AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
SCALE: State-Calibrated Latent Embeddings for JEPA Planning in the Right Geometry
Reinforcement Learning
Robotics
Optimization
- SCALE improves the geometric properties of latent embeddings for better planning.
- The method correlates latent distances with task-relevant state differences.
- SCALE consistently outperforms LeWM across various tasks and compute budgets.
- Decodability of task-relevant information and its geometric influence are distinct properties.
Read more
SCALE: State-Calibrated Latent Embeddings for JEPA Planning in the Right Geometry
Summary
The paper introduces SCALE (State-CAlibrated Latent Embeddings), a novel approach to enhance the planning capabilities of joint-embedding predictive world models (JEPA) by improving the geometric properties of latent embeddings. The authors identify that while existing models like DINO-WM and LeWorldModel (LeWM) effectively retain task-relevant state information, they differ in how this information influences planning. DINO-WM's latent distances better reflect variations in task state, which is crucial for effective planning. SCALE addresses this by introducing a training objective that aligns pairwise latent distances with distances in a standardized task-relevant state space, without altering LeWM's learned encoder. The empirical results demonstrate that SCALE consistently outperforms LeWM across multiple tasks and planning solvers, highlighting the importance of not just encoding task-relevant information but also ensuring it shapes the geometry used by planners.
Methodology
SCALE employs a state-calibration regularizer that reshapes the latent-space geometry to ensure that pairwise latent distances reflect task-relevant state differences. This is achieved while retaining the existing prediction and anti-collapse objectives of LeWM, thus enhancing the model's planning capabilities without significant computational overhead.
Results
The empirical evaluation shows that SCALE improves all 15 task-solver averages over LeWM across five different tasks, three planning solvers, and five compute budgets. A control experiment using a latent-to-state regression objective matched or exceeded SCALE's decodability but failed to achieve consistent planning improvements, underscoring the effectiveness of SCALE's geometric intervention.
Implications
The findings suggest that enhancing the geometric properties of latent embeddings can significantly improve planning in reinforcement learning and predictive modeling contexts. This could lead to more efficient and effective planning algorithms in various applications, including robotics and autonomous systems.
One Residual with Three Reuses: A Wristband Front End for Gesture Sensing
Robotics
Efficient ML
Multimodal
- Integration of IMU and FMCW radar for robust gesture sensing.
- Single on-chip residual generator used for multiple functions, enhancing efficiency.
- Achieved high detection probabilities with low false alarm rates.
- Significant energy savings in classifier invocations.
Read more
One Residual with Three Reuses: A Wristband Front End for Gesture Sensing
Summary
This paper presents a novel wristband architecture designed for continuous hand gesture sensing and motor symptom monitoring. The proposed system integrates a MEMS inertial measurement unit (IMU) with a 60 GHz frequency-modulated continuous-wave (FMCW) radar, aiming to operate within a coin-cell power budget while maintaining robustness against occlusion and on-body drift. The architecture features a single on-chip residual generator that facilitates classifier wake-up gating, modality routing, and innovation-based measurement reweighting in an extended Kalman filter (EKF). This design allows for significant energy savings and improved detection accuracy. The implementation runs on an Ambiq Apollo4 Blue Plus microcontroller, demonstrating a detection probability of 0.72/0.80 at a 1% false alarm rate across multiple datasets, while achieving a 47% reduction in classifier invocation energy at 90% gesture detection recall. The study highlights the potential of combining multiple functions into a single residual generator, optimizing both performance and energy efficiency in wearable gesture sensing systems.
Methodology
The wristband architecture employs a recurrent forecaster to predict conditioned IMU and radar streams, generating a one-step prediction error used for classifier wake-up gating, modality selection, and EKF measurement reweighting. The system is designed to operate efficiently on a low-power microcontroller, with a focus on minimizing energy consumption while maximizing detection accuracy.
Results
The proposed wristband achieved a detection probability of 0.72 on the IPN Hand dataset and 0.80 on the SHREC 2021 dataset at a 1% false alarm rate. The architecture demonstrated a 47% reduction in classifier invocation energy at 90% gesture detection recall, and it significantly outperformed previous methods in terms of detection accuracy and energy efficiency.
Implications
The findings suggest that integrating multiple sensing modalities and functions into a single architecture can lead to more efficient wearable devices for gesture recognition and health monitoring. This approach could pave the way for advanced applications in rehabilitation, clinical assessments, and user interface design.
Hoeffding adaptive splitting trees for data stream classification with concept drift and ensemble learning
Theory
Efficient ML
Time Series
- Introduction of Hoeffding Adaptive Splitting Trees (HASTs) that enhance ensemble diversity and performance.
- Identification of limitations in existing adaptive-splitting decision trees as ensemble base learners.
- Comprehensive experimental analysis showing HASTs achieve state-of-the-art results in data stream classification.
- Insights into the computational costs and adaptability of the proposed models in the presence of concept drift.
Read more
Hoeffding adaptive splitting trees for data stream classification with concept drift and ensemble learning
Summary
This paper addresses the challenges of data stream classification, particularly in the context of concept drift and ensemble learning. Traditional Hoeffding Trees, commonly used as base learners in ensemble methods, employ a periodic splitting mechanism that has been found to lack adaptability to changes in data distribution. The authors propose two novel decision tree models, termed Hoeffding Adaptive Splitting Trees (HASTs), which combine the strengths of periodic splitting with adaptive mechanisms that respond to performance degradation. These models utilize change detection algorithms to identify when splits should occur, thereby enhancing the diversity of the ensemble and improving overall performance. The paper presents a comprehensive experimental evaluation, demonstrating that HASTs outperform existing methods in terms of accuracy, computational efficiency, and adaptability to concept drift. The findings suggest that HASTs can significantly improve ensemble learning in dynamic environments, paving the way for future research in adaptive learning systems.
Methodology
The authors developed two new decision tree architectures that integrate a periodic splitting strategy with adaptive mechanisms. These trees utilize change detection algorithms to monitor performance and determine optimal split points. A comprehensive experimental protocol was employed to evaluate the performance of HASTs against existing ensemble methods, focusing on accuracy, computational cost, and adaptability to concept drift.
Results
The experimental results indicate that Hoeffding Adaptive Splitting Trees significantly enhance ensemble performance, achieving state-of-the-art results across various benchmarks. The models demonstrated improved adaptability to concept drift and reduced computational overhead compared to traditional methods.
Implications
The proposed HASTs can be applied in real-time data stream classification scenarios, particularly in domains where concept drift is prevalent, such as finance, healthcare, and online services. The findings may influence the design of future adaptive learning systems and ensemble methods.
Resource-Adaptive Primal-Dual Learning for One-Warehouse Multi-Store Systems with Censored Demand
Optimization
Theory
- Introduction of Resource-Adaptive Primal-Dual Learning (RAPDL) framework for OWMS systems.
- Dynamic tracking of the Primal-Dual re-solving path improves adaptability to changing resource states.
- Achieves logarithmic expected regret, outperforming existing square-root-order guarantees.
- Numerical evaluations show RAPDL's superior performance in cost reduction compared to established benchmarks.
Read more
Resource-Adaptive Primal-Dual Learning for One-Warehouse Multi-Store Systems with Censored Demand
Summary
This paper introduces a novel learning framework called Resource-Adaptive Primal-Dual Learning (RAPDL) for managing inventory in One-Warehouse Multi-Store (OWMS) systems with censored demand. Traditional OWMS learning policies rely on fixed targets based on initial resource rates, which fail to adapt as sales change the available inventory. RAPDL addresses this limitation by dynamically tracking the Primal-Dual re-solving path, allowing for real-time adjustments to store allocations based on the evolving resource state. The framework utilizes remaining inventory and censored sales data to update allocation targets and dual variables, leading to a logarithmic expected regret guarantee, which is a significant improvement over existing square-root-order guarantees. The analysis of RAPDL separates regret into three components, each contributing logarithmically to the overall regret. Numerical experiments demonstrate that RAPDL consistently outperforms the state-of-the-art Double Binary Search policy across various inventory scenarios, indicating its effectiveness in finite-horizon settings.
Methodology
The methodology involves developing a resource-adaptive learning framework that replaces fixed-target learning with online tracking of the Primal-Dual re-solving path. It uses remaining inventory and censored sales to update allocation targets and dual variables, ensuring that the learning process adapts to the evolving state of resources without needing to estimate demand distributions.
Results
The results indicate that RAPDL achieves lower mean total costs in all 24 paired comparisons against the Double Binary Search policy, demonstrating its effectiveness across both homogeneous and heterogeneous inventory regimes. The logarithmic regret guarantee signifies a substantial improvement over existing methods in the literature.
Implications
The findings suggest that RAPDL can significantly enhance inventory management in various applications, including general merchandise distribution, perishable product logistics, and emergency supply allocation. The framework's adaptability may also inform other online learning and control problems involving shared resources.
Do Uncertainty Signals Help? A Systematic Study of Uncertainty-Aware Decoding with Rollback Mechanisms
NLP
Large Language Models
- Introduces uncertainty-aware rollback decoding as a strategy to improve LLM-based code generation.
- Demonstrates that the complete rollback framework significantly enhances performance compared to traditional methods.
- Finds that information-theoretic uncertainty signals outperform sampling-based alternatives in guiding rollback.
- Identifies feedback-guided rollback as the main contributor to performance improvements.
Read more
Do Uncertainty Signals Help? A Systematic Study of Uncertainty-Aware Decoding with Rollback Mechanisms
Summary
This paper investigates the role of uncertainty signals in enhancing code generation through a novel approach called uncertainty-aware rollback decoding. The authors highlight that while prediction uncertainty is a valuable metric for model confidence, its application in large language models (LLMs) for code generation has been underexplored. The proposed framework utilizes uncertainty signals to identify unreliable generation regions, allowing the model to roll back to earlier valid prefixes without requiring retraining. The study evaluates this framework across seven code LLMs and five code generation benchmarks, employing eight different token-level uncertainty signals. The findings indicate that the complete rollback framework outperforms equal-budget restart methods, achieving significant improvements in pass rates and safety metrics. Notably, information-theoretic measures like token entropy and negative log-likelihood consistently yield the best results. The authors also conduct ablation studies to assess the contributions of rollback mechanisms and uncertainty localization, revealing that feedback-guided rollback is the primary driver of improvement. Overall, this research provides a comprehensive empirical analysis of uncertainty in code generation and establishes a new decoding strategy that mitigates error propagation.
Methodology
The authors developed a unified framework for uncertainty-aware rollback decoding, which incrementally generates code, checks for errors, and rolls back to earlier prefixes based on uncertainty signals. They compared various uncertainty measures, including information-theoretic and sampling-based signals, across multiple benchmarks and model scales. The study involved controlled ablation experiments to isolate the effects of different components of the framework.
Results
The complete rollback framework showed improvements over equal-budget restart methods, with gains of up to 0.26 in pass@1 and 0.35 in AvgTestPassRate on functional code generation benchmarks. Additionally, an absolute improvement of up to 6.4% in Patch-Aligned Safe Rate was observed on the Dsec-Python benchmark. Information-theoretic signals consistently provided better performance compared to sampling-based signals.
Implications
The findings suggest that incorporating uncertainty signals into code generation processes can significantly enhance model reliability and performance. This approach may have broader applications in areas where precision and correctness are critical, such as software development and automated programming tools.
Training Fair Tabular Foundation Models
Theory
- Introduces FairTFM, a scalable training strategy for fair predictions in TFMs.
- Incorporates fairness as a first-class objective during TFM pretraining.
- Employs synthetic fairness tasks and a fairness-aware architecture to enhance model fairness.
- Demonstrates significant improvements in fairness metrics with minimal accuracy trade-offs.
Read more
Training Fair Tabular Foundation Models
Summary
This paper addresses the fairness properties of Tabular Foundation Models (TFMs), which have gained popularity for tabular predictive tasks due to their ability to leverage in-context learning (ICL). Despite their effectiveness, the fairness of TFMs in high-stakes decision-making contexts has not been thoroughly investigated. The authors propose FairTFM, a novel training framework that incorporates fairness constraints directly into the TFM training process, allowing for fair predictions in a single forward pass. The approach tackles two main challenges: limited access to sensitive attributes in training data and the incompatibility of existing fairness techniques with the ICL paradigm. FairTFM employs synthetic fairness task generation and a fairness-aware architecture that utilizes a gradient reversal layer to encourage the model to learn representations invariant to sensitive attributes. The authors evaluate their method on 132 fairness tasks, demonstrating that FairTFM consistently improves fairness metrics while maintaining competitive accuracy, thus establishing a new standard for fairness in TFMs.
Methodology
The authors developed a framework that integrates fairness constraints into the training of TFMs. This includes generating synthetic fairness tasks by designating input features as sensitive attributes and employing a fairness-aware architecture that incorporates a gradient reversal layer. This architecture enables the model to learn representations that are invariant to sensitive attributes, facilitating fair predictions during inference.
Results
The FairTFM framework was evaluated on 132 fairness tasks, showing improvements in fairness metrics by 32–75% relative to the strongest baseline TFM, with accuracy costs ranging from 2–11%. The results indicate that FairTFM effectively balances fairness and predictive performance.
Implications
The findings suggest that FairTFM can be applied in high-stakes decision-making scenarios, such as finance and healthcare, where fairness is critical. By enabling fair predictions without the need for post-hoc corrections, this approach could enhance the ethical deployment of machine learning models in sensitive applications.
A Privacy Study of Sparse Collaborative Inference
Computer Vision
Federated Learning
Theory
- Sparsification reduces communication costs but does not significantly lower privacy risks.
- The positions of sparse activations are a major source of privacy leakage, enabling input reconstruction and re-identification.
- A novel probing framework is introduced to measure the contribution of values and positions to privacy leakage.
- Privacy risks persist even at low transmission rates and task utility levels.
Read more
A Privacy Study of Sparse Collaborative Inference
Summary
This paper investigates the privacy implications of Sparse Collaborative Inference (CI), a method that splits model computations between edge devices and servers. While sparsification is often employed to reduce communication costs and enhance privacy, the authors reveal that the privacy benefits of sparsity are overstated. They decompose the sparse activations into retained values and their positions, demonstrating that the positions carry significant privacy risks, enabling adversaries to reconstruct inputs and re-identify individuals. The study employs a probing framework to isolate and measure the leakage from both components, revealing that while sparsification reduces transmission costs, it does not proportionately decrease privacy risks. The findings indicate that privacy auditing must account for the sensitive nature of positional data in sparse activations, which can lead to serious compliance issues under regulations like GDPR.
Methodology
The authors developed a probing framework that isolates the retained values and positions of sparse activations to measure their respective contributions to privacy leakage. They conducted experiments across various datasets and architectures, employing both optimization audits and learned inverse attacks to assess the privacy risks associated with each component.
Results
The study found that the positions of sparse activations account for the majority of privacy leakage, enabling high-fidelity reconstructions and re-identifications. The leakage from positions persists even when transmission costs and task utility are low, indicating a disconnect between communication efficiency and privacy protection.
Implications
The findings suggest that privacy auditing in collaborative inference must rigorously evaluate the sensitivity of positional data. This has implications for the design of privacy-preserving machine learning systems, particularly in contexts where compliance with data protection regulations is critical.
QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction
Large Language Models
Optimization
Efficient ML
- QUASAR continuously minimizes loss-aware reconstruction error during QAT to improve training trajectories.
- The method uses saliency estimates and searches for optimal clipping ranges to enhance model quality.
- Empirical results indicate QUASAR achieves lower KL divergence and higher accuracy than competitive QAT methods.
- QUASAR supports standard deployment formats without requiring changes at inference time.
Read more
QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction
Summary
The paper introduces QUASAR, a novel method for quantization-aware training (QAT) aimed at improving model quality during low-precision inference. Traditional QAT suffers from a mismatch between the loss computed using lossy reconstructions of full-precision weights and the updates applied to latent weights, leading to suboptimal training trajectories and a higher loss floor. QUASAR addresses this issue by continuously performing lightweight, loss-aware reconstruction during the training process. It utilizes an exponential moving average of squared gradients to estimate saliency, searches over a small set of clipping ranges, and fits affine dequantizers using saliency-weighted least squares. The authors demonstrate that minimizing the loss-aware reconstruction error is crucial for controlling the loss of the final quantized model. QUASAR modifies only the training procedure, supporting standard deployment formats without inference-time changes. Empirical results show that QUASAR significantly reduces KL divergence and improves accuracy across various tasks compared to existing QAT methods.
Methodology
QUASAR employs a two-stage reconstruction process involving quantization and dequantization, utilizing saliency-weighted least squares to fit dequantizers. It approximates the Hessian using an exponential moving average of squared gradients to derive per-parameter saliency scores, allowing for efficient optimization of clipping ranges and reconstruction parameters during training.
Results
QUASAR consistently outperforms existing QAT methods, achieving the lowest held-out KL divergence across various bit widths (2, 3, and 4 bits) and improving average accuracy by 3.5–4.3 percentage points at 2 bits. It reduces KL divergence by at least 10% at 3 and 4 bits and approximately 30% when applied to NVFP4.
Implications
The findings suggest that QUASAR can significantly enhance the deployment of large language models in low-precision formats, ensuring better performance and quality. This method could be pivotal for applications requiring efficient inference in resource-constrained environments.
Probabilistic indirect models for undrained shear strength: addressing significant data missing and variability with advanced imputation and machine learning techniques
Theory
Optimization
Efficient ML
- Developed probabilistic indirect models for predicting undrained shear strength (su) using a global database.
- Evaluated three imputation methods to address high missing data rates and variability.
- Introduced a novel MHA-PNN model that enhances information extraction from limited data.
- Demonstrated superior performance of the MN-enhanced MHA-PNN model in prediction accuracy and uncertainty quantification.
Read more
Probabilistic indirect models for undrained shear strength: addressing significant data missing and variability with advanced imputation and machine learning techniques
Summary
This paper addresses the challenge of accurately predicting undrained shear strength (su) in geotechnical engineering, which is often hindered by significant data missingness and variability. Utilizing the CLAY/10/7490 global database, the authors develop probabilistic indirect models that leverage Atterberg limits and piezocone cone penetration (CPTU) measurements. The study first evaluates three imputation methods—multivariate normal (MN), multiple imputation by chained equations (MICE), and miss forest (MF)—to handle the high missing data rate in the dataset. A Probabilistic Extreme Gradient Boosting (PXGB) model is then constructed to assess the effectiveness of these imputation techniques by comparing the model's performance on imputed datasets versus the original incomplete data. Furthermore, the authors introduce a novel model that integrates a multi-head attention (MHA) mechanism into an artificial neural network (ANN), resulting in the MHA-based probabilistic neural networks (MHA-PNN) model. The performance of the MHA-PNN model is evaluated against a conventional MN-based prediction model using metrics such as root mean square error (RMSE), coefficient of determination (R²), mean absolute percentage error (MAPE), conditional interval width (wCI), and coverage rate (CR). The results indicate that the MN-enhanced MHA-PNN model significantly outperforms other models in both prediction accuracy and uncertainty quantification, showcasing the potential of this integrated strategy for robust probabilistic modeling in geotechnical applications, especially in the context of sparse and incomplete datasets.
Methodology
The study employs advanced imputation techniques (MN, MICE, MF) to fill missing data in the CLAY/10/7490 database. A PXGB model is developed to validate the effectiveness of these imputation methods. Additionally, a multi-head attention mechanism is integrated into an ANN to create the MHA-PNN model, which is then compared against conventional models using various performance metrics.
Results
The MN-enhanced MHA-PNN model outperformed other models in terms of prediction accuracy and uncertainty quantification, as evidenced by improved RMSE, R², MAPE, wCI, and CR metrics. This indicates the effectiveness of the proposed approach in handling incomplete and variable datasets.
Implications
The findings suggest that integrating advanced imputation techniques with machine learning can significantly improve the reliability of geotechnical predictions, particularly in scenarios with sparse data. This approach could be applied to enhance modeling in various geotechnical applications, leading to better design and risk assessment.
Hybrid Quantum-inspired Kolmogorov-Arnold Networks for Privacy-Aware Federated Biosignal Learning
Federated Learning
Time Series
Efficient ML
- HQKAN improves classification performance on ECG data compared to MLP.
- The model uses 37.35% fewer trainable parameters and reduces communication costs by 24.89% on the MIT-BIH dataset.
- HQKAN is designed to handle client-side data limitations and class imbalances effectively.
- The study demonstrates the feasibility of quantum-inspired models in practical federated learning scenarios.
Read more
Hybrid Quantum-inspired Kolmogorov-Arnold Networks for Privacy-Aware Federated Biosignal Learning
Summary
This paper presents a novel approach to federated learning for electrocardiogram (ECG) classification using a hybrid quantum-inspired Kolmogorov-Arnold network (HQKAN). The authors address the challenges of privacy-sensitive biosignal data, which limits the sharing of raw ECG signals for centralized model training. Federated learning allows for collaborative model training while keeping data local, but it faces issues such as limited client-side samples, imbalanced arrhythmia labels, and non-IID data distributions. The HQKAN model is evaluated against a multilayer perceptron (MLP) for classifying arrhythmias in two datasets: MIT-BIH and INCART. The results show that HQKAN outperforms MLP in terms of aggregate and minority-class metrics while being more parameter-efficient and reducing communication costs significantly. This indicates that HQKAN is a robust alternative for privacy-aware federated learning in biosignal applications.
Methodology
The authors employed a hybrid quantum-inspired Kolmogorov-Arnold network (HQKAN) architecture, which combines a fully connected encoder and decoder with a QKAN latent feature processor. The model was tested under federated averaging (FedAvg) across multiple client configurations, utilizing two datasets (MIT-BIH and INCART) with varying client counts and data distributions (IID and non-IID).
Results
HQKAN achieved superior performance metrics across both datasets, including higher macro-F1 scores and lower Brier scores, while using significantly fewer parameters than the MLP baseline. Specifically, on the MIT-BIH dataset, HQKAN used 11,581 parameters compared to 18,485 for MLP, and on INCART, it used 15,147 parameters compared to 27,443 for MLP. Communication costs were also reduced by 24.89% on MIT-BIH and 36.41% on INCART.
Implications
The findings suggest that HQKAN can be effectively utilized in privacy-sensitive applications such as healthcare, where federated learning is essential for maintaining data confidentiality while still enabling collaborative model training. This approach could enhance the deployment of machine learning in medical diagnostics and monitoring systems.
Post-training Quantization for Hybrid Iterative Generative Models
Generative Models
Efficient ML
- Introduces HyGenQ, a PTQ framework specifically designed for hybrid IGMs.
- Addresses the challenges of Excessive Outliers and Amplified Anomalies in quantization.
- Utilizes Hierarchical Cluster Decoupling to effectively manage outlier channels.
- Implements Scaling Recalibration to stabilize activation fluctuations during inference.
Read more
Post-training Quantization for Hybrid Iterative Generative Models
Summary
This paper addresses the challenges of applying Post-training Quantization (PTQ) to Hybrid Iterative Generative Models (IGMs), which combine autoregressive and diffusion paradigms for high-fidelity image generation. The authors identify two main issues: Excessive Outliers (EOs) in activations that hinder effective quantization and Amplified Anomalies (AAs) that arise from quantization errors, leading to model collapse. To tackle these challenges, they propose a novel PTQ framework called HyGenQ, which includes two key components: Hierarchical Cluster Decoupling (HCD) and Scaling Recalibration (SR). HCD isolates outlier channels through a multi-stage clustering process, allowing for independent quantization of normal and outlier values, thus preserving generation quality. SR dynamically adjusts the scaling of anomalies based on a Gaussian Bound, preventing instability during inference. The experimental results demonstrate that HyGenQ can quantize hybrid IGMs to 8-bit precision while significantly outperforming existing methods, showcasing its robustness across various model families.
Methodology
The methodology involves two main components: Hierarchical Cluster Decoupling (HCD) for isolating outlier channels through a multi-stage clustering process, and Scaling Recalibration (SR) for dynamically adjusting the scaling of anomalies based on their distribution, thereby maintaining stability during the iterative generation process.
Results
Extensive experiments show that HyGenQ successfully quantizes hybrid IGMs to 8-bit precision (W8A8) and significantly outperforms existing baselines, validating its effectiveness and robustness across different model families.
Implications
The proposed HyGenQ framework has the potential to enhance the efficiency of hybrid IGMs, making them more feasible for real-time applications in image generation and other domains requiring high-fidelity outputs with reduced computational overhead.
A Banach-Space Theory of Markovian Halpern Iteration for Non-Expansive Maps
Theory
Optimization
Reinforcement Learning
- Introduces a variance-reduced PAGE-Halpern method for non-expansive operators under Markovian sampling.
- Achieves O(ϵ−3) sample complexity in finite-dimensional Banach spaces.
- Establishes high-probability guarantees for the proposed methods using auxiliary smooth norms.
- Extends the analysis of Halpern iteration to non-smooth geometries relevant in reinforcement learning.
Read more
A Banach-Space Theory of Markovian Halpern Iteration for Non-Expansive Maps
Summary
This paper investigates the stochastic approximation of fixed points for non-expansive operators using a Markovian framework. The authors present a block-minibatch implementation of Halpern iteration, which achieves an expected last-iterate residual of O(log N/N) but incurs a high complexity of ˜O(ϵ−5) in terms of Markovian samples. To address this, they propose a variance-reduced PAGE-Halpern method that utilizes refresh and same-state difference blocks, analyzed through the Poisson equation. The findings extend to finite-dimensional Banach spaces, demonstrating that the displacement-level Halpern bound can yield a sample complexity of ˜O(ϵ−3) without the need for restarting the Markov chain or imposing stationarity at block boundaries. The paper also provides a high-probability guarantee for the Banach result, leveraging an auxiliary smooth norm to ensure martingale concentration. Overall, the work significantly enhances the understanding of stochastic fixed-point iterations in non-expansive settings, particularly in reinforcement learning contexts.
Methodology
The authors develop a Markovian block baseline and a PAGE-Halpern method, analyzing the performance through perturbation-level and displacement-level bounds. They utilize Poisson equation decomposition to derive sample complexity results and extend their findings to Banach spaces, employing norm smoothing techniques to handle non-smooth geometries.
Results
The paper shows that the PAGE-Halpern method achieves a sample complexity of ˜O(ϵ−3) in finite-dimensional Banach spaces, significantly improving upon the traditional O(ϵ−5) complexity associated with standard Halpern iteration. Additionally, it establishes high-probability guarantees for the accuracy of the estimator, demonstrating robustness in the presence of Markovian noise.
Implications
The findings have significant implications for stochastic optimization and reinforcement learning, particularly in scenarios where non-expansive operators are prevalent. The methods developed can enhance the efficiency and reliability of algorithms that rely on fixed-point iterations in noisy environments.
Measuring Structured Predictability in Neural Training Dynamics: A Cross-Regime Study
Optimization
Theory
Efficient ML
- Introduces a novel framework for measuring short-horizon temporal structure in neural training dynamics.
- Identifies distinct dynamics between auxiliary and bulk parameters, with the latter containing localized predictable pockets.
- Demonstrates that probe readings provide a state-sensitive view of optimization, revealing role-specific temporal events.
- Highlights the influence of architecture and training recipes on the predictability of parameter updates.
Read more
Measuring Structured Predictability in Neural Training Dynamics: A Cross-Regime Study
Summary
This paper investigates the temporal organization of neural training dynamics, focusing on short-horizon predictability as a measure of temporal redundancy in parameter updates. The authors introduce a multiscale framework utilizing three probe families—displacement-direction, subspace-residual, and predictor-based probes—to analyze the predictability of parameter motion during training. The study is conducted on multi-pass vision training using CIFAR and one-pass pretraining of the Pythia model. Results reveal that auxiliary parameters exhibit simpler dynamics compared to bulk parameters, which display localized, time-varying predictable pockets. The findings indicate that the short-horizon predictability is influenced by architecture and training conditions, providing a retrospective diagnostic tool for understanding training dynamics and optimizing predictive methods.
Methodology
The authors developed three complementary probe families to measure short-horizon predictability: displacement-direction probes assess the persistence of recent movement directions, subspace-residual probes evaluate whether future displacements remain within the span of recent updates, and predictor-based probes fit local dynamics to capture richer temporal structures. These probes were applied to analyze multi-pass vision training on CIFAR and one-pass pretraining of the Pythia model, using convention-aware and null-calibrated readouts for comparison.
Results
The study found that auxiliary parameters exhibit simpler short-horizon dynamics compared to bulk parameters, which are largely unpredictable in aggregate but contain localized predictable pockets that shift over time. The results also showed that the structure of predictability is influenced by the optimization position, architecture, and training recipe, with distinct temporal events observed across different parameter roles.
Implications
The findings suggest that understanding short-horizon predictability can enhance predictive acceleration methods in neural training, allowing for more efficient optimization strategies. This research could inform the design of training algorithms that leverage temporal redundancy for improved performance in deep learning models.
CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing
NLP
Large Language Models
Generative Models
- CForce enhances early-stage prediction reliability in dLLMs under aggressive parallel decoding.
- The method aligns predictions from different stages using self-rollout trajectories.
- Confidence Adaptive KL Divergence is introduced as a novel distillation objective.
- CForce shows significant improvements in speed-quality trade-offs for both non-edit and edit-capable models.
Read more
CForce: Boosting Parallel Decoding for dLLMs via Consistency Forcing
Summary
The paper introduces Consistency Forcing (CForce), a novel distillation method aimed at enhancing the reliability of early-stage predictions in diffusion large language models (dLLMs) during aggressive parallel decoding. Traditional dLLMs generate text by predicting multiple masked tokens in a single forward pass, but they often struggle with unreliable predictions at early denoising stages, which can lead to errors propagating through later stages. CForce addresses this by aligning early-stage mask predictions with those from later stages, utilizing self-rollout trajectories for training. The method employs Confidence Adaptive KL Divergence as a distillation objective, which balances the benefits of forward and reverse KL divergence. Theoretical analysis supports the design, indicating that early-stage prediction errors can be minimized through this approach. The experiments demonstrate that CForce significantly improves the speed-quality trade-off in both non-edit and edit-capable dLLMs, particularly under high-parallelism conditions, leading to enhanced performance metrics across various benchmarks.
Methodology
CForce trains dLLMs on self-rollout trajectories, aligning early-stage predictions with later-stage predictions using Confidence Adaptive KL Divergence and a cross-entropy anchor. The method partitions trajectories into stages based on revealed masked tokens, allowing for efficient training without relying on a frozen teacher model.
Results
The implementation of CForce resulted in an increase in average Tokens Per Forward (TPF) from 6.94 to 9.08 for the edit-capable LLaDA2.1-mini model, while improving accuracy from 85.57 to 86.41. For the non-edit LLaDA2.0-mini model, TPF increased from 3.60 to 6.42, demonstrating a clear speed-quality trade-off.
Implications
CForce has the potential to significantly enhance the efficiency and reliability of text generation in dLLMs, making it applicable in various natural language processing tasks where high-speed and high-quality outputs are critical. This could lead to advancements in applications such as conversational agents, content generation, and code synthesis.
TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH
Time Series
Reinforcement Learning
Optimization
- TRACE-CASH is a hybrid optimizer that combines history-conditioned candidate generation with structured search rules.
- The method addresses the challenges of time-series forecasting by automating model and hyperparameter selection.
- TRACE-CASH outperforms six alternative search methods in a controlled evaluation across diverse time-series tasks.
- The approach emphasizes the importance of trial history in guiding the search process for optimal configurations.
Read more
TRACE-CASH: Trial-History-Conditioned Reinforcement Learning for Adaptive Configuration Exploration in Time-Series CASH
Summary
The paper introduces TRACE-CASH, a novel hybrid sequential optimizer designed for the combined algorithm selection and hyperparameter optimization (CASH) problem in time-series forecasting. The authors highlight the challenges posed by the heterogeneity of time-series tasks, which necessitates an automated approach to model and hyperparameter selection. TRACE-CASH employs a task-local mechanism that utilizes trial history to guide the search process. It integrates a grouped actor-critic framework for candidate generation, where different actors are responsible for proposing model configurations based on previous outcomes. The methodology emphasizes model coverage, validation-guided exploitation, and exploration strategies to recover from stalled progress. The authors conduct a comprehensive evaluation comparing TRACE-CASH against six other search methods across 41 dataset-frequency task variants. The results demonstrate that TRACE-CASH consistently outperforms its competitors in terms of mean absolute scaled error (MASE) and weighted quantile loss (WQL), establishing its effectiveness in adaptive configuration exploration for time-series forecasting.
Methodology
TRACE-CASH utilizes a grouped actor-critic framework where a model actor proposes an initial forecasting model, and three model-conditioned actors generate actions related to temporal, architectural, and training decisions. The method is reinitialized for each task-seed run and updates its search strategy based on accumulated trial history, employing fixed rules for model coverage and validation-guided exploitation.
Results
TRACE-CASH achieved the lowest mean rank on both MASE and WQL metrics across the evaluated methods. It also demonstrated the lowest window-averaged test-MASE rank in predefined evaluation windows, indicating its superior performance in time-series forecasting tasks.
Implications
The findings suggest that TRACE-CASH can significantly enhance the efficiency and effectiveness of automated machine learning systems in time-series forecasting, potentially leading to better predictive performance across various applications in finance, healthcare, and environmental monitoring.
Amortised Post-Hoc Explanation with Exact Preservation for Dynamic Graph Anomaly Detectors
Graph Learning
Interpretability
- Introduction of X-StrGNN, the first post-hoc explanation layer for StrGNN, providing dual attributions for flagged edges.
- Exact preservation of detection performance with no degradation in accuracy metrics.
- Efficient attribution computation at 0.66 ms per edge, significantly faster than traditional methods.
- First controlled design study comparing attribution strategies, demonstrating X-StrGNN's superior stability.
Read more
Amortised Post-Hoc Explanation with Exact Preservation for Dynamic Graph Anomaly Detectors
Summary
This paper addresses the critical need for explainability in dynamic graph anomaly detection, particularly in high-stakes applications such as financial fraud analysis and intrusion detection. The authors introduce X-StrGNN, a post-hoc explanation layer that enhances the existing StrGNN model, which is known for its strong performance in anomaly detection but lacks any form of explanation for its predictions. X-StrGNN provides two types of attributions for flagged edges: structural attribution, which identifies the relevant contextual interactions within the enclosing subgraph, and temporal attribution, which indicates the historical snapshot that contributed to the decision. Importantly, this explanation layer maintains detection accuracy with no loss in performance, as verified by metrics showing no change in AUC, AP, or precision. The authors also present an amortised spatio-temporal parameterisation that allows for efficient computation of attributions, achieving a processing time of 0.66 ms per flagged edge, making it feasible to explain entire alarm lists. Additionally, the paper includes a controlled design study comparing various attribution strategies, revealing that X-StrGNN offers the highest stability at a significantly lower cost compared to other methods. The authors also document defects in the existing StrGNN implementation, highlighting the importance of accurate temporal attribution.
Methodology
The authors developed X-StrGNN as a post-hoc explanation layer for the StrGNN model, which operates on dynamic graphs. They implemented two types of attributions—structural and temporal—using multiplicative masks that ensure detection performance remains unchanged. The authors conducted a controlled design study to compare different attribution strategies, including gradient attribution and per-instance optimization, under a unified protocol.
Results
X-StrGNN achieved a detection performance with no change in AUC, AP, or precision (∆AUC = 0.0000). The attribution process was completed in 0.66 ms per flagged edge, making it 268 times faster than per-instance optimization. The stability of X-StrGNN's attributions was measured at 0.913, outperforming other strategies and exceeding a random baseline.
Implications
The findings suggest that X-StrGNN can enhance the interpretability of dynamic graph anomaly detection systems, making them more suitable for regulated environments where explanations are necessary. This could lead to broader adoption of such systems in critical applications, improving accountability and trust in automated decision-making.
Generalised Transportability via Causal Abstractions
Theory
Optimization
- Introduces a model-level approach to transportability using Causal Abstraction theory.
- Provides a framework for simultaneous transport of multiple queries rather than one at a time.
- Develops methods for approximate transportability that yield certified query intervals.
- Addresses limitations of existing transportability theories in scenarios with non-transportable queries or absent target data.
Read more
Generalised Transportability via Causal Abstractions
Summary
This paper addresses the problem of transporting causal conclusions from a source population to a target population, a key issue in causal inference. The authors critique existing transportability theories that provide a binary verdict on whether a causal query can be transported, highlighting their limitations in scenarios where queries are non-transportable or when no target data is available. To overcome these challenges, the authors propose a model-level perspective grounded in Causal Abstraction theory, which allows for a more holistic approach to transportability. They introduce a causal abstraction map that aligns the source and target populations based on their interventional behaviors, enabling the simultaneous transport of all target queries. The paper's main contribution is in the approximate case, where the authors develop a framework that yields certified query intervals even when no exact transportability map exists. This involves distributionally robust optimization to account for perturbations in the target environment, providing bounds for non-transportable queries and guarantees in target-agnostic settings. The proposed framework is evaluated on synthetic benchmarks and a real ecological dataset, demonstrating that the certified intervals effectively bracket the true interventional query.
Methodology
The authors utilize Causal Abstraction theory to create a causal abstraction map that aligns interventional behaviors between source and target populations. They formulate the transportability problem as a distributionally robust optimization problem, allowing for the derivation of certified query intervals even in the absence of an exact transportability map.
Results
The framework successfully provides certified intervals for target queries, effectively bracketing the true interventional query values in both synthetic and real ecological datasets. The results indicate that the proposed method can handle cases where traditional transportability methods fail, particularly in scenarios with non-transportable queries or when no target data is available.
Implications
This work has significant implications for causal inference in various fields, including epidemiology and social sciences, where understanding the transferability of causal effects across different populations is crucial. The ability to provide certified intervals enhances the reliability of causal conclusions drawn from experimental data, potentially improving decision-making in public health and policy.
Retrieval-guided Twin Fusion with Similarity-aware Contrast for Molecule-Text Alignment
Multimodal
- Introduction of RISEN, a novel framework for molecule-text alignment.
- Utilization of retrieval-guided twin fusion to enhance substructure representations.
- Implementation of similarity-aware contrast to reduce false negatives in learning.
- Demonstrated superior performance on benchmark datasets compared to existing methods.
Read more
Retrieval-guided Twin Fusion with Similarity-aware Contrast for Molecule-Text Alignment
Summary
This paper addresses the challenge of molecule-text alignment, which involves projecting molecular structures and their textual descriptions into a shared latent space for applications like molecule search and property prediction. Traditional methods often overlook the fine-grained semantic relationships between molecular substructures and their corresponding texts, leading to suboptimal performance. The authors propose a novel framework called Retrieval-guided Twin Fusion with Similarity-aware Contrast (RISEN) that enhances representation learning by constructing a latent twin molecule for each substructure through cross-modal retrieval. This approach retrieves relevant textual descriptions and samples similar molecules, aggregating their representations via attention pooling to create a twin latent representation. This representation is then fused with the original substructure to enrich its context. Additionally, the framework employs a similarity-aware contrastive learning strategy that replaces hard one-hot targets with soft targets based on embedding similarities, mitigating the issue of false negatives. Extensive experiments demonstrate that RISEN outperforms existing baselines in molecule-text retrieval and molecular property prediction tasks, showcasing its effectiveness in improving alignment and representation learning.
Methodology
The RISEN framework employs a two-step approach: first, it retrieves semantically related complete molecules for each substructure to create a twin latent representation through attention pooling. Second, it applies similarity-aware contrastive learning, replacing hard targets with soft targets derived from embedding similarities to better preserve semantic neighborhoods.
Results
The experiments conducted on molecule-text retrieval and molecular property prediction benchmarks show that RISEN consistently achieves significant performance improvements over strong baseline models, validating its effectiveness in enhancing molecule-text alignment.
Implications
The proposed RISEN framework has potential applications in drug discovery, materials science, and molecular property prediction, where accurate molecule-text alignment is crucial for understanding chemical information and facilitating cross-modal reasoning.
Sequence prediction under a lying oracle
Theory
- Introduces a perturbed loss function that accounts for lies from an oracle in sequential prediction.
- Establishes logarithmic upper bounds on regret for both stochastic and adversarial environments.
- Analyzes the performance of the add-β family of estimators and the Exponentially Weighted Online Optimizer (EWOO).
- Demonstrates the impact of the number of allowed lies on prediction performance through numerical experiments.
Read more
Sequence prediction under a lying oracle
Summary
This paper addresses the problem of sequential prediction of m-ary sequences in the presence of a 'lying oracle' that provides unreliable information. The authors propose a novel loss function that incorporates the complexity introduced by the oracle's lies, extending traditional log-loss to account for additional query complexity. They analyze the performance of the add-β family of estimators in a stochastic setting where outcomes are generated independently from an unknown distribution, and also in an adversarial setting where outcomes can be determined arbitrarily. The paper establishes logarithmic upper bounds on the regret for both settings, demonstrating that the proposed algorithms can effectively manage the challenges posed by the oracle's lies. Theoretical results are complemented by numerical experiments that illustrate the dependence of regret on various parameters, including the number of allowed lies and the choice of estimator.
Methodology
The authors develop algorithms for both stochastic and adversarial settings, utilizing the add-β family of estimators for the stochastic case and the Exponentially Weighted Online Optimizer (EWOO) for the adversarial case. They derive a new loss function that combines traditional log-loss with an additional term that reflects the complexity introduced by the oracle's lies. The performance is evaluated through theoretical analysis and numerical simulations.
Results
The paper shows that the proposed algorithms achieve logarithmic regret bounds in both stochastic and adversarial settings. The add-β estimator is shown to maintain optimal or near-optimal regret under the new perturbed loss function, while the EWOO effectively manages the adversarial outcomes. Numerical experiments validate the theoretical findings, revealing how regret varies with the number of allowed lies and the choice of estimator.
Implications
The findings have significant implications for online learning and prediction tasks where information may be unreliable. The proposed methods can be applied in various domains, including finance, healthcare, and any scenario where decision-making relies on potentially deceptive information sources.
L-FNO: Lorentzian Fourier Neural Operator for Stochastic Event Dynamics
Time Series
- L-FNO incorporates a Lorentzian spectral kernel to model self-exciting point process dynamics.
- The model is trained using a Poisson negative log-likelihood objective, improving calibration for rare events.
- L-FNO provides interpretable parameters that quantify the strength and persistence of self-excitation.
- The approach demonstrates superior performance on both synthetic and real-world datasets compared to traditional methods.
Read more
L-FNO: Lorentzian Fourier Neural Operator for Stochastic Event Dynamics
Summary
The paper introduces the Lorentzian Fourier Neural Operator (L-FNO), a novel stochastic neural operator designed to model event dynamics characterized by uncertainty and sparsity. Traditional neural operators, which are typically trained as regression models, struggle with rare events due to their inability to capture the likelihood structure of discrete events. L-FNO addresses this limitation by integrating a Lorentzian spectral kernel that corresponds to an exponential excitation function, akin to the dynamics of Hawkes processes. This allows L-FNO to model both covariate-driven effects and history-dependent dynamics effectively. The authors evaluate L-FNO on eight synthetic benchmarks and three real-world datasets, including disease outbreak prediction and semiconductor defect detection. The results demonstrate that L-FNO outperforms existing regression- and likelihood-based neural operator baselines in terms of event likelihood, calibration diagnostics, and rare-event detection, highlighting the importance of structured spectral memory and likelihood-based learning in stochastic event modeling.
Methodology
The L-FNO extends the Fourier Neural Operator architecture by introducing a Lorentzian spectral kernel, which is equivalent to an exponential excitation function. The model is trained using a Poisson negative log-likelihood loss to learn an event intensity function, capturing both covariate effects and historical dynamics. The authors also analyze the limitations of regression-based approaches under sparse event conditions and propose L-FNO as a solution.
Results
L-FNO significantly improves event likelihood, calibration diagnostics, and rare-event detection across multiple benchmarks. The model's performance surpasses that of regression- and likelihood-based neural operator baselines, demonstrating its effectiveness in capturing the dynamics of stochastic events.
Implications
The findings suggest that L-FNO can be applied to various fields where understanding the dynamics of rare and unpredictable events is crucial, such as epidemiology, finance, and manufacturing. The model's ability to quantify predictive temporal structures could enhance decision-making processes in operational systems facing uncertainty.
FirstDiff: One-Step Diffusion-Based Anomaly Detection for Multivariate Time Series via Initial Noise Prediction
Time Series
Generative Models
Efficient ML
- FirstDiff enables one-step anomaly detection by extracting information from the initial reverse-diffusion evaluation.
- The predicted diffusion noise serves as an effective representation for identifying anomalies in multivariate time series.
- The framework significantly reduces computational costs and inference latency compared to traditional diffusion-based methods.
- A Diffusion Transformer is employed to capture complex temporal and inter-sensor dependencies.
Read more
FirstDiff: One-Step Diffusion-Based Anomaly Detection for Multivariate Time Series via Initial Noise Prediction
Summary
The paper introduces FirstDiff, a novel framework for anomaly detection in multivariate time series using diffusion models. Unlike traditional methods that rely on completing the reverse diffusion process to detect anomalies, FirstDiff leverages the predicted diffusion noise at the initial reverse-diffusion evaluation. This approach allows for anomaly inference from a single evaluation of the denoising network, significantly reducing computational costs and inference latency. The framework employs a Diffusion Transformer (DiT) to effectively model complex temporal and inter-sensor dependencies, capturing long-range patterns and intricate correlations across multiple sensors. The authors establish the predicted diffusion noise as a robust representation for characterizing normal and anomalous behavior, utilizing statistical measures to identify deviations from a reference distribution derived from normal validation data. Extensive experiments on five benchmark datasets demonstrate that FirstDiff achieves state-of-the-art performance while maintaining efficiency in inference time.
Methodology
FirstDiff utilizes a Diffusion Transformer to model the statistical distribution of predicted diffusion noise during the initial reverse-diffusion evaluation. Anomalies are detected by comparing the predicted noise against a reference distribution derived from normal behavior using validation data. The framework reduces the need for completing the full reverse diffusion trajectory, allowing for efficient anomaly detection.
Results
The experiments conducted on five public benchmark datasets (SMAP, MSL, SMD, SWaT, and PSM) demonstrate that FirstDiff achieves state-of-the-art performance in anomaly detection while significantly lowering the computational cost associated with traditional diffusion-based methods.
Implications
FirstDiff's approach can enhance the reliability and safety of various cyber-physical systems by enabling efficient and accurate anomaly detection in real-time. Its application could be particularly beneficial in industrial control systems, healthcare monitoring, and financial platforms where timely detection of anomalies is critical.
TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity
Time Series
Efficient ML
- TinyCast computes periodicity explicitly, reducing the need for parameterized learning.
- It is the smallest zero-shot forecasting model on the GIFT-Eval board, with 146,505 parameters.
- The model emits a predictive distribution while maintaining probabilistic accuracy below 1.4 million parameters.
- TinyCast is optimized for embedded hardware, requiring no per-signal fitting.
Read more
TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity
Summary
TinyCast is introduced as a novel attention-free zero-shot forecasting model designed to provide probabilistic predictions with a minimal parameter count of 146,505. The model operates on the principle that the periodic structure of time series data can be computed rather than learned, thus optimizing its capacity for forecasting. It employs a zero-parameter spectral detector to identify dominant periods, which informs the folding of context based on their phase. The architecture consists of a dilated convolutional encoder and a block-autoregressive quantile decoder, enabling it to emit a predictive distribution without requiring extensive parameterization. TinyCast outperforms existing zero-shot models on the GIFT-Eval board, being the only model under 1.4 million parameters to provide a predictive distribution while maintaining probabilistic accuracy. The model is designed for deployment on embedded devices, requiring no per-signal fitting and operating efficiently with static INT8 computations. Controlled experiments validate its performance and identify key contributors to its accuracy, establishing TinyCast as a significant advancement in the field of time series forecasting.
Methodology
TinyCast utilizes a zero-parameter spectral detector to compute periodicity from a Fourier transform, which informs the context folding. The architecture includes a dilated convolutional encoder and a block-autoregressive quantile decoder, allowing it to generate probabilistic forecasts without attention mechanisms. The model is trained on a diverse dataset and evaluated on several benchmarks for zero-shot forecasting.
Results
TinyCast achieves superior performance on the GIFT-Eval board and other benchmarks like Chronos-ZS and fev-bench, outperforming larger models by a significant margin in terms of parameter efficiency and probabilistic accuracy. It is the only model under 1.4 million parameters to emit a predictive distribution without test-data leakage.
Implications
The development of TinyCast suggests a new direction for time series forecasting, emphasizing the importance of model size and computational efficiency. Its ability to operate on embedded devices without extensive fitting processes could facilitate real-time forecasting applications in various domains such as energy management, logistics, and environmental monitoring.
Learning Stock Trading Policies via Barycenter-Based Adversarial Inverse Reinforcement Learning
Reinforcement Learning
Time Series
Optimization
- Introduction of BRaG, a framework that aggregates expert strategies for stock trading.
- Utilization of a performance-weighted Wasserstein barycenter for stable pseudo-expert representation.
- Pretraining of trading policies through adversarial imitation learning to address exploration challenges.
- Incorporation of control barrier functions for risk-aware decision-making.
Read more
Learning Stock Trading Policies via Barycenter-Based Adversarial Inverse Reinforcement Learning
Summary
This paper presents BRaG, a novel barycenter-based adversarial inverse reinforcement learning framework designed to learn effective stock trading policies from multiple heterogeneous expert strategies. The authors address the challenges of delayed and noisy rewards, poor exploration, and the enforcement of risk constraints in reinforcement learning for trading. BRaG utilizes a performance-weighted Wasserstein barycenter to aggregate expert demonstrations, creating a stable pseudo-expert representation that captures shared structures across diverse trading styles. This representation is employed to pretrain a trading policy through adversarial imitation learning, which helps mitigate exploration instability. The pretrained policy is then refined using reinforcement learning with actual market rewards. To ensure risk-aware decision-making, the framework incorporates control barrier functions that constrain actions and regularize policy learning to adhere to drawdown limits. The proposed approach is evaluated across four major global equity markets, demonstrating superior performance compared to classical trading rules and recent deep reinforcement learning methods, while also exhibiting more stable risk characteristics.
Methodology
The methodology involves aggregating multiple expert trading strategies using a performance-weighted Wasserstein barycenter to create a stable pseudo-expert distribution. This distribution is then used to pretrain a trading policy via adversarial imitation learning, followed by refinement through reinforcement learning with actual market rewards. Control barrier functions are integrated to ensure compliance with risk constraints during policy learning.
Results
The BRaG framework was evaluated on four major global equity markets (US, UK, Indian, and Taiwanese indices) and achieved stronger performance than both classical trading rules and recent deep reinforcement learning methods. Additionally, it exhibited more stable risk characteristics, indicating its effectiveness in real-world trading scenarios.
Implications
The proposed framework has significant implications for automated trading systems, providing a robust method for learning trading policies that can adapt to diverse market conditions while ensuring risk management. It can be applied in financial technology solutions and algorithmic trading strategies.
More Correct Mass, Worse Answers: Why Power Sampling Can Fail and How to Fix It
NLP
Large Language Models
Theory
- Power Sampling can increase probability mass on correct trajectories but degrade overall inference accuracy.
- Two mismatches—dose mismatch and coverage mismatch—are identified as the primary causes of this paradox.
- A new support-preserving Power target is proposed to mitigate the issues associated with traditional Power Sampling.
- The modified sampling method shows improved performance over standard multi-sample inference techniques.
Read more
More Correct Mass, Worse Answers: Why Power Sampling Can Fail and How to Fix It
Summary
This paper investigates the paradox of Power Sampling in language models, which aims to enhance reasoning by sharpening the distribution of generation trajectories. Despite its intention to improve downstream inference, the authors find that Power Sampling can lead to significant accuracy drops, with reductions of up to 18.5 percentage points across various models and benchmarks. The authors identify two primary mismatches contributing to this paradox: dose mismatch, where a fixed exponent leads to varying distributional changes across problems, and coverage mismatch, where global sharpening concentrates probability mass on a limited set of dominant paths, undermining the collective support necessary for effective aggregation. To address these issues, the authors propose a new method that replaces uniform trajectory exponentiation with a deformation-controlled, support-preserving Power target. This approach calibrates sharpening across different problems while maintaining the diversity of moderate-probability paths. The results demonstrate that this modified sampling method reverses the accuracy losses caused by traditional Power Sampling and outperforms standard multi-sample inference methods across reasoning benchmarks.
Methodology
The authors analyze the effects of Power Sampling on downstream inference accuracy, identifying mismatches in trajectory distribution. They propose a new sampling method that calibrates sharpening across problems and preserves support for moderate-probability paths, which is validated through experiments on various reasoning benchmarks.
Results
The proposed method successfully reverses the accuracy losses observed with traditional Power Sampling, achieving better performance than standard multi-sample inference across multiple reasoning benchmarks.
Implications
The findings suggest that careful calibration of trajectory distributions is crucial for enhancing the performance of language models in reasoning tasks. The proposed method could serve as a more effective front end for various downstream sampling methods, improving the reliability of multi-trajectory reasoning.
Identifiability and Order-Dimension Limits of In-Context Learning on Partial Orders
Theory
Large Language Models
Graph Learning
- Formalizes relational ICL using version spaces on partial orders.
- Proves a completion trichotomy for open-world prompts with positive and negative demonstrations.
- Characterizes optimal teaching prompts through cover labels and blocker-set hitting problems.
- Establishes exact capability boundaries for prompt-dependent coordinate decoders.
Read more
Identifiability and Order-Dimension Limits of In-Context Learning on Partial Orders
Summary
This paper develops a theoretical framework for in-context learning (ICL) on partial orders, addressing the limitations of traditional function-learning approaches. The authors introduce a version-space semantics that distinguishes between open- and closed-world assumptions, allowing for a nuanced understanding of relational prompts. They prove a completion trichotomy for finite open-world prompts with positive and negative comparisons, categorizing queries as forced true, forced false, or genuinely ambiguous. The paper characterizes the optimal teaching number for open-world prompts, revealing that the maximum teaching cost for n-element posets is n(n−1), achieved uniquely by antichains. Additionally, the authors formalize prompt-dependent coordinate decoders and establish an exact capability boundary, indicating that a dimension of at most s is both necessary and sufficient for effective learning. The paper contributes to the understanding of relational reasoning in large language models and provides a structured approach to teaching and representation in ICL.
Methodology
The authors utilize a theoretical approach that combines version space semantics with a focus on partial orders. They analyze the logical identifiability of relational prompts and derive completion classifications based on positive and negative demonstrations. The study employs combinatorial techniques to characterize teaching costs and establish capability boundaries for coordinate decoders.
Results
The paper presents an exact completion trichotomy for queries based on finite open-world prompts, identifies the maximum teaching cost for n-element posets as n(n−1), and demonstrates that this maximum is uniquely attained by antichains. It also establishes that a dimension of at most s is necessary and sufficient for effective learning with prompt-dependent coordinate decoders.
Implications
The findings have significant implications for the design of large language models and their ability to perform relational reasoning tasks. The framework can inform the development of more efficient ICL strategies and enhance the understanding of how background knowledge influences learning outcomes in complex relational settings.
Connected Subspace Clustering: Hardness, a Scalable Heuristic, and an Application to Sea Level Geodesy
Optimization
Time Series
Theory
- Introduction of Connected Subspace Clustering, combining clustering with connectivity constraints.
- Proven NP-hardness of the problem for certain configurations.
- Development of a heuristic method that ensures connected clusters while minimizing reconstruction error.
- Extensive experimental validation showing the method's effectiveness in sea level geodesy.
Read more
Connected Subspace Clustering: Hardness, a Scalable Heuristic, and an Application to Sea Level Geodesy
Summary
This paper introduces the Connected Subspace Clustering problem, which integrates the need for both similarity among measurements and physical coherence in clustering. The authors demonstrate that this problem is NP-hard to approximate, even in simplified scenarios. They propose a heuristic method that combines subspace fitting with a connectivity enforcement strategy, ensuring that the resulting clusters are connected. The method is evaluated through extensive experiments on global sea level time series data, showing superior performance in maintaining connectivity and reducing reconstruction costs compared to existing clustering methods. The approach is particularly relevant for applications in geodesy, where it aids in analyzing climate-driven sea level variability, but it also has broader implications for other fields involving spatially embedded multivariate time series data.
Methodology
The authors propose a Lloyd-style heuristic that alternates between fitting affine subspaces to data points and merging components to maintain connectivity. This method operates on a given connectivity graph and is designed to run efficiently on bounded-degree graphs.
Results
In experiments involving 160 configurations of global sea level time series, the proposed method outperformed competitors in 73.75% of cases, effectively isolating signals that correlate with established climate indices. The method significantly reduced the number of disconnected fragments compared to unconstrained clustering approaches.
Implications
The Connected Subspace Clustering approach has significant implications for analyzing spatially embedded multivariate time series data, particularly in geodesy and climate science. It can also be adapted for use in other fields such as remote sensing, neuro-imaging, and sensor networks.
Learning Generalizable Reconstruction of High-Dimensional Neural Dynamics
Time Series
- PCA-DMD outperforms traditional DMD methods in reconstructing high-dimensional neural dynamics.
- The framework achieves high correlation and low divergence metrics in zero-shot generalization across subjects.
- PCA-DMD maintains accuracy in out-of-sample temporal predictions, demonstrating robustness across different temporal regions.
- The method is computationally scalable, allowing for effective reconstruction from large datasets.
Read more
Learning Generalizable Reconstruction of High-Dimensional Neural Dynamics
Summary
The paper addresses the challenge of accurately reconstructing long-duration neural recordings, specifically local field potentials (LFPs), which exhibit high temporal resolution, multichannel structure, and variability across subjects. Traditional Dynamic Mode Decomposition (DMD) methods struggle with high-dimensional signals, while deep learning approaches often require significant computational resources and lack dynamical interpretability. The authors introduce PCA-DMD, a scalable operator-theoretic framework that segments LFP recordings into overlapping windows, projects them into a principal-component space, learns a linear approximation of the Koopman evolution in this latent space, and reconstructs the continuous signal through inverse projection and overlap-add aggregation. The framework is evaluated through various experiments, demonstrating superior performance compared to classical DMD methods and achieving high correlation and low divergence metrics in cross-subject zero-shot generalization. The results indicate that PCA-DMD can accurately predict temporally held-out LFP segments and maintain stable performance across different sample sizes, establishing it as an interpretable and generalizable method for reconstructing high-dimensional neural dynamics.
Methodology
The PCA-DMD framework segments multichannel LFP recordings into overlapping windows, projects them into a compact principal-component space, learns a linear approximation of the Koopman evolution in this latent space, and reconstructs the continuous signal through inverse projection and overlap-add aggregation.
Results
PCA-DMD achieved a Kullback–Leibler divergence (KLD) of 0.0761 and a Hellinger distance (HD) of 0.0847 on 200,000-sample hippocampal recordings. In cross-subject zero-shot generalization at 300,000 samples, it achieved correlations of 0.9504–0.9800, with HD values of 0.0010–0.0072 and KLD values of 0.0005–0.0022. The method also demonstrated stable zero-shot reconstruction from 400,000 to 900,000 samples with mean correlation around 0.965–0.968. External validation on independent recordings yielded a mean channel-wise correlation of 0.7427.
Implications
The PCA-DMD framework has potential applications in computational neuroscience, particularly in the reconstruction and analysis of neural dynamics, which could aid in understanding neural mechanisms and disorders such as epilepsy and Alzheimer's disease.
LLM-based Framework for Generating and Verifying Parallel DEVS Statecharts
Large Language Models
NLP
Theory
- The framework leverages LLMs to generate and verify PDEVS statecharts from natural language descriptions.
- A controlled-correction mechanism is implemented to ensure logical consistency of generated plausible facts.
- The verification process improves statechart properties such as deadlock and phase reachability.
- Experimental results demonstrate the framework's generalizability across various LLMs and model complexities.
Read more
LLM-based Framework for Generating and Verifying Parallel DEVS Statecharts
Summary
This paper presents a novel framework that utilizes Large Language Models (LLMs) to assist in the generation and verification of Parallel DEVS (PDEVS) statecharts, which are essential for modeling the behavior of atomic Parallel Discrete Event System Specification (PDEVS) models. The framework operates in two stages: first, it generates plausible facts from a natural language description of the system, and second, it uses these facts along with a PDEVS schema to create statecharts. A significant challenge addressed is the logical inaccuracies that can arise from incorrect plausible facts, which can lead to flawed statecharts. To mitigate this, the authors introduce a controlled-correction mechanism that verifies the logical consistency of the generated facts through propositional logic entailment. The verification process is evaluated for completeness and correctness, demonstrating improvements in properties such as deadlock and phase reachability through model checking with Timed Automata. The framework's effectiveness is validated through experiments with various LLMs, showcasing its generalizability across different model complexities. The study also introduces a basic correctness metric to quantify the accuracy of the generated statecharts, highlighting the framework's capability to refine and enhance the modeling process iteratively.
Methodology
The methodology involves generating plausible facts from system descriptions using LLMs, followed by the verification of these facts through propositional logic entailment. The framework includes a controlled-correction loop that allows for iterative refinement of the generated facts and statecharts, ensuring logical consistency and accuracy.
Results
The evaluation of the proposed verification mechanism showed significant improvements in the logical consistency of the generated PDEVS statecharts. Model checking revealed enhanced properties such as reduced deadlock occurrences and improved phase reachability. The framework's effectiveness was confirmed through experiments with multiple LLMs, demonstrating its applicability across different scenarios.
Implications
This research has potential implications for the fields of modeling and simulation, particularly in automating the generation of accurate models from natural language descriptions. It can aid modelers in reducing errors and improving the reliability of simulation models, which is crucial in various domains such as system design, software engineering, and complex systems analysis.
Learning to Run Power Networks: Effective AlphaZero-inspired Topological Control
Reinforcement Learning
- The integration of renewable energy sources increases strain on power grids, necessitating effective management strategies.
- AlphaZero-inspired approaches using MCTS can significantly enhance the stability and operational efficiency of power networks.
- Optimized reward functions and observation configurations are crucial for improving the performance of RL agents in grid management.
- Simpler binary rewards can outperform complex reward structures in guiding search processes.
Read more
Learning to Run Power Networks: Effective AlphaZero-inspired Topological Control
Summary
This paper addresses the challenges posed by the integration of volatile renewable energy sources into power grids, which can lead to instability and congestion. It explores the use of Reinforcement Learning (RL) for autonomous topological reconfiguration, proposing a model-based approach inspired by AlphaZero that employs Monte Carlo Tree Search (MCTS) for proactive grid management. The authors systematically evaluate the impact of various design factors, including reward functions, observation density, and search guidance, on the performance of the RL agent. The study finds that an optimized AlphaZero approach achieves a peak survivability of 98.43%, significantly outperforming traditional methods like Proximal Policy Optimization (PPO). Notably, the research highlights that simpler binary rewards can provide more effective search guidance than complex multi-objective functions. The findings suggest that while AlphaZero is a powerful framework for topological control, successful implementation requires a minimalist integration of domain-specific heuristics and a focused observation space.
Methodology
The authors implemented a model-based AlphaZero-inspired approach that utilizes Monte Carlo Tree Search (MCTS) for proactive grid management. They conducted systematic evaluations of various design factors, including reward shaping, observation density, and search guidance, using the IEEE-14 bus system as a controlled environment to analyze the performance of the RL agent.
Results
The optimized AlphaZero approach achieved a peak survivability of 98.43%, significantly outperforming the PPO variant. The study also found that conducting MCTS without prior policy or value function guidance improved training efficiency, and that binary survival rewards provided more effective search guidance than complex reward functions.
Implications
The findings suggest that integrating AlphaZero-inspired methods into power grid management can lead to more efficient and stable operations, particularly as renewable energy sources become more prevalent. This research could inform future developments in autonomous grid management systems and contribute to the transition toward carbon neutrality.
The Query Knows What to Forget: A Second Erase Direction for Linear Attention
NLP
Large Language Models
Efficient ML
- Introduction of Query-derived Erase Direction (QED) to enhance linear attention mechanisms.
- QED adds a second erase direction that is orthogonal to the key and derived from the query.
- Significant improvements in retrieval performance, with up to 39 percentage points increase beyond training length.
- Controlled ablations indicate consistent benefits from using the query for defining the erase direction.
Read more
The Query Knows What to Forget: A Second Erase Direction for Linear Attention
Summary
This paper addresses the limitations of linear attention mechanisms, particularly the interference that arises when retrieving information from a fixed-size state as context length increases. The authors introduce a novel approach called Query-derived Erase Direction (QED), which enhances the existing Gated DeltaNet-2 (GDN-2) model by incorporating a second erase direction derived from the query. This new direction is orthogonal to the key, allowing the model to effectively cancel out old-state content that interferes with retrieval. The paper demonstrates that QED significantly improves retrieval performance beyond the training length, effectively doubling the usable context length on the S-NIAH-1 dataset. The authors conduct extensive experiments with 340M-parameter models trained on 15 billion tokens, showing that QED consistently benefits retrieval accuracy across multiple training seeds. The findings suggest that the integration of query-derived information into the erase mechanism can mitigate interference and enhance the efficiency of linear attention models.
Methodology
The authors propose a modification to the GDN-2 model by adding a gated term derived from the query to the erase vector, projecting it orthogonal to the key. This adjustment allows the model to cancel out old-state content that interferes with the retrieval process. The methodology includes training large models on extensive datasets and conducting controlled ablation studies to evaluate the impact of the new erase direction.
Results
The implementation of QED resulted in a notable improvement in retrieval accuracy, with enhancements of up to 39 percentage points beyond the training length on the S-NIAH-1 dataset. Additionally, the model showed improvements in FDA scores by 3.4 and 3.7 points across different training seeds. The controlled ablation studies confirmed the effectiveness of using the query to define the added erase direction.
Implications
The findings suggest that incorporating query-derived information into attention mechanisms can significantly enhance the performance of linear attention models, particularly in scenarios with long contexts. This could lead to advancements in various applications of natural language processing and other fields that rely on efficient retrieval of information from large datasets.
M-LINKX: Multiview Graph Learning for Brain Cognitive Disease Detection
Graph Learning
Time Series
- M-LINKX is a multi-view graph learning framework specifically designed for EEG-based dementia classification.
- The framework constructs multiple functional-connectivity graph views to capture diverse channel interaction patterns.
- M-LINKX simplifies the modeling of node features and connectivity without relying on complex message passing.
- Experiments show that M-LINKX outperforms existing EEG classification methods on two different datasets.
Read more
M-LINKX: Multiview Graph Learning for Brain Cognitive Disease Detection
Summary
The paper presents M-LINKX, a novel multi-view graph learning framework aimed at improving the classification of dementia-related conditions using electroencephalogram (EEG) data. The authors highlight the challenges in EEG-based classification due to the noisy and non-stationary nature of EEG signals, which vary across subjects. M-LINKX addresses these challenges by extracting channel-level node features from EEG segments and constructing multiple functional-connectivity (FC) graph views, each defined by different combinations of connectivity metrics, frequency bands, and topology filters. Unlike traditional graph models that rely on message passing, M-LINKX employs a simpler design that models node features and adjacency-based connectivity representations. The framework fuses the graph-view representations using global trainable view weights, and subject-level predictions are derived by averaging segment-level probabilities. The effectiveness of M-LINKX is demonstrated through experiments on two EEG datasets (CAUEEG and AHEAP), where it achieves superior subject-level performance compared to existing methods. The study concludes that integrating multi-view functional connectivity with an appropriate graph-learning architecture can significantly enhance EEG-based dementia classification.
Methodology
M-LINKX extracts channel-level features from EEG segments and constructs multiple functional-connectivity graphs based on different metrics, frequency bands, and topologies. It fuses these graph representations using trainable weights and computes subject-level predictions by averaging segment-level probabilities.
Results
M-LINKX achieved the best subject-level classification performance on two EEG datasets (CAUEEG and AHEAP), demonstrating its effectiveness in distinguishing between healthy controls and various dementia-related conditions.
Implications
The findings suggest that multi-view graph learning can enhance the accuracy of EEG-based diagnostics for cognitive diseases, potentially leading to better early detection and treatment strategies for dementia-related conditions.
MedMix: Specialization-Consistent Federated Sparse MoEs under Modality Heterogeneity
Federated Learning
Multimodal
- Identifies fragility in federated multimodal MoEs due to dual-level modality heterogeneity.
- Introduces MedMix, a framework that ensures semantic consistency in routing and expert specialization.
- Demonstrates significant performance improvements over existing federated multimodal and MoE baselines.
- Highlights the importance of cross-client specialization consistency for effective aggregation.
Read more
MedMix: Specialization-Consistent Federated Sparse MoEs under Modality Heterogeneity
Summary
The paper addresses the challenges of modality heterogeneity in federated multimodal medical AI, where clients may lack access to certain modalities and individual records may contain incomplete modality subsets. The authors propose MedMix, a semantic-alignment framework for federated sparse Mixture-of-Experts (MoEs) that enhances routing and expert specialization across clients. MedMix incorporates three main components: modality-context-aware routing for expert selection based on modality identity and context, consensus-guided routing alignment to create shared modality patterns, and client-adaptive expert aggregation to match and combine similar experts while maintaining their specialization. The framework is evaluated on two real-world medical datasets, ADNI and MIMIC-IV, demonstrating superior performance in terms of average F1 scores, especially under conditions of severe modality heterogeneity. The findings highlight the importance of consistent routing and expert specialization in improving the effectiveness of federated multimodal learning.
Methodology
The methodology involves a three-component framework: (1) modality-context-aware routing that guides expert selection based on the modality identity and context of each token; (2) consensus-guided routing alignment that constructs consensus anchors for shared modality patterns and aligns local routing distributions; and (3) client-adaptive expert aggregation that matches and aggregates experts based on client-specific modality patterns.
Results
MedMix achieved the best average F1 score across various scenarios of modality heterogeneity and incompleteness in the ADNI and MIMIC-IV datasets, with particularly notable improvements under severe modality heterogeneity conditions.
Implications
The findings suggest that maintaining semantic consistency in routing and expert specialization can significantly enhance the performance of federated learning systems in medical AI, potentially leading to more accurate and robust multimodal models in clinical settings.
Convex losses and their applications to SVM, SVR, and Shallow Neural Networks
Theory
Optimization
- Introduction of new convex loss functions for SVM and shallow neural networks.
- Demonstration of the applicability of these losses in primal SVM formulations.
- Use of Particle Swarm Optimization for solving the modified SVM problem.
- Evaluation through Nested Cross-Validation on small datasets shows consistent generalization measures.
Read more
Convex losses and their applications to SVM, SVR, and Shallow Neural Networks
Summary
This paper introduces novel convex loss functions for Support Vector Machines (SVM) and shallow neural networks, specifically targeting binary classification tasks. The author highlights the limitations of applying these losses in dual SVM models but demonstrates their effectiveness in the primal formulation of SVM and in neural networks. The proposed losses generalize standard loss functions and are evaluated using the Particle Swarm Optimization algorithm. The study includes a theoretical foundation for the new losses, showing that they can enhance generalization performance by incorporating pattern correlations. The evaluation is conducted through a Nested Cross-Validation procedure on several small datasets, revealing that the generalization measures remain consistent regardless of the loss function used. The paper also discusses the mathematical derivation of dual objective functions for SVM and SVR, emphasizing the computational challenges in solving these problems. The findings suggest that while the new losses theoretically improve performance, empirical results indicate no significant difference in generalization capabilities compared to traditional losses.
Methodology
The paper employs a mathematical approach to derive new convex loss functions and integrates them into the primal formulations of SVM and SVR. The Particle Swarm Optimization algorithm is utilized to solve the modified SVM problem. The performance of the proposed losses is evaluated using a Nested Cross-Validation procedure on various small datasets.
Results
The results indicate that while the new convex losses are a theoretical generalization of standard losses, empirical evaluations show no significant improvement in generalization performance across the tested datasets. The study confirms that the generalization measures are similar whether using the new losses or traditional ones.
Implications
The findings suggest that while the proposed convex losses may offer theoretical advantages, their practical application does not necessarily lead to improved performance in binary classification tasks. This raises questions about the effectiveness of modifying loss functions in machine learning models and encourages further exploration of loss function design.
RecipeNet: A Hierarchical Transformer for Recipe Data
Theory
NLP
Optimization
- RecipeNet effectively models the hierarchical and sequential nature of recipe data.
- The architecture outperforms traditional tabular learning methods across various datasets.
- Field-level tokenization and step-level fusion are key components of RecipeNet's design.
- The model accommodates variable schemas inherent in recipe data without losing structural relationships.
Read more
RecipeNet: A Hierarchical Transformer for Recipe Data
Summary
RecipeNet introduces a novel hierarchical Transformer architecture designed specifically for recipe data, which is prevalent in various domains such as materials synthesis and pharmaceutical formulation. Traditional tabular learning methods struggle with recipe data due to its unique characteristics, including variable schemas, hierarchical structures, and sequential dependencies. RecipeNet addresses these challenges by preserving the inherent structure of recipes, allowing for effective modeling of both intra-step field interactions and inter-step dependencies. The architecture consists of three main stages: field-level tokenization, step-level fusion, and recipe-level sequence encoding. Through extensive experiments on multiple recipe datasets, RecipeNet demonstrates superior performance compared to existing tabular models, underscoring the importance of hierarchical and sequential modeling in recipe representation learning.
Methodology
RecipeNet employs a hierarchical Transformer architecture that consists of three stages: field-level tokenization to process individual fields within each recipe step, step-level fusion to combine information from multiple fields, and recipe-level sequence encoding to capture dependencies across procedural steps. This approach allows for the preservation of the native hierarchical structure of recipes, enabling effective representation learning.
Results
Experimental results indicate that RecipeNet consistently outperforms existing tabular learning methods across multiple recipe datasets and downstream prediction tasks, demonstrating its effectiveness in capturing the complex relationships within recipe data.
Implications
The development of RecipeNet has significant implications for various fields that utilize recipe data, such as industrial manufacturing, pharmaceuticals, and materials science. By providing a robust framework for modeling complex procedural data, RecipeNet can enhance the efficiency and accuracy of recipe-based predictions and optimizations.
PL-Guard: Probabilistic Logic Reasoning for LLM Guardrails
NLP
Large Language Models
Interpretability
- PL-Guard operationalizes LLM guardrails as probabilistic logical policy-consistency checking.
- It separates neural predicate grounding from symbolic policy reasoning for improved safety.
- The architecture uses ProbLog for explicit rule-level policy reasoning, enhancing interpretability.
- Evaluation shows a drastic reduction in unsafe compliance, highlighting the effectiveness of the approach.
Read more
PL-Guard: Probabilistic Logic Reasoning for LLM Guardrails
Summary
The paper introduces PL-Guard, a neurosymbolic architecture designed to enhance the safety of large language models (LLMs) by separating the tasks of semantic grounding and policy reasoning. Traditional approaches to LLM guardrails often conflate these tasks, leading to issues such as unsafe compliance with harmful prompts or unnecessary refusals of benign requests. PL-Guard addresses this by utilizing a symbolic policy interface with predicates and ProbLog rules, where a local LLM grounds prompt-response pairs into predicate probabilities, and ProbLog conducts explicit probabilistic rule inference over the symbolic policy. The authors evaluate PL-Guard on the XSTest benchmark, demonstrating a significant reduction in unsafe compliance from 22.0% to 0.5%, while noting a trade-off with higher over-refusal rates compared to an LLM-as-a-judge baseline. This separation of grounding and reasoning not only improves safety but also provides a more transparent and auditable process for understanding the guardrail's decision-making.
Methodology
PL-Guard consists of two phases: an offline phase where a natural language policy is translated into a symbolic representation using ProbLog, and a forward phase where a local LLM processes user requests to generate initial responses, which are then evaluated against the symbolic policy to determine compliance and guide final outputs.
Results
On the XSTest benchmark, PL-Guard achieved a reduction in unsafe compliance from 22.0% for the base model to 0.5%, outperforming the LLM-as-a-judge baseline which had a 6.0% rate. However, it also exhibited a higher over-refusal rate of 14.4% compared to 5.2% for the baseline.
Implications
The findings suggest that separating grounding from reasoning can enhance the safety and usability of LLMs in real-world applications, providing a framework for more reliable and interpretable guardrails that can be audited for compliance with safety policies.
Iterative Refinement Diffusion for Super-Resolved Data Assimilation of Multiscale Physical Systems
Generative Models
Time Series
Theory
- Introduces a multiresolution framework for super-resolved data assimilation.
- Combines learned data assimilation with classical forecast-update cycles.
- Demonstrates improved performance in recovering high-resolution states.
- Outperforms traditional one-shot super-resolution methods.
Read more
Iterative Refinement Diffusion for Super-Resolved Data Assimilation of Multiscale Physical Systems
Summary
This paper introduces a novel multiresolution forecast-analysis diffusion framework aimed at enhancing super-resolved data assimilation for complex physical systems. The authors propose an Iterative Refinement (IR) approach that integrates learned data assimilation techniques with classical forecast-update cycles. By decomposing the reconstruction task into a series of resolution-wise operations, the method leverages a shared neural operator for dynamical prior information and a conditional diffusion corrector for refining states across different resolutions. The framework is evaluated on two benchmarks: the one-dimensional Burgers equation and two-dimensional Kraichnan turbulence, demonstrating significant improvements in recovering high-wavenumber energy and overall state accuracy compared to traditional methods and one-shot super-resolution techniques. The results indicate that the IR approach is particularly advantageous in multiscale and underdetermined scenarios, providing a more effective strategy for data assimilation in scientific applications.
Methodology
The proposed methodology involves an iterative refinement process that decomposes the state reconstruction task into a series of forecast-analysis steps across a multiresolution hierarchy. A shared neural operator with spectral mode slicing serves as a dynamical prior, while a conditional diffusion corrector refines the state estimates at each resolution level. This approach allows for better conditioning of the inverse problem, enhancing the assimilation of low-resolution observations.
Results
In the evaluation on the Kraichnan turbulence benchmark, the IR method achieved an RMSE of 0.184 and an SSIM of 0.836, outperforming various baselines including spectral interpolation and one-shot diffusion super-resolution. For the Burgers equation, the IR method achieved an RMSE of 0.00436, remaining competitive with the best method, which had an RMSE of 0.00278. These results highlight the effectiveness of the IR approach in multiscale data assimilation tasks.
Implications
The proposed framework has significant implications for scientific machine learning and data assimilation, particularly in fields such as numerical weather prediction and geophysical fluid dynamics. It offers a more efficient and effective means of reconstructing high-resolution states from low-resolution observations, potentially improving the accuracy of forecasts in complex physical systems.
Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling
Time Series
- Introduction of BaguanHR, a framework for high-resolution weather forecasting that emphasizes data transfer over model transfer.
- Demonstration that super-resolution techniques provide a more robust and effective means of generating high-resolution training data.
- Establishment of power-law scaling laws indicating that performance improves significantly with increased data volume.
- BaguanHR outperforms existing ML models and traditional forecasting methods, achieving notable reductions in RMSE.
Read more
Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling
Summary
This paper addresses the challenges of high-resolution weather forecasting using machine learning (ML) by proposing a novel framework called BaguanHR. Traditional methods have struggled due to the limited availability of high-resolution data, as existing reanalysis datasets are primarily at a coarser resolution. The authors argue that transferring models from coarse to fine resolutions is hindered by information loss and propose a shift in focus from model transfer to data transfer. By employing super-resolution (SR) techniques, they synthesize extensive high-resolution data from existing datasets, demonstrating that SR is more effective than direct forecasting due to its lower conditional entropy and greater robustness. The results show that BaguanHR significantly outperforms both ML-based methods and traditional forecasting systems, achieving superior performance across over 85% of lead times within 72 hours. The study also establishes a power-law scaling effect, indicating that increasing the training data volume leads to substantial reductions in forecasting error. Overall, the findings suggest that the key to advancing high-resolution ML forecasting lies in effective data construction through variable-wise super-resolution.
Methodology
The authors developed BaguanHR, which utilizes variable-wise super-resolution to synthesize high-resolution weather data from coarser datasets. They conducted experiments comparing the performance of super-resolution against traditional forecasting methods, analyzed the robustness of their approach, and established data scaling laws to quantify the impact of training data volume on forecasting accuracy.
Results
BaguanHR achieved superior performance in over 85% of lead times within 72 hours, with a 4.0% reduction in RMSE compared to the IFS-HRES model. The study found that increasing the training dataset from 7 to 18 years resulted in a 4.9% reduction in RMSE for long lead times, demonstrating the effectiveness of their data scaling approach.
Implications
The findings suggest that improving high-resolution weather forecasting relies heavily on the availability and quality of training data. The proposed framework could be applied to other domains where high-resolution predictions are needed, potentially transforming how data-driven models are developed in meteorology and beyond.
Adversarial Learning of Classifier-Free Guidance Schedules
Generative Models
Computer Vision
Multimodal
- Introduces a method for learning dynamic guidance schedules based on marginal consistency.
- Utilizes an adversarial framework for density ratio estimation to optimize guidance scales.
- Demonstrates improved image quality and text alignment compared to static and heuristic CFG methods.
- Empirical results validate the effectiveness of the proposed approach on standard benchmarks.
Read more
Adversarial Learning of Classifier-Free Guidance Schedules
Summary
This paper addresses the limitations of static Classifier-Free Guidance (CFG) schedules in text-to-image diffusion models, which often lead to suboptimal image quality and artifacts. The authors propose a novel approach to learn dynamic guidance schedules that adapt based on diffusion time, conditioning, and the current noisy sample. By framing the problem as density ratio estimation, they utilize an adversarial framework where a discriminator estimates the log-density ratio between true and guided distributions, while a generator predicts optimal guidance scales. The proposed method emphasizes marginal consistency, allowing for better alignment of generated images with text prompts. Empirical results demonstrate that this approach outperforms both heuristic CFG schedules and previous methods for learning dynamic guidance, achieving superior performance on text-to-image generation benchmarks.
Methodology
The authors employ an adversarial learning framework where a discriminator estimates the log-density ratio between real and classifier-free guided samples. A generator network learns to predict guidance scales that maximize this ratio, optimizing for marginal consistency across the sampling trajectory. This approach allows for time-dependent, conditioning-aware guidance scales that adapt to the current state of the diffusion process.
Results
The proposed method significantly outperforms constant guidance weights and manually designed schedules in terms of image fidelity and text alignment, as evidenced by improved Fréchet Inception Distance (FID) scores on text-to-image generation benchmarks.
Implications
This work has the potential to enhance the performance of generative models in various applications, particularly in generating high-fidelity images from textual descriptions. The dynamic guidance schedules could lead to more robust and flexible generative systems that better understand and adapt to complex prompts.
CutClean: Neural Network Pruning for Privacy-Preserving Inference
Efficient ML
Theory
- CutClean is the first structured pruning framework explicitly designed to reduce private attribute leakage in neural networks.
- The method allows for significant reductions in privacy leakage without the need for adversarial minimax optimization.
- CutClean achieves high structured sparsity while preserving target task accuracy.
- The approach highlights the importance of model-centric privacy interventions beyond dataset bias mitigation.
Read more
CutClean: Neural Network Pruning for Privacy-Preserving Inference
Summary
The paper introduces CutClean, a novel structured pruning framework designed to mitigate privacy leakage in neural networks while enhancing model sparsity. It addresses the growing concerns of privacy risks associated with neural networks deployed in sensitive applications, highlighting that privacy leakage can occur even without explicit dataset biases. CutClean employs auxiliary linear privacy heads at various network blocks to measure information leakage and applies structured pruning to remove components that contribute to the predictability of sensitive attributes. The authors validate their approach through experiments on both synthetic and real-world datasets, demonstrating that CutClean effectively reduces private information flow while maintaining high classification accuracy and achieving significant sparsity. This work positions pruning as a privacy-preserving intervention rather than merely a tool for model compression or robustness, revealing a compatibility between privacy reduction and efficient model deployment.
Methodology
CutClean introduces auxiliary linear privacy heads at intermediate blocks of the neural network to quantify privacy leakage. The framework progressively applies structured pruning to remove components that contribute to private attribute predictability, while continuously monitoring both target task performance and privacy leakage.
Results
Experiments demonstrate that CutClean effectively minimizes private information flow, achieving high sparsity rates and maintaining classification accuracy across both synthetic and real-world datasets.
Implications
The findings suggest that structured pruning can serve as a viable strategy for enhancing privacy in neural networks, making it particularly relevant for applications in high-stakes domains such as healthcare and finance, where privacy protection is paramount.
Revisiting Energy-based Tabular Anomaly Detection: Energy and Reconstruction are Complementary
Theory
- The DBM's mean-field energy matches or outperforms classical anomaly detection baselines.
- Combining DBM energy with Autoencoder reconstruction error yields significant performance improvements.
- The study highlights the complementary nature of energy-based and reconstruction-based methods in anomaly detection.
Read more
Revisiting Energy-based Tabular Anomaly Detection: Energy and Reconstruction are Complementary
Summary
This paper addresses the gap in tabular anomaly detection methods by revisiting Energy-Based Models (EBMs), specifically the Deep Boltzmann Machine (DBM). The author hypothesizes that the mean-field energy from the DBM can effectively complement reconstruction-based scores, such as those from Autoencoders (AEs). The study evaluates a two-hidden-layer DBM on two distinct datasets: UCI Bank Marketing and NSL-KDD, comparing its performance against eight classical and modern anomaly detection baselines. The results show that the DBM's mean-field energy matches the performance of the strongest baseline (Autoencoder) on one dataset and statistically outperforms it on another. Furthermore, when combined with the Autoencoder through rank fusion, the DBM energy significantly enhances anomaly detection performance, demonstrating the value of integrating energy-based and reconstruction-based approaches. The findings suggest that classical EBMs, like the DBM, should be considered a vital component in the tabular anomaly detection toolbox.
Methodology
The paper employs a two-hidden-layer Deep Boltzmann Machine (DBM) and evaluates its mean-field energy against eight baseline methods across two datasets using a multi-seed protocol. The performance of the DBM is compared to classical density-proxy methods and reconstruction-based methods, and hybrid models are created by combining DBM energy with Autoencoder outputs.
Results
The DBM's mean-field energy statistically ties with the Autoencoder on the Bank Marketing dataset and outperforms it on the NSL-KDD dataset. When fused with the Autoencoder, the DBM energy shows a statistically significant improvement in anomaly detection performance, with AUROC increases of +0.014 and +0.002 on the respective datasets.
Implications
The findings suggest that incorporating energy-based models like the DBM can enhance the effectiveness of anomaly detection in tabular data, providing a new perspective that complements existing reconstruction-based methods. This could lead to more robust anomaly detection systems in various applications, including finance and cybersecurity.
Structure-Guided Spatiotemporal Attention Graph Neural Network for Traffic Flow Prediction
Graph Learning
Time Series
Interpretability
- SGSAN explicitly learns a static Directed Dependency Graph (DDG) to capture stable traffic state propagation paths.
- The model incorporates an InfoNCE-based soft-coupling mechanism to align dynamic attention with structural priors.
- A decoupled two-stage optimization framework resolves conflicts between structural learning and predictive accuracy.
- SGSAN achieves high predictive accuracy while enhancing interpretability, addressing the 'black-box' nature of traditional models.
Read more
Structure-Guided Spatiotemporal Attention Graph Neural Network for Traffic Flow Prediction
Summary
This paper introduces the Structure-Guided Spatiotemporal Attention Graph Neural Network (SGSAN) aimed at enhancing traffic flow prediction by integrating graph convolutions and attention mechanisms while addressing the interpretability challenges of existing models. Traditional models often lack transparency, leading to reliance on spurious correlations that can undermine trust in safety-critical urban systems. The SGSAN overcomes this by learning a static Directed Dependency Graph (DDG) that captures invariant macroscopic traffic state propagation paths. This structural prior anchors the model's dynamic attention mechanism, ensuring that predictions are robust and aligned with the underlying traffic network's physical logic. The authors propose a two-stage optimization framework that separates structural discovery from predictive error minimization, thus stabilizing the learning process. Extensive experiments on real-world datasets demonstrate that SGSAN achieves state-of-the-art predictive accuracy while providing interpretability that reflects the inherent traffic dynamics.
Methodology
The SGSAN employs a two-stage training process: the first stage focuses on learning a static Directed Dependency Graph (DDG) to capture long-term spatial dependencies, while the second stage utilizes a spatiotemporal attention mechanism to adapt to dynamic local variations. The model incorporates an InfoNCE contrastive learning approach to guide attention towards the DDG, ensuring robust predictions without overfitting to local noise.
Results
The SGSAN demonstrated state-of-the-art predictive accuracy on multiple real-world traffic datasets, outperforming existing models. Additionally, it provided built-in interpretability, allowing for a clearer understanding of the model's decision-making process in relation to traffic dynamics.
Implications
The findings suggest that incorporating structural dependencies into traffic flow prediction models can significantly enhance their interpretability and reliability, making them more suitable for deployment in safety-critical urban environments. This approach could be extended to other domains where understanding the underlying structure is crucial for model trustworthiness.
Hard Cases, Bad Labels: Testing Error Exposure and Error Location in Uncertainty Sampling Under Bounded Label Noise
Theory
Efficient ML
- Uncertainty sampling can lead to improved model performance under clean labels but may acquire more corrupted labels under noisy conditions.
- The advantage of uncertainty sampling diminishes more significantly under difficulty-dependent noise compared to random classification noise.
- No evidence was found to support the hypothesis that structured error locations are more harmful than random noise.
- The performance of uncertainty sampling is influenced by the dataset, annotation budget, and noise structure.
Read more
Hard Cases, Bad Labels: Testing Error Exposure and Error Location in Uncertainty Sampling Under Bounded Label Noise
Summary
This study investigates the effectiveness of uncertainty sampling in active learning, particularly under conditions of bounded label noise. The research addresses two main concerns: the potential for uncertainty sampling to acquire more corrupted labels and the impact of error location on model performance. The author compares margin-based uncertainty sampling with random sampling across three binary tabular datasets, employing various noise models, including random classification noise (RCN) and difficulty-dependent noise. The methodology involves a controlled experimental design with paired seeds, varying noise rates, and annotation budgets. The findings reveal that uncertainty sampling generally improves normalized balanced-accuracy area under the learning curve (AULC) compared to random sampling under clean labels. However, the advantage diminishes under difficulty-dependent noise, particularly in certain datasets. The study also finds no significant evidence that structured error locations are more detrimental than random noise, challenging some existing assumptions in the field. Overall, the results highlight the nuanced performance of uncertainty sampling based on dataset characteristics, noise structure, and evaluation metrics.
Methodology
The study employs a controlled experimental design comparing margin-based uncertainty sampling and random sampling across three binary tabular datasets. It utilizes 100 paired seeds, nine expected noise rates (0 to 0.30), and varying annotation budgets (20 to 120). Logistic regression with cross-validated regularization is applied, and a leave-one-seed-out exposure-matched control is used to assess performance differences after aligning mean final acquired corruption.
Results
Under clean labels, uncertainty sampling improved AULC by 1.09 to 1.77 percentage points across all datasets. Difficulty-dependent noise reduced this advantage more than RCN in six of eight rates on the Breast Cancer Wisconsin dataset, but not on Banknote Authentication or MAGIC Gamma Telescope. The exposure-matched analysis showed no significant evidence for the predicted additional harm of structured error locations. No mean AULC advantage fell below zero through 0.30 noise levels.
Implications
The findings suggest that while uncertainty sampling is label-efficient, its robustness is contingent on various factors, including dataset characteristics and noise structure. This has implications for the design of active learning systems, particularly in domains where label noise is a concern.
Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays
NLP
Large Language Models
Theory
- Identification of three failure modes in naive per-field selective risk control for document extraction.
- Introduction of a validity ladder as a structured reporting standard for risk control.
- Demonstration of effective risk management through fit/val split protocols and Mondrian Learn-then-Test methods.
- Empirical results showing the effectiveness of conditioning on provenance for improving risk certification.
Read more
Valid Per-Field Selective Risk Control for Document Extraction: Three Failure Modes, a Validity Ladder, and When Conditioning Pays
Summary
This paper addresses the challenges of ensuring selective risk control in document extraction systems, particularly focusing on the per-field accept/review mechanism. The author identifies three critical failure modes that undermine the reliability of naive implementations: document clustering, score-refit leakage, and tie-mass pathology. These issues lead to a significant violation of the expected error rates among accepted fields. To mitigate these problems, the paper proposes a 'validity ladder' as a reporting standard, which includes a practical tier using a fit/val split protocol and more rigorous tiers employing Mondrian Learn-then-Test methods. The results demonstrate that the proposed methods can achieve controlled selective risk, with empirical evidence showing that conditioning on provenance-derived taxonomies can enhance certification in challenging scenarios. The findings highlight the importance of proper calibration and the need for robust methodologies in document extraction systems.
Methodology
The study employs a combination of empirical analysis and theoretical frameworks to diagnose failure modes in document extraction systems. It introduces a validity ladder to categorize different levels of risk control, utilizing a fit/val split protocol and Mondrian Learn-then-Test methods to achieve selective risk guarantees. The methodology includes counterfactual experiments to quantify the impact of identified failure modes.
Results
The proposed methods achieved a coverage of 0.318 at an achieved risk of 0.096 under nominal conditions, with significant improvements in risk control demonstrated through rigorous testing. The study also found that conditioning on provenance-derived support bins consistently outperformed other methods across various rigor tiers, achieving statistically significant results (p < 10^-4).
Implications
The findings have significant implications for the development of reliable document extraction systems, particularly in applications requiring high trust and accuracy. The structured approach to risk control can enhance the deployment of AI in sensitive areas such as finance, healthcare, and legal documentation, where errors can have serious consequences.
Robust XGBoosting for Regression
Theory
Optimization
Efficient ML
- XGBoost is sensitive to outliers and leverage points when using the standard quadratic loss.
- The authors propose MM-XGBoost, a two-step procedure that incorporates robust loss functions.
- Alternative loss functions based on M-, S-, and Ï„-estimators improve robustness without sacrificing predictive accuracy.
- Simulation studies validate the effectiveness of the proposed robust extensions of XGBoost.
Read more
Robust XGBoosting for Regression
Summary
This paper investigates the robustness of XGBoost, a widely used gradient boosting method for regression, particularly in the presence of outliers and leverage points. The authors highlight that the standard quadratic loss function can lead to significant performance degradation when the data contains vertical outliers or influential observations. While XGBoost offers a Huber loss as a more robust alternative, it still remains sensitive to leverage points. To enhance the robustness of XGBoost, the authors propose a two-step procedure called MM-XGBoost, which incorporates alternative loss functions derived from robust regression estimators, including M-, S-, and Ï„-estimators. Through simulation studies, the authors demonstrate that MM-XGBoost achieves a better balance between robustness and predictive accuracy compared to traditional XGBoost implementations. The paper concludes by emphasizing the potential of these robust loss functions in improving the performance of XGBoost in real-world regression tasks.
Methodology
The authors adapt and extend existing robust regression techniques to create new loss functions for XGBoost. They conduct simulation studies to evaluate the performance of these robust loss functions in terms of their ability to mitigate the influence of outliers while maintaining predictive accuracy.
Results
The simulation studies indicate that MM-XGBoost outperforms traditional XGBoost in terms of robustness against outliers and leverage points, achieving a favorable trade-off between robustness and prediction accuracy.
Implications
The findings suggest that incorporating robust loss functions into XGBoost can significantly enhance its applicability in real-world scenarios where data may contain outliers, thereby improving the reliability of predictions in various regression tasks.
Invariant Pretraining for Robust Code Representations
NLP
Theory
Efficient ML
- Identifies a robustness gap in encoder-based code models when dealing with invariant programs.
- Introduces Invariant Pretraining (InvPT) as a solution that utilizes semantics-preserving transformations.
- Demonstrates significant improvements in robustness across various model-dataset comparisons.
- Shows that multi-positive invariant contrastive learning is crucial for enhancing performance.
Read more
Invariant Pretraining for Robust Code Representations
Summary
This paper addresses the robustness of encoder-based code representation models, which are commonly used for tasks like clone detection and code classification. The authors identify a significant robustness gap when these models encounter invariant programs—semantically equivalent code expressed in different syntactic forms. The study empirically evaluates the performance degradation of four encoder baselines across two tasks and four datasets. To mitigate this issue, the authors propose a method called Invariant Pretraining (InvPT), which employs semantics-preserving transformations and a multi-positive supervised contrastive learning approach. This method allows the model to learn from both original and transformed code snippets without the need for paired natural language data. The results indicate that InvPT improves robustness on transformed test sets, achieving a median increase of 8.1 percentage points in clone detection and 3.6 in code classification, while maintaining or enhancing standard accuracy. The findings highlight the importance of multi-positive invariant contrastive learning as a key factor in improving model robustness.
Methodology
The methodology involves continued pretraining of code representation models using a code-only dataset. InvPT applies semantics-preserving transformations to the training corpus and employs a masked language modeling objective combined with a multi-positive supervised contrastive learning approach. This approach treats all augmentations of the same source function as positives, mixing self-contrast pairs with invariant-contrast pairs to provide varying difficulty levels during training.
Results
The implementation of InvPT resulted in a median improvement of 8.1 percentage points in clone detection tasks and 3.6 points in code classification tasks across all evaluated models and datasets. The method also achieved up to 6.57 points of standard accuracy on the original clone-detection benchmark. Ablation studies indicated that the multi-positive invariant contrastive learning was the primary contributor to these gains.
Implications
The findings suggest that InvPT can enhance the robustness of code representation models, making them more reliable for practical applications in software engineering tasks such as clone detection and code classification. This approach could lead to improved performance in real-world scenarios where code may be syntactically varied but semantically equivalent.
Lipschitz Bandits with Arbitrary Feedback Delays
Theory
Optimization
- Introduces algorithms for Lipschitz bandits with arbitrary feedback delays.
- Achieves regret bounds that account for both the zooming dimension and feedback delays.
- Demonstrates that feedback delays contribute an additional penalty to regret.
- Extends previous work by analyzing both stochastic and adversarial reward settings.
Read more
Lipschitz Bandits with Arbitrary Feedback Delays
Summary
This paper addresses the Lipschitz bandit problem, which extends traditional multi-armed bandit frameworks to continuous action spaces under the Lipschitz condition. The authors investigate the scenario where feedback is delayed, meaning that rewards are not received immediately after actions are taken. They propose two algorithms: an elimination-based algorithm for stochastic rewards and an EXP3-based algorithm for adversarial rewards, both designed to handle arbitrary feedback delays. The algorithms achieve a regret bound of ˜O(T^(d_z + 1)/(d_z + 2) + √D) over a time horizon T with total delay D, where d_z represents the zooming dimension. This work fills a gap in existing literature by analyzing Lipschitz bandits under more general delay models, showing that the additional penalty introduced by feedback delays is of order ˜O(√D). The results match existing delay-free regret guarantees, demonstrating the robustness of the proposed methods.
Methodology
The authors develop an elimination-based algorithm for stochastic rewards and an EXP3-based algorithm for adversarial rewards. They utilize a greedy-based analysis to characterize the impact of worst-case delay patterns and carefully select zooming schedules and weight-update procedures to optimize performance under feedback delays.
Results
The proposed algorithms achieve a regret bound of ˜O(T^(d_z + 1)/(d_z + 2) + √D) for both stochastic and adversarial settings. These bounds recover optimal rates from delay-free settings and highlight the additional regret incurred due to feedback delays, quantified as ˜O(√D).
Implications
The findings have significant implications for real-world applications where feedback delays are common, such as online learning systems, recommendation engines, and adaptive control systems. The results provide a theoretical foundation for designing more effective algorithms in environments with delayed feedback.
Variational Outlier-Robust Gaussian Process Regression with Generative Modeling
Theory
Generative Models
Robotics
- Introduces a generative GPR model that adapts to observation-specific outliers.
- Utilizes a variational generalized expectation-maximization approach for learning.
- Demonstrates improved prediction accuracy over robust GPR baselines in various contamination scenarios.
- Maintains cubic computational complexity, ensuring efficiency in large datasets.
Read more
Variational Outlier-Robust Gaussian Process Regression with Generative Modeling
Summary
This paper addresses the issue of outliers in Gaussian Process Regression (GPR), which can significantly distort model learning and predictions due to the conventional Gaussian observation likelihood. The authors propose a novel generative GPR model that captures observation-specific contamination and adaptively mitigates the influence of outliers. The methodology employs a variational generalized expectation-maximization procedure to learn latent variables and GPR model parameters. The proposed method is evaluated through experiments on both synthetic and real datasets under various contamination settings, demonstrating competitive performance against existing robust GPR baselines in terms of prediction accuracy. Additionally, the method maintains cubic computational scaling similar to traditional GPR methods, making it efficient for practical applications.
Methodology
The proposed methodology integrates observation-specific precision modeling with the latent Gaussian process function and its associated parameters. A variational generalized expectation-maximization procedure is employed to learn the latent variables and model parameters, with closed-form updates for most quantities except kernel hyperparameters, which are optimized using gradient-based methods.
Results
The experiments reveal that the proposed method outperforms several robust GPR baselines in terms of prediction accuracy across synthetic and real datasets, particularly under different contamination settings. The method's computational efficiency is also highlighted, as it retains cubic scaling similar to traditional GPR methods.
Implications
The findings suggest that the proposed outlier-robust GPR model can be effectively applied in various signal processing and machine learning applications where data contamination is a concern, enhancing the reliability and accuracy of predictions in real-world scenarios.
KV Cache Compression Through the Lens of Transform Coding
Large Language Models
Efficient ML
Theory
- Introduction of an attention-aware distortion measure that separates key and value quantization errors.
- Development of Attention-Aware Transform Coding (AATC) for optimal bit allocation in KV cache compression.
- AATC achieves near-lossless accuracy at approximately 5.8× compression on standard benchmarks.
- The method provides a unifying perspective on existing KV cache compression techniques.
Read more
KV Cache Compression Through the Lens of Transform Coding
Summary
This paper addresses the memory bottleneck posed by key-value (KV) caches in long-context inference for large language models (LLMs). Traditional quantization methods focus on minimizing reconstruction error within the cache but often overlook the propagation of this error through attention mechanisms. The authors introduce an attention-aware distortion measure that decomposes the expected distortion into contributions from keys and values, which vary across tokens and channels. They develop a novel method called Attention-Aware Transform Coding (AATC), which employs techniques from signal processing, specifically transform coding and reverse water-filling, to allocate bits optimally across a calibration set. The method is validated on two LLMs, Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct, across multiple benchmarks, demonstrating that AATC achieves near-lossless accuracy with approximately 5.8× compression, outperforming existing methods that degrade in various settings.
Methodology
The authors derive an attention-aware distortion measure that accounts for the impact of quantization on output. They utilize a reverse water-filling algorithm for bit allocation, combined with a whitening step to decorrelate key and value features. This approach allows for optimal channel-wise allocation of bits under the attention-aware distortion framework.
Results
AATC achieves near-lossless accuracy across all evaluated benchmarks (LongBench, RULER, GSM8K, MMLU-Pro, MATH-500) while providing approximately 5.8× compression. In contrast, baseline methods (KIVI, KVQuant, PALU) show degradation in performance on at least one benchmark.
Implications
The findings suggest that incorporating attention mechanisms into quantization strategies can significantly enhance the efficiency of KV cache compression in LLMs, potentially leading to more scalable and effective deployment of these models in real-world applications.