, allowing for faster tree training without sacrificing performance. The method integrates seamlessly into existing decision tree algorithms, including random forests and gradient-boosting models.
Results
The experiments conducted demonstrate that DICS achieves similar accuracy levels to traditional exhaustive split search methods while significantly reducing training time across various datasets. This highlights the effectiveness of using data-informed priors in the split selection process.
Implications
The findings suggest that DICS can be a valuable tool for practitioners looking to improve the efficiency of decision tree-based models, particularly in scenarios involving large and complex datasets. Its integration into existing frameworks could lead to faster model training and deployment in real-world applications.
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
Large Language Models
- LLM-Detector utilizes in-context learning to derive anomaly detection logic from structured normal-state knowledge.
- The framework eliminates the need for LLM fine-tuning, reducing computational costs.
- Extensive evaluations show LLM-Detector outperforms existing statistical, deep learning, and LLM-based methods.
- The approach effectively handles mixed-type data and captures relational dependencies among features.
Read more
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
Summary
This paper introduces LLM-Detector, a novel framework for tabular anomaly detection that leverages the in-context learning capabilities of Large Language Models (LLMs). Unlike traditional methods that rely on geometric or reconstruction signals, LLM-Detector synthesizes anomaly detection logic from structured normal-state knowledge without requiring LLM fine-tuning. The framework converts normal training data into statistical summaries, causal dependencies, and distilled prototypes, which are organized into a prompt for code generation. The resulting scoring engine evaluates statistical deviation, structural inconsistency, and density-based abnormality to compute anomaly scores for test samples. Evaluations on 24 tabular datasets demonstrate that LLM-Detector consistently outperforms 15 state-of-the-art baselines across various settings, highlighting its effectiveness and efficiency in real-world applications.
Methodology
The methodology involves two main phases: first, extracting statistical profiles, causal relations, and representative prototypes from normal training data to create a structured prompt. This prompt induces the LLM to generate a deterministic program for anomaly detection. Second, the synthesized program evaluates unseen samples by aggregating various signals to produce anomaly scores.
Results
The results indicate that LLM-Detector achieves superior detection performance compared to 15 state-of-the-art methods across 24 diverse tabular datasets, demonstrating consistent improvements in both mixed-type and continuous-only settings.
Implications
The implications of this research suggest that LLM-Detector can be effectively applied in critical domains such as finance, healthcare, and cybersecurity, where accurate anomaly detection is essential for preventing costly failures and risks. Its efficiency and ability to handle complex relational dependencies make it a valuable tool for practitioners.
Triangular Fuzzy Rescaling Distance
Theory
- Introduction of the Triangular Fuzzy Rescaling Distance (dTR) for comparing TFNs.
- dTR integrates Linear Rescaling directly into the distance calculation.
- Proven to satisfy all properties of a metric, including scale and origin invariance.
- Suitable for applications in heterogeneous fuzzy data contexts.
Read more
Triangular Fuzzy Rescaling Distance
Summary
This paper addresses the challenge of measuring distances between Triangular Fuzzy Numbers (TFNs) in decision-making processes that involve uncertain or imprecise information. Traditional distance measures often require normalization to handle heterogeneous attributes with varying scales, which can distort results. The authors propose a novel metric called the Triangular Fuzzy Rescaling Distance (dTR), which integrates Linear Rescaling (LRE) directly into the distance calculation, ensuring that normalization occurs during the comparison of fuzzy numbers. The paper rigorously proves that dTR satisfies all properties of a metric, including non-negativity, identity, symmetry, and the triangle inequality. Additionally, dTR is shown to be bounded, scale-invariant, and origin-invariant, making it particularly suitable for applications involving heterogeneous fuzzy data. The authors illustrate the calculation of dTR through numerical examples and case studies, demonstrating its practical utility in constructing synthetic indicators and supporting multicriteria decision-making.
Methodology
The authors define the Triangular Fuzzy Rescaling Distance (dTR) mathematically and prove its properties as a metric. They utilize Linear Rescaling (LRE) to ensure normalization in the distance calculation. The paper includes numerical examples and a case study to demonstrate the application of dTR.
Results
The paper successfully establishes that dTR is a valid metric for comparing n-tuples of TFNs, demonstrating its boundedness and invariance properties. The numerical examples and case studies validate the effectiveness of dTR in practical applications.
Implications
The proposed dTR metric has significant implications for decision-making processes that involve fuzzy data, particularly in fields such as synthetic indicator construction, distance-based machine learning algorithms, and multicriteria decision aiding. Its ability to handle heterogeneous data without distortion enhances its applicability in real-world scenarios.
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
NLP
Interpretability
- BERT-LER integrates laboratory test results into a transformer model while maintaining interpretability.
- The model achieves competitive performance on both benchmark datasets and real-world clinical tasks.
- Token-level attributions align with clinically known risk factors, enhancing the model's utility in clinical settings.
- The approach demonstrates the feasibility of combining predictive performance with interpretability in healthcare applications.
Read more
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
Summary
This paper presents BERT-LER, a novel BERT-style transformer model designed for clinical prediction tasks using structured electronic health records (EHRs). The model uniquely encodes laboratory test results as discrete tokens while preserving graded information through percentile-based binning. Additionally, it employs Integrated Gradients for token-level attributions, enhancing interpretability by linking model predictions to specific medical events. BERT-LER was evaluated on the EHRShot benchmark suite and in a real-world asthma severity progression study, demonstrating competitive predictive performance, particularly in laboratory-related tasks. The findings highlight the potential for integrating strong predictive capabilities with interpretable outputs, which is crucial for clinical validation and decision-making. This work addresses a significant methodological gap in EHR modeling by combining laboratory data representation with explainability, offering insights that can be generalized across various therapeutic areas and prediction tasks.
Methodology
BERT-LER is a BERT-style model that encodes laboratory results as discrete tokens using percentile-based binning. It utilizes Integrated Gradients for generating token-level attributions that connect model predictions to input medical events. The model was pretrained and fine-tuned on a large dataset of 75 million de-identified EHRs and evaluated on both a public benchmark and a specific clinical study.
Results
BERT-LER demonstrated predictive performance that is competitive with existing benchmark models, often exceeding them in laboratory-related tasks. The model's attributions were clinically relevant, aligning with known risk factors, thus validating its interpretability.
Implications
The integration of strong predictive performance with interpretable outputs can facilitate the adoption of machine learning models in clinical workflows, enhancing decision-making processes in healthcare. The methodology can be applied to various clinical prediction tasks, potentially improving patient outcomes.
Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay
Reinforcement Learning
Large Language Models
Theory
- Step-level credit signals in LLM training do not reliably indicate causally significant actions.
- Causal contributions are sparse and vary significantly between different model architectures.
- Implicit credit scores correlate more with policy fluency than with actual causal contributions.
- A confidence-only routing rule can reduce costs while maintaining performance at chance levels.
Read more
Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay
Summary
This paper investigates the effectiveness of step-level credit assignment signals used in training large language model (LLM) agents, specifically in the context of a single-agent tool environment called ALFWorld. The study reveals that none of the evaluated credit signalsβsuch as LLM-judge scores, outcome-conditioned log-probability ratios, or the policy's own confidenceβcan reliably identify causally significant steps better than chance. The authors establish a causal ground truth through executed replay, which highlights that causal contributions are sparse and model-dependent. The findings indicate that implicit credit often reflects the fluency of the policy rather than its actual contribution to outcomes. A novel confidence-only routing decision rule is proposed, which reduces judge costs while maintaining chance-level pivotal recall. The paper emphasizes the need for a structured approach to measuring causal contributions in LLM training and introduces protocols for credit comparison, underscoring that existing credit assignment methods may not effectively capture the true impact of individual steps.
Methodology
The authors utilize executed replay to measure causal contributions at each decision point in a trajectory. By resampling alternative actions supported by the policy and rolling them forward, they assess how these changes affect outcome distributions. This method allows for the establishment of a causal ground truth against which credit assignment signals can be audited.
Results
The audit reveals that none of the credit signals outperform their own shuffled controls, indicating that they do not effectively capture causal contributions. The analysis shows that only 30.5% of decision points have measurable effects, and the ability to measure contributions varies significantly between model families. The confidence-only routing rule achieves a reduction in judge costs while maintaining performance at chance levels.
Implications
The findings suggest that current methods for credit assignment in LLM training may need reevaluation, as they do not accurately reflect the causal contributions of individual steps. This could lead to more effective training strategies that better align with actual agent performance and decision-making processes. The introduced protocols for credit comparison may serve as a foundation for future research in this area.
When to Retrain: An Empirical Study of Retraining Policies for Streaming ML Under Concept Drift, Budget, and Latency Constraints
Theory
Efficient ML
Time Series
- Incremental learning is crucial; it often negates the need for retraining policies.
- Periodic retraining significantly outperforms reactive policies under abrupt and gradual drift.
- Reactive policies exhibit systematic failures, such as budget exhaustion and drift blindness.
- Latency constraints can halve effective retraining budgets, impacting model performance.
Read more
When to Retrain: An Empirical Study of Retraining Policies for Streaming ML Under Concept Drift, Budget, and Latency Constraints
Summary
This paper addresses the critical challenge of determining when to retrain machine learning models in production environments, particularly under conditions of concept drift, budget limitations, and latency constraints. The author conducts a comprehensive empirical study comparing three retraining policiesβperiodic retraining, error-threshold triggering, and statistical drift-triggered retraining using ADWINβagainst a no-retrain baseline. The study spans 3,933 experimental runs across various drift regimes (abrupt, gradual, recurring), budget levels, latency levels, and datasets, including both synthetic and real-world scenarios. The findings reveal that the most significant factor influencing model performance is whether the deployed model can learn incrementally. For models with incremental learning, no retraining policy significantly outperforms the no-retrain approach. Conversely, without incremental updates, periodic retraining outperforms reactive policies, particularly under abrupt and gradual drift conditions. The paper also identifies systematic failures in reactive policies and highlights the impact of latency on retraining budgets. The research contributes an open-source simulator and datasets for reproducibility, providing valuable insights for practitioners in the field of streaming machine learning.
Methodology
The study employs a controlled experimental framework using an open-source streaming simulator that incorporates various drift types, retraining policies, budget constraints, and latency mechanisms. It evaluates the performance of different retraining strategies through a factorial design across multiple datasets and statistical significance testing.
Results
The results indicate that with incremental updates, no retraining policy significantly improves upon the no-retrain baseline. Without incremental updates, periodic retraining outperforms reactive policies by 15-55 percentage points in post-drift accuracy, particularly under abrupt and gradual drift scenarios. Systematic failures of reactive policies were documented, including pre-drift budget exhaustion and inefficiencies in handling gradual drift.
Implications
The findings suggest that practitioners should prioritize incremental learning capabilities in their models to mitigate the need for frequent retraining. Additionally, the study provides a framework for evaluating retraining strategies, which can inform better decision-making in production ML systems facing concept drift.
A Standardized Framework for Machine Learning in Power System Protection
Theory
Optimization
Time Series
- Proposes a standardized framework for evaluating ML in power system protection.
- Defines seven critical dimensions for study design to enhance reproducibility.
- Demonstrates the framework with a case study using the PROTECT-90 benchmark.
- Achieves high performance metrics with a multi-layer perceptron for fault classification and localization.
Read more
A Standardized Framework for Machine Learning in Power System Protection
Summary
This paper addresses the challenges of evaluating machine learning (ML) approaches in power system protection, where reported performance metrics often lack context due to varying evaluation settings. The authors propose a standardized framework that emphasizes the importance of clearly defined evaluation designs as part of the scientific contribution. The framework outlines seven essential dimensions for study design: protection objective, physical system scope, observability and measurements, timing and decision windows, targets and valid samples, training and validation protocol, and evaluation outputs. To demonstrate the framework, the authors conduct a case study using the PROTECT-90 benchmark, which includes 9022 simulated episodes for fault classification and localization in a 90 kV double-line topology. The study employs a multi-layer perceptron (MLP) and achieves a high mean macro-averaged F1 score of 0.991 for classification and a mean absolute error of 10.20% for localization. The results indicate that extending the decision horizon maintains performance asymmetry, while reduced observability impacts localization accuracy. The proposed framework aims to enhance the reproducibility and comparability of ML-based protection research, ultimately contributing to the certification and trustworthiness of these systems in safety-critical applications.
Methodology
The authors developed a standardized evaluation framework and applied it in a case study using the PROTECT-90 electromagnetic transient benchmark. They utilized a multi-layer perceptron (MLP) for fault classification and localization, analyzing performance across various decision windows and observability conditions.
Results
The MLP achieved a five-fold mean macro-averaged F1 score of 0.991 Β± 0.001 for classification and a mean absolute localization error of 10.20 Β± 0.25% of line length. Performance was maintained with extended decision horizons, but reduced observability led to increased localization errors.
Implications
The proposed framework can guide future research in ML-based power system protection, ensuring that evaluations are transparent and comparable. This can facilitate the certification of ML systems in safety-critical environments, enhancing trust in their deployment.
Auditing Recorded Predictive Lead Service-Line Classifications Against Physical Verification: A Statewide Study of New York
Theory
- 49% of localities recorded a single material value, with some inconsistencies in physical verification.
- New York City's predictive model classified a large number of addresses as 'Known Other' without verification.
- The study highlights the importance of physical verification in validating predictive models for public health.
- Significant discrepancies were found between predictive model outputs and actual physical verifications.
Read more
Auditing Recorded Predictive Lead Service-Line Classifications Against Physical Verification: A Statewide Study of New York
Summary
This paper investigates the reliability of predictive models used by water utilities in New York State to classify service line materials under the US Lead and Copper Rule Revisions. The study audits the recorded classifications against physical verifications conducted by the utilities themselves. The analysis covers 153 localities, revealing that 49% of them recorded a single distinct material value for service lines, with 68 of these being consistent with physical verification. However, seven localities, including five boroughs of New York City, showed contradictions in their classifications. The findings highlight significant discrepancies in the predictive model's outputs, particularly in New York City, where a large number of addresses were classified as 'Known Other' without any physical verification. The study also examines the construction era of the buildings, finding that the predictive model fails to identify lead service lines across various eras, while physical verification does find lead in older constructions. The paper emphasizes the need for rigorous validation of predictive models used in public health contexts, particularly regarding lead exposure risks.
Methodology
The study screened all 153 localities in New York that classified at least 100 addresses using predictive models. It compared the recorded outputs against physical verifications, focusing on zero output variance and contradictions in classifications. The analysis included spatial joins to assess the impact of construction era on classification accuracy.
Results
The analysis revealed that 49% of the localities had zero variance in their recorded material classifications, with only a small fraction being contradicted by physical verification. New York City's model showed a significant lack of lead identification, with a 95% upper bound on the rate of lead being 0.0085%. In contrast, physical verification found lead in 12.21% of addresses across the state, primarily in Poughkeepsie. The study also identified issues with the classification of addresses built before 1940, where the model's outputs did not align with state guidance.
Implications
The findings suggest that reliance on predictive models without sufficient physical verification can lead to underreporting of lead service lines, posing risks to public health. The study calls for improved validation processes for predictive models in water utilities and highlights the need for adherence to regulatory standards.
Towards On-Board Implementation of ML-Based Helicopter Weight Estimator
Time Series
Robotics
Efficient ML
- Development of a supervised ML model for helicopter weight estimation during takeoff.
- Alignment with EASA guidelines and Eurocae ED-324 standards for ML applications in aviation.
- Implementation of an LSTM recurrent neural network architecture for the weight estimator.
- Verification of the model's compliance with safety requirements on legacy avionics systems.
Read more
Towards On-Board Implementation of ML-Based Helicopter Weight Estimator
Summary
This paper presents the development and implementation of a supervised Machine Learning (ML) model designed to estimate helicopter weight during takeoff, utilizing extensive datasets from Airbus's global fleet. The authors detail a learning assurance process that aligns with the European Union Aviation Safety Agency (EASA) guidelines and the ongoing Eurocae ED-324 standards. The proposed model is based on a Long Short-Term Memory (LSTM) recurrent neural network architecture. The paper outlines the Machine Learning Requirements (MLCR), the Machine Learning Model Description (MLMD), and the implementation process, which has been verified on legacy avionics computers. The implementation is deemed suitable for deployment on airborne systems, enabling critical functions such as on-board alerting. The study emphasizes the importance of accurate weight estimation for enhancing safety and optimizing operational costs, particularly within Condition-Based Maintenance (CBM) frameworks. The authors also address the challenges of ensuring compliance with safety standards and the verification of ML models in aviation contexts.
Methodology
The authors employed a W-shape development process as outlined in EASA guidelines, focusing on defining Machine Learning Requirements (MLCR) and Data Processing Requirements (DPR). They implemented a Long Short-Term Memory (LSTM) model, verified its performance through statistical methods, and ensured compliance with safety standards through a structured verification process involving High Level Tests (HLT).
Results
The implementation of the ML-based weight estimator was successfully demonstrated on legacy avionics computers, confirming its suitability for on-board deployment. The model was verified against high-level requirements, ensuring its operational reliability and safety in real-world scenarios.
Implications
The successful implementation of this ML-based weight estimator could significantly enhance helicopter safety through improved Helicopter Terrain Awareness and Warning Systems (HTAWS). Additionally, it may lead to cost savings in maintenance and operational efficiency by enabling more accurate weight predictions and optimizing the retirement time of mechanical parts.
CLaST: Context-aware Contrastive VAE for Probabilistic Time Series Forecasting
Generative Models
Time Series
Optimization
- CLaST introduces a context-aware contrastive loss function for improved probabilistic time series forecasting.
- The framework captures contextual similarity, enhancing the quality of latent representations.
- Theoretical properties of the loss function are derived under the LINTS process, ensuring optimality.
- Empirical results show CLaST consistently outperforms strong baseline methods across diverse datasets.
Read more
CLaST: Context-aware Contrastive VAE for Probabilistic Time Series Forecasting
Summary
The paper introduces CLaST, a novel variational autoencoder (VAE) framework designed for probabilistic multivariate time series forecasting. Traditional forecasting models often struggle to capture internal temporal dependencies, leading to inadequate latent representations. CLaST addresses this by employing a contrastive loss function that preserves contextual similarity between observations, enhancing the expressiveness of learned embeddings. The authors define a new class of time series called Lag-Invariant Non-stationary Time Series (LINTS), which allows for a more realistic modeling of temporal dynamics. The proposed method is validated through extensive experiments on nine benchmark datasets, demonstrating significant improvements over existing models in both short-term and long-term forecasting tasks. CLaST achieves up to 16.4% improvement in Continuous Ranked Probability Score (CRPS) and 14.4% in Normalized Mean Absolute Error (NMAE) for short-term forecasts, and up to 48.6% and 25.1% improvements in CRPS and NMAE, respectively, for long-term predictions.
Methodology
CLaST employs a variational autoencoder architecture that integrates a novel contrastive loss function to maintain contextual similarity between time series observations. The method is grounded in the LINTS framework, which allows for modeling non-stationary time series with lag-dependent covariance structures. This approach replaces traditional mutual information estimators with a more stable contrastive-based penalty, enhancing training stability and representation quality.
Results
CLaST demonstrated superior performance in forecasting tasks, achieving up to 16.4% improvement in CRPS and 14.4% in NMAE for short-term predictions, and up to 48.6% and 25.1% improvements in CRPS and NMAE, respectively, for long-term predictions across nine benchmark datasets.
Implications
The findings suggest that CLaST can be effectively applied in various domains requiring probabilistic time series forecasting, such as finance, energy systems, and healthcare, where capturing temporal dependencies and contextual similarities is crucial for accurate predictions.
Orthogonal JEPA: Factorized Predictive States for Latent World Models
Computer Vision
Time Series
Robotics
- Orthogonal JEPA introduces a factorized approach to predictive state modeling, enhancing the learning of latent states in complex systems.
- The framework employs multiple learned components for target states, each with dedicated prediction branches, improving predictive accuracy.
- Regularization techniques are integrated to maintain variation and prevent encoder collapse, ensuring robust learning.
- The methodology is validated across diverse applications, demonstrating its versatility and effectiveness in various domains.
Read more
Orthogonal JEPA: Factorized Predictive States for Latent World Models
Summary
The paper introduces Orthogonal JEPA, a novel framework for latent world modeling that enhances the predictive capabilities of joint-embedding predictive architectures (JEPAs). Traditional JEPAs utilize a single target embedding and prediction pathway, which can lead to inefficiencies in complex systems where multiple signals interact. Orthogonal JEPA addresses this limitation by employing orthogonal predictive factorization, allowing the decomposition of target states into multiple components, each predicted by its own dedicated branch. This approach not only preserves the necessary factor magnitudes for state synthesis but also incorporates regularization techniques to maintain diversity in predictions and prevent encoder collapse. The framework is versatile, applicable across various domains including vision, single-cell transcriptomics, health records, control systems, and molecular dynamics. Experimental evaluations demonstrate the effectiveness of Orthogonal JEPA in improving representation quality, forecasting accuracy, planning capabilities, and stability over long horizons.
Methodology
The Orthogonal JEPA framework utilizes learned basis matrices to decompose target states into multiple components, each predicted through separate branches. It incorporates an orthogonality objective to ensure distinct predictive directions, along with factor-activity regularization and online variance regularization to maintain diversity and prevent coordinate-wise collapse in the encoder. The framework is adaptable to different domains through specific context-target interfaces.
Results
Experiments conducted across various domains, including controlled vision tasks, single-cell transcriptomics, longitudinal health records, continuous control, and molecular dynamics, show that Orthogonal JEPA significantly improves representation quality, forecasting accuracy, and planning capabilities. The framework also demonstrates enhanced stability in long-horizon predictions compared to traditional JEPA approaches.
Implications
The advancements presented in Orthogonal JEPA have the potential to improve predictive modeling in various fields, including healthcare, robotics, and molecular dynamics. By enabling more effective latent state learning, this framework can enhance decision-making processes in complex systems and contribute to the development of more sophisticated AI models.
End-to-end Early Classification of Time Series in Non-Stationary Environments
Reinforcement Learning
Time Series
Optimization
- Introduces DQeND, an end-to-end framework for ECTS that optimizes classification and triggering jointly.
- Demonstrates the limitations of separable designs in adapting to non-stationary environments.
- Provides empirical evidence that end-to-end learning improves robustness and adaptability in ECTS.
- Highlights the importance of joint optimization in enhancing decision-making processes.
Read more
End-to-end Early Classification of Time Series in Non-Stationary Environments
Summary
This paper addresses the challenge of Early Classification of Time Series (ECTS) in non-stationary environments, where the underlying data distribution can change over time. Traditional methods often assume stationarity and optimize classification and triggering decisions independently, which limits their adaptability to concept drift. The authors propose a novel end-to-end framework called DQeND that leverages Reinforcement Learning to jointly learn representation, classification, and triggering decisions. Through systematic experiments, the paper compares the performance of separable and end-to-end architectures under controlled drifting scenarios. The results demonstrate that DQeND significantly outperforms separable baselines in terms of robustness and adaptability, particularly in dynamic environments. An ablation study reveals that the joint optimization of representation and decision-making modules is crucial for achieving these improvements. Overall, the findings suggest that end-to-end learning can enhance ECTS systems' performance in real-world applications where timely and accurate decision-making is essential.
Methodology
The authors developed DQeND, an end-to-end architecture that employs Reinforcement Learning to simultaneously optimize representation, classification, and triggering decisions. They conducted controlled experiments involving covariate shifts and concept drifts to evaluate the performance of DQeND against traditional separable methods.
Results
The experiments revealed that DQeND consistently outperformed separable baselines across various non-stationary scenarios, demonstrating enhanced robustness and adaptability. The ablation study confirmed that jointly updating the representation and decision modules was critical for achieving superior performance.
Implications
The findings suggest that end-to-end learning frameworks can significantly improve the performance of ECTS systems in dynamic environments, making them more suitable for real-world applications such as social media monitoring, where timely and accurate classification is crucial. This work encourages further exploration of end-to-end approaches in other machine learning domains.
RecPFN: Prior-Fitted Networks for In-Context-Based Recommendations
Efficient ML
Theory
NLP
- RecPFN introduces a synthetic framework for generating clickstream sequences for training.
- It employs a lightweight transformer for next-item prediction based on in-context examples.
- Achieves state-of-the-art zero-shot performance while being competitive in low-data environments.
- Demonstrates robustness to domain shifts and efficiency in deployment.
Read more
RecPFN: Prior-Fitted Networks for In-Context-Based Recommendations
Summary
The paper introduces RecPFN, a novel prior-fitted network designed to enhance sequential recommendation systems through in-context learning (ICL). RecPFN is pretrained on synthetic clickstream data generated from a broad structural causal prior, allowing it to perform Bayesian-style inference effectively with minimal data. At inference time, a lightweight decoder-only transformer utilizes a small set of domain sequences to predict the next item in a single forward pass without requiring weight updates. The authors demonstrate that RecPFN achieves state-of-the-art zero-shot performance across eight public benchmarks, while also being competitive with supervised methods in low-compute and low-data scenarios. This approach is noted for its deployment efficiency and robustness to domain shifts, outperforming traditional zero-shot baselines that depend on extensive real-interaction datasets. The findings suggest that RecPFN paves the way for more generalizable and data-efficient recommendation systems, with potential for future enhancements through richer priors and multimodal extensions.
Methodology
RecPFN utilizes a prior-fitted network architecture that is pretrained on synthetic data generated from a structural causal model. The model leverages Bayesian inference principles to adapt to new domains using a small number of examples during inference, employing a transformer architecture to predict the next item in a sequence efficiently.
Results
The results indicate that RecPFN outperforms existing zero-shot recommendation models and rivals supervised methods, particularly in scenarios with limited computational resources and data. The model's ability to generalize across different domains without extensive retraining is a significant highlight.
Implications
The development of RecPFN suggests a shift towards more efficient and adaptable recommendation systems that can operate effectively in diverse environments with minimal data. This has implications for industries relying on personalized recommendations, such as e-commerce and content platforms, and encourages further exploration of synthetic data in machine learning.
PETA: Parameter-Efficient Test-Time Adaptation for Virtual Screening
Efficient ML
- PETA formulates Test-Time Adaptation for Virtual Screening (TTA-VS), allowing specialization of pretrained models to unseen protein pockets without full retraining.
- The framework constructs target-conditioned supervision using information available at test time, updating only a small fraction of model parameters.
- PETA achieves superior performance compared to both frozen and fully retrained models, highlighting its efficiency.
- The method emphasizes the importance of suppressing high-scoring invalid candidates in ranking tasks.
Read more
PETA: Parameter-Efficient Test-Time Adaptation for Virtual Screening
Summary
The paper addresses the challenge of accurately ranking active ligands for target protein pockets from extensive chemical libraries in virtual screening. While existing methods like DrugCLIP have improved efficiency by encoding protein pockets and molecules into a shared embedding space, they often require full model retraining for performance enhancements, which is computationally expensive. The authors propose PETA, a parameter-efficient framework that adapts pretrained models at test time without the need for extensive retraining. PETA constructs pocket-specific negative samples through molecular diffusion and chemical validity filtering, and uses embedding-space mixup to create challenging ranking tasks. A cost-sensitive ranking objective is employed to suppress high-scoring invalid candidates, facilitating lightweight adaptation by updating only LayerNorm parameters, which constitute about 0.03% of the full model. Experimental results show that PETA outperforms both pretrained and fully retrained baselines, demonstrating a balance between screening performance and adaptation efficiency.
Methodology
PETA employs a test-time adaptation approach that updates only LayerNorm parameters of a pretrained virtual screening model. It generates pocket-specific negative samples and uses a reference ligand for adaptation. The framework incorporates embedding-space mixup and a cost-sensitive ranking objective to enhance the ranking process.
Results
PETA achieved an EF0.5% score of 42.34, surpassing the frozen DrugCLIP (37.90) and fully retrained BindCLIP (39.82) models, while only updating approximately 0.03% of the model parameters.
Implications
The proposed method can significantly reduce computational costs and time in virtual screening processes, making it a valuable tool for drug discovery and development, particularly in scenarios with large chemical libraries.
DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule
Optimization
- DELTAMOMENTUM introduces direction-aware momentum updates based on the frequency of gradient direction queries.
- The method utilizes a key-value structure of gradients to optimize the forgetting rate for each direction.
- Proven to be a valid momentum update, DELTAMOMENTUM applies curvature correction without the need for matrix inversion.
- Experimental results show substantial improvements in training efficiency, achieving lower validation loss in fewer steps compared to traditional optimizers.
Read more
DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule
Summary
The paper introduces DELTAMOMENTUM, a novel momentum update method for optimizers that addresses the anisotropic nature of gradient inputs during neural network training. Traditional optimizers utilize an exponential moving average (EMA) for momentum, which applies a fixed forgetting rate across all directions, leading to inefficiencies in gradient tracking. DELTAMOMENTUM leverages a key-value structure inherent in the gradient of linear layers, where the input acts as a key and the output-side error as a value. This method updates the momentum buffer using the delta rule, allowing for direction-specific forgetting rates based on how frequently each direction is queried. The authors prove that DELTAMOMENTUM is a valid momentum update, applies curvature correction without matrix inversion, and clears stale directions faster than EMA. It is designed to be a drop-in replacement for existing momentum buffers in various optimizers, maintaining compatibility with methods like AdamW and Muon. Experimental results demonstrate that DELTAMOMENTUM significantly reduces the number of steps required to reach validation loss benchmarks in language model pretraining, outperforming traditional methods while maintaining low computational overhead.
Methodology
The authors propose DELTAMOMENTUM, which updates the momentum buffer based on the delta rule, treating the input as a key and the output error as a value. This allows for a dynamic forgetting rate that adapts to the frequency of direction queries, improving gradient tracking and efficiency in training.
Results
In experiments with language model pretraining at scales of 67M, 370M, and 1B parameters, DELTAMOMENTUM (DeltaAdamW) achieved validation loss benchmarks in up to 46.39% fewer steps at 67M and 22.12% fewer steps at 370M compared to AdamW. The performance gains were consistent across multiple seeds and persisted at larger model sizes.
Implications
DELTAMOMENTUM has the potential to enhance the efficiency of training large neural networks, making it feasible to train larger models on more data within fixed compute budgets. Its compatibility with existing optimizers allows for broader adoption in various machine learning applications.
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
Reinforcement Learning
Large Language Models
Optimization
- MileGPO reveals the unreliability of intermediate credit based on final-goal distance.
- The proposed method learns from rollout graphs and on-policy rewards without external annotations.
- MileGPO achieves state-of-the-art performance on ALFWorld and WebShop benchmarks.
- The method effectively calibrates intermediate credit, enhancing learning efficiency.
Read more
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
Summary
The paper addresses the challenge of credit assignment in long-horizon reinforcement learning, particularly for large language model (LLM) agents, where supervision is often limited to final rewards. Existing methods struggle to assign meaningful credit to intermediate milestones. The authors propose MileGPO, a novel approach that derives process-level credit from grouped on-policy rollouts through three main components: Milestone Discovery (MD), which identifies significant milestones and traps; Reliability-Calibrated Shaping (RCS), which weights these candidates based on outcome confidence; and Progress-Contrastive Calibration (PCC), which evaluates local progress and compares transitions from the same state. MileGPO operates without the need for auxiliary models or additional environment interactions. Experimental results on the ALFWorld and WebShop benchmarks demonstrate that MileGPO achieves state-of-the-art performance, with a notably small in-distribution to out-of-distribution gap, indicating strong generalization capabilities. The findings highlight the importance of calibrated intermediate credit in improving performance over traditional methods.
Methodology
MileGPO employs a three-part framework: Milestone Discovery identifies key milestones and traps from successful and failed rollouts; Reliability-Calibrated Shaping weights these milestones based on their reliability; and Progress-Contrastive Calibration assesses local progress and compares sibling transitions to refine credit assignment. This approach leverages on-policy rollouts and does not require auxiliary models or additional interactions with the environment.
Results
MileGPO demonstrated superior performance on the ALFWorld and WebShop benchmarks, achieving state-of-the-art results. The in-distribution to out-of-distribution gap on ALFWorld was only 1.69 points, outperforming previous methods like GiGPO and GraphGPO. Further analysis indicated that the method's improvements stem from its calibrated intermediate credit rather than indiscriminate milestone propagation.
Implications
The findings suggest that MileGPO can enhance the training of LLM agents in complex, long-horizon tasks by providing more reliable credit assignment. This could lead to more effective reinforcement learning applications in areas such as web navigation, embodied instruction following, and other sequential decision-making tasks.
Answer-Level Trust Selection for Physical Vision-Language Reasoning
Multimodal
Computer Vision
Interpretability
- ATS is a model-agnostic framework for evaluating individual predictions from VLMs.
- It aggregates multiple behavioral diagnostic scores into a unified trust score.
- ATS can identify hidden failure modes that traditional evaluation methods may miss.
- The framework enhances the reliability assessment of quantitative predictions in VLMs.
Read more
Answer-Level Trust Selection for Physical Vision-Language Reasoning
Summary
This paper addresses the challenge of evaluating the reliability of individual predictions made by Vision-Language Models (VLMs) in the context of quantitative physical reasoning. Traditional benchmarks focus on overall model performance against ground truth, which does not account for the trustworthiness of specific predictions when ground truth is unavailable. The authors propose a novel framework called Answer-Level Trust Selection (ATS), which is a post-hoc, model-agnostic method that assesses the reliability of individual predictions without requiring fine-tuning or additional verifiers. ATS aggregates eight interpretable behavioral diagnostic scores derived from repeated queries and controlled interventions into a unified trust score. The evaluation of ATS is conducted on the Qwen2.5-VL-7B model and across 20 different VLM backbones, revealing that ATS can effectively identify stable-but-wrong predictions and those influenced by textual priors. However, the trade-off exists where improved rejection of failure cases may lead to a decrease in the retention of correct predictions. The findings suggest that ATS provides a complementary approach to model-level evaluations by focusing on answer-level reliability, which is crucial for practical deployment of VLMs in real-world applications.
Methodology
The authors developed the Answer-Level Trust Selection (ATS) framework, which operates post-hoc and is minimally invasive. It does not require fine-tuning of the VLM or access to internal model logits. Instead, ATS evaluates predictions based on eight distinct behavioral diagnostic scores derived from repeated queries and controlled interventions, which are then aggregated to form a trust score for each prediction.
Results
The evaluation of ATS on the Qwen2.5-VL-7B model and across 20 VLM backbones demonstrated its effectiveness in identifying stable-but-wrong predictions and those influenced by textual priors. The results indicated that while ATS improves the rejection of unreliable predictions, it may also lower the retention rate of correct predictions, highlighting a critical trade-off in its application.
Implications
The ATS framework has significant implications for the deployment of VLMs in real-world scenarios where individual prediction reliability is paramount. It provides a systematic approach to assess the trustworthiness of numerical outputs, which is essential for applications in robotics, decision support systems, and any domain requiring accurate quantitative reasoning from visual data.
Uncovering the Limits of Proof Sharing for Neural Networks
Theory
Efficient ML
- Introduces a systematic framework to evaluate the effectiveness of template-based proof sharing for neural network verification.
- Develops a novel metric, jointly stable neurons, to explain the variability in template reuse success.
- Presents FastCert, a technique for optimizing template distribution across layers to enhance performance.
- Demonstrates that FastCert achieves an average speedup of 1.13Γ over existing methods in L0-verification tasks.
Read more
Uncovering the Limits of Proof Sharing for Neural Networks
Summary
This paper investigates the effectiveness of proof sharing techniques in the robustness verification of neural networks, particularly focusing on template-based acceleration methods. The authors conduct a systematic study to understand the limits of template reuse across various neural network architectures, properties, datasets, and training methods. They introduce a novel metric called jointly stable neurons to characterize the potential for template reuse and demonstrate that the effectiveness of template-based techniques can vary significantly depending on the scenario. The paper presents FastCert, a new method that automatically distributes templates across neural network layers to optimize performance, avoiding templates when they are unlikely to yield speedups. FastCert is shown to achieve an average speedup of 1.13Γ over existing template-based techniques in L0-verification tasks, significantly reducing verification time.
Methodology
The authors perform a limit study to analyze the effectiveness of template-based proof sharing, using a novel metric of jointly stable neurons to assess the potential for speedup. They implement FastCert, which profiles and generates templates before verification, predicting subsumption rates based on a small set of queries.
Results
FastCert achieves an average speedup of 1.13Γ over existing template-based techniques for L0-verification tasks, resulting in a significant reduction in wall-clock time by approximately 7 hours in certain cases. The study reveals that the potential for template-based acceleration varies widely across different scenarios.
Implications
The findings suggest that while template-based proof sharing can enhance verification efficiency, its effectiveness is highly context-dependent. FastCert provides a practical solution for optimizing verification processes in neural networks, which could be beneficial in critical applications requiring robust neural network performance.
Physical-Support Confidence Sets for Highly Coherent Dictionaries
Theory
Interpretability
Efficient ML
- Formulation of physical-support uncertainty after dictionary learning.
- Derivation of a minimax law for optimal physical resolution based on calibration signals.
- Introduction of active endpoint bracketing (AEB) for efficient candidate evaluation.
- Demonstration of AEB's ability to recover justified conclusions with reduced computational effort.
Read more
Physical-Support Confidence Sets for Highly Coherent Dictionaries
Summary
This paper addresses the challenges of sparse pursuit after dictionary learning, particularly when dealing with highly coherent dictionaries that can lead to ambiguous physical interpretations of selected atom supports. The author develops a resolution-aware physical-support inference method that accounts for uncertainties in both the learned dictionary and the deployment signal representation. The proposed approach includes an exact cross-dictionary confidence correspondence that retains calibration-compatible dictionaries and deployment-compatible sparse representations, projecting the surviving explanations onto a physical-support space. The paper derives a minimax physical resolution law, demonstrating that the optimal resolution is influenced by the number of calibration signals and the separation scale of coherent blocks. Additionally, the author introduces an adaptive finite-bank method called active endpoint bracketing (AEB), which efficiently evaluates candidates that can impact the physical report, ensuring statistical validity while minimizing unnecessary computations. Experimental results indicate that AEB can provide justified conclusions with fewer evaluations compared to traditional methods.
Methodology
The methodology involves developing a resolution-aware physical-support inference framework that utilizes an exact cross-dictionary confidence correspondence. The author derives a minimax law for optimal physical resolution and introduces AEB, an adaptive finite-bank method that selectively evaluates candidates based on their relevance to the physical report.
Results
The results show that the proposed AEB method can achieve finer physical interpretations than traditional exhaustive evaluations, while also avoiding unsupported refinements. The minimax resolution law indicates that optimal physical resolution is governed by the orientation-information scale, which is critical for understanding the physical structure within coherent blocks.
Implications
The findings have significant implications for applications in fields such as array source localization, hyperspectral unmixing, and M/EEG source imaging, where accurate physical interpretations of sparse representations are crucial. The proposed methods can enhance the reliability and efficiency of these applications by providing clearer insights into the physical supports associated with deployment signals.
Green BOA: Determining the environmental break-even point for ML-based data compression
Efficient ML
- The break-even point for ML-based data compression is sensitive to the carbon intensity of the energy source.
- ML-based compression algorithms can achieve better compression ratios but may have higher carbon costs compared to traditional methods.
- The study provides a proof-of-principle calculation for assessing the environmental impact of ML training and inference.
- Future work will explore a broader range of scenarios and improve throughput of ML-based compression.
Read more
Green BOA: Determining the environmental break-even point for ML-based data compression
Summary
This paper presents a study on the environmental sustainability of machine learning (ML)-based data compression algorithms, specifically focusing on the break-even point where the carbon costs of ML training and inference equal the carbon savings from reduced disk storage. The authors analyze a lossless compression algorithm developed at the University of Manchester, comparing its carbon-equivalent costs with those of traditional storage methods. The study highlights the significant carbon footprint associated with ML training and the operational costs of data storage, particularly in the context of large-scale data generation in experiments like those at the Large Hadron Collider. The results indicate that the break-even point is highly sensitive to the carbon intensity of the energy sources used in different countries, and that while ML-based compression can achieve better compression ratios, it may also incur higher carbon costs per unit of data processed compared to standard algorithms. The findings suggest that further exploration is needed to determine if the environmental benefits of ML-based compression justify its deployment at scale.
Methodology
The authors estimated the energy usage of a ML-based data compression algorithm on an Nvidia T4 GPU, considering various country-specific energy mixes to convert energy consumption into carbon-equivalent costs. They compared these costs with the carbon footprint of traditional disk storage methods, specifically HDDs and tape drives, over a five-year lifespan. The break-even point was defined as the dataset size at which the carbon costs of ML training and execution equaled the embodied carbon of the additional storage required for uncompressed data.
Results
The analysis revealed that the break-even point varies significantly based on the carbon intensity of the energy used for ML operations. The study found that tape storage has a lower carbon footprint than HDDs, albeit with slower access speeds. The results also indicated that while the ML-based compression algorithm offers better compression ratios, it incurs higher carbon costs per unit of data processed compared to standard algorithms.
Implications
The findings suggest that as data generation increases, particularly in scientific fields, understanding the environmental impact of data compression methods is crucial. This research could guide future developments in sustainable ML practices and inform decisions on the deployment of ML-based compression techniques in large-scale data environments.
FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration
Large Language Models
Optimization
Efficient ML
- FleetSieve focuses on decision-critical profiling, selecting measurements based on their impact on resource allocation.
- It models capacity and tail latency jointly to ensure configurations meet SLO requirements.
- The method achieves a reduction in profiling GPU-seconds, particularly for chat workloads.
- FleetSieve provides a three-state result for profiling outcomes: certified feasible, undecided, or certified infeasible.
Read more
FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration
Summary
The paper introduces FleetSieve, a novel approach for optimizing the configuration of large language model (LLM) serving fleets by addressing the challenges of selecting tensor-parallel degrees and replica counts under varying loads. Traditional exhaustive profiling methods are inefficient as they often measure configurations that do not impact final resource allocation. FleetSieve improves upon this by focusing on decision-critical measurements that directly influence resource-coupled, service-level objective (SLO)-aware decisions. It jointly models capacity and tail latency, allowing it to compare conservative and optimistic resource allocations. The profiling process stops when the decision gap is below a specified tolerance. Evaluations on a 31B-parameter model demonstrate that FleetSieve achieves significant reductions in GPU-seconds used for profiling compared to uniform random profiling, while ensuring compliance with SLOs. The method is particularly effective for workloads like chat, though it shows less benefit for code-related tasks. FleetSieve's approach emphasizes the importance of selecting measurements that can change the allocation decisions, making it a targeted solution rather than a universal one.
Methodology
FleetSieve employs a decision-focused profiling strategy that prioritizes measurements based on their expected impact on resource allocation decisions. It maintains uncertainty over capacity and tail latency, solving fleet allocation under both conservative and optimistic scenarios. The profiling process continues until the decision gap is minimized below a specified threshold.
Results
FleetSieve demonstrated a reduction of 6.9% in GPU-seconds compared to uniform random profiling on a fixed measurement grid, with a mean saving of 5.4% across 200 random reveal orders. The method achieved a 21.5% saving for chat workloads, while showing no reduction for code tasks. It effectively avoided configurations that would violate SLOs, such as a completion time exceeding 30 seconds.
Implications
FleetSieve's approach can significantly enhance the efficiency of LLM serving systems by optimizing resource allocation based on critical decision-making metrics. This can lead to improved performance and resource utilization in environments with diverse workload demands, making it applicable in various real-world LLM deployment scenarios.
An Inclusive and Lightweight Approach to Federated Continual Learning for Cultural Heritage
Federated Learning
- Introduction of FedCurv-DR, a lightweight FCL strategy for cultural heritage.
- Focus on minimizing communication and computation overhead in federated settings.
- Evaluation on WikiArt dataset shows effective genre classification with evolving styles.
- Demonstrated reduction in forgetting and improved fairness and energy efficiency.
Read more
An Inclusive and Lightweight Approach to Federated Continual Learning for Cultural Heritage
Summary
This paper presents FedCurv-DR, a novel lightweight and regularisation-based Federated Continual Learning (FCL) strategy specifically designed for cultural heritage applications. The authors highlight the challenges posed by the distributed and evolving nature of cultural heritage data, which often face ownership and access restrictions. FedCurv-DR addresses these challenges by allowing multiple institutions to collaboratively train AI models on their private collections without sharing raw data. The method focuses on accumulating parameter-importance estimates across clients to protect learned knowledge while minimizing communication and computation overhead by updating these estimates at fixed intervals. The evaluation of FedCurv-DR is conducted using the WikiArt image dataset, where genre classification is treated as a fixed objective while introducing evolving artistic styles as new experiences. The results demonstrate that FedCurv-DR effectively reduces forgetting, balances performance, fairness, and energy efficiency, aligning with the principles of sustainable AI in cultural heritage.
Methodology
The authors propose FedCurv-DR, which is a regularisation-based FCL framework that aggregates parameter-importance information across clients periodically. This approach avoids the use of replay buffers or synthetic data generation, making it lightweight and efficient for continual learning scenarios.
Results
The evaluation of FedCurv-DR on the WikiArt dataset indicates significant improvements in maintaining performance on past data while adapting to new experiences. The method shows a reduction in forgetting and achieves a balance between performance, fairness, and energy efficiency, making it suitable for sustainable AI applications.
Implications
The findings suggest that FedCurv-DR can facilitate collaborative AI model training in cultural heritage institutions, promoting inclusivity and fairness while addressing ethical concerns related to data ownership and representation. This approach can enhance the accessibility and effectiveness of AI in the cultural heritage sector.
Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution
Multimodal
Robotics
Theory
- Introduces the Disjoint-Bridge Operator-Substitution Certificate (DBOSC) for certifying multimodal representations.
- Demonstrates that audio and acceleration modalities can interchangeably represent the same physical surface.
- Establishes a hierarchy of capabilities: attribute access, response substitution, fusion closure, and ordered execution.
- Shows that the executor's design significantly impacts the performance of ordered execution.
Read more
Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution
Summary
This paper addresses the challenge of certifying the meaning of multimodal representations in physical interactions, proposing a novel framework called the Disjoint-Bridge Operator-Substitution Certificate (DBOSC). The authors argue that existing methods fail to confirm whether different sensor modalities convey the same executable meaning. The DBOSC evaluates whether independently trained modality compilers can interchangeably access a frozen response chart based on evidence outside their training data. The study employs two experiments: the first, using the Cluster Haptic dataset, demonstrates that audio and acceleration representations of unseen surfaces are significantly closer in response space than incorrect pairings. The second experiment tests ordered execution in a controlled elastoplastic system, revealing that the DBOSC serves as an instrument for assessing the executor's capability rather than merely providing a verdict. The findings indicate that while fusion improves performance, distinct achievements in attribute access, response substitution, fusion closure, and ordered execution can be separately validated. This work emphasizes the importance of operational semantics in establishing the meaning of multimodal representations in physical contexts.
Methodology
The authors developed the DBOSC framework to evaluate the interchangeability of modality compilers trained on disjoint evidence panels. They conducted experiments using the Cluster Haptic dataset and a controlled elastoplastic system to assess the performance of different modalities in terms of response substitution and ordered execution.
Results
The results indicate that audio and acceleration modalities are 4.5 times closer in response space for unseen surfaces compared to incorrect pairings. In the elastoplastic system, the DBOSC confirmed that ordered execution was only certified once the executor was adequately trained, with a reported oracle NMSE of 0.18. Fusion of modalities improved performance, but two registered checks failed due to a diagonal restriction in the fused information matrix.
Implications
The findings suggest that the DBOSC can serve as a robust tool for validating multimodal systems in robotics and physical interactions. This work could lead to improved designs of multimodal interfaces that ensure consistent meaning across different sensory inputs, enhancing the reliability of robotic systems in real-world applications.
Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification
NLP
Large Language Models
Theory
- Introduction of TUP, a policy that truncates low-ranked completions and upweights high-ranked ones.
- Theoretical support for the effectiveness of lower-tail truncation in improving performance.
- Demonstration of TUP's competitive performance against established offline alignment methods.
- Decoupling of lower-tail truncation and upper-tail sharpness for better policy training.
Read more
Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification
Summary
This paper introduces TUP (Truncate-bad, Upweight-good Policy), a novel approach to BoN-style distillation that enhances the generation quality of language models by optimizing the selection of completions based on their ranks. Traditional methods retain all completions but apply smooth reweighting, which can lead to inefficiencies and reliance on potentially flawed rankings. TUP addresses this by truncating low-ranked completions entirely from the support and applying a sharper reweighting to the remaining higher-ranked completions. This method allows for a more effective training signal by separating the decisions of how much of the lower-ranked tail to remove and how sharply to upweight the retained completions. The authors provide theoretical justifications for this approach and demonstrate its effectiveness through empirical evaluations against leading offline alignment baselines, showing that TUP can match or exceed their performance while simplifying the training process.
Methodology
The TUP policy employs a shifted-truncated win-rate transform to assign zero mass to completions below a certain win-rate threshold while softly reweighting the retained completions. The training is conducted offline using binary cross-entropy loss, with shifted-truncated win-rates as soft labels and log-likelihood ratios as logits.
Results
TUP was evaluated on the QRPO benchmark using various language models and reward models, showing competitive performance compared to four leading offline alignment baselines. The results indicate that TUP effectively balances the trade-off between retaining useful completions and discarding low-quality ones.
Implications
The proposed TUP policy has significant implications for improving the efficiency and effectiveness of language model training and deployment, particularly in scenarios where computational resources are limited. It provides a framework for better aligning language models with human preferences by focusing on the quality of generated completions.
Time-Uniform Self-Normalized Concentration for Discounted Least Squares: Limits and Corrections
Theory
Reinforcement Learning
- Identifies a critical flaw in the claimed time-uniform self-normalized inequality for discounted least-squares estimators.
- Establishes a lower bound for valid anytime boundaries that must grow with time horizon T.
- Clarifies the distinction between fixed-time and time-uniform validity in concentration inequalities.
- Provides valid corrections for confidence bounds in both finite and infinite horizons.
Read more
Time-Uniform Self-Normalized Concentration for Discounted Least Squares: Limits and Corrections
Summary
This paper addresses the limitations of self-normalized concentration inequalities in the context of discounted least-squares estimators, particularly in non-stationary problems. The author critiques a widely used time-uniform guarantee for these estimators, demonstrating through a scalar Gaussian counterexample that the claimed bounded radius is violated with probability one. The paper establishes that for fixed discount and regularization parameters, any valid anytime boundary must grow at least as fast as Rβlog(T/Ξ΄) for sufficiently large T/Ξ΄ when Ξ΄ β€ 1/2. The author identifies a proof error in the existing literature, clarifying that different terminal times utilize distinct Gaussian mixing distributions, which invalidates the proposed stopping-time argument. Despite these issues, the paper confirms that the weighted inequality remains valid at fixed deterministic times and provides corrections for both finite- and infinite-horizon scenarios. The implications of these findings are significant for downstream analyses in bandit and reinforcement-learning literature, as they refine the understanding of confidence bounds in non-stationary settings.
Methodology
The author employs a counterexample approach using a scalar Gaussian distribution to demonstrate the failure of the claimed time-uniform inequality. The analysis includes deriving lower bounds for valid anytime boundaries and revisiting the proof of classical self-normalized inequalities to establish valid extensions.
Results
The paper shows that the claimed time-uniform concentration inequality is incorrect and provides a rigorous lower bound for the growth of valid anytime boundaries. It confirms that the weighted inequality holds at fixed deterministic times and offers corrections for both finite and infinite horizons.
Implications
The findings have significant implications for the design and analysis of algorithms in bandit and reinforcement-learning frameworks, particularly in non-stationary environments where parameters change over time. The corrections provided can enhance the reliability of confidence bounds used in these analyses.
Inadvertent Context Leakage in Language Models
NLP
Large Language Models
Theory
- Inadvertent context leakage can occur in LLMs, allowing sensitive information to be reconstructed from benign outputs.
- Controlled experiments show high accuracy in reconstructing 2-digit and 4-digit numeric secrets from model outputs.
- More capable models are more susceptible to leakage, indicating a correlation between model capability and privacy risks.
- The study introduces a formal threat model for understanding and mitigating inadvertent context leakage.
Read more
Inadvertent Context Leakage in Language Models
Summary
This paper investigates the phenomenon of inadvertent context leakage in large language models (LLMs), particularly in scenarios where sensitive user information is present in the model's context window. The authors explore whether benign outputs from LLMs can inadvertently encode sensitive information, even when the model refuses direct extraction attempts. Through controlled experiments on eight proprietary models, they demonstrate that adaptive adversaries can reconstruct numeric secrets with high accuracy from seemingly innocuous outputs. The study reveals that more capable models exhibit greater leakage, suggesting that this issue is a byproduct of model capability rather than a fixable bug. The authors propose a threat model formalizing this leakage as a predicate inference game and present two practical attack scenarios: inferring semantic predicates about user memories and extracting Social Security Numbers from a production-style agent. The findings highlight the need for improved privacy measures in LLM applications.
Methodology
The authors conducted controlled experiments on eight proprietary language models, using benign prompts to assess the models' ability to leak sensitive information. They employed an adaptive adversary model that learns to exploit specific channels of leakage unique to each model. The study also analyzed the effects of suppression mechanisms and the relationship between model capability and leakage.
Results
The experiments revealed that 2-digit secrets could be reconstructed with near-perfect accuracy, while 4-digit secrets achieved an 82% exact match. The study found that conditioning prompts on previously recovered digits significantly improved recovery rates. Additionally, the authors demonstrated that adversaries could infer semantic predicates about user memories and successfully extract Social Security Numbers from model outputs.
Implications
The findings underscore the importance of addressing privacy concerns in LLM applications, particularly as these models are increasingly used in sensitive contexts. The proposed threat model and insights into leakage mechanisms can inform the development of more robust privacy-preserving techniques in AI systems.