AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
69
Papers today
8h
Update frequency
7
Days of history
An Analysis of Residual-Stream Geometry Across Transformer Depth
NLP
Large Language Models
Interpretability
- Relative displacement of representations is structured by layer depth, with larger updates in early and late layers.
- The magnitude of global Procrustes rotation remains nearly constant across depth, while residuals peak at the final layer.
- Non-English targets exhibit greater final-layer displacement and residual compared to English targets.
- The findings are statistically robust and suggest that depth curves are largely stable across conditions and models.
Read more
An Analysis of Residual-Stream Geometry Across Transformer Depth
Summary
This paper presents a geometric analysis of the residual streams in transformer models, focusing on how representations change across different layers. The authors introduce a framework that utilizes relative displacement to measure the movement of representations between consecutive layers and applies orthogonal Procrustes analysis to decompose this movement into rigid rotations and non-rigid residuals. The study is conducted on six instruction-tuned models across tasks such as code generation and cross-lingual translation. Key findings include structured patterns of relative displacement that vary with depth, a nearly constant rotation magnitude across layers, and depth-modulated Procrustes residuals. The results indicate that the geometry of residual streams is model-dependent and stable across different conditions, suggesting a compact descriptive framework for understanding how transformer models process information.
Methodology
The authors employ a transition-centered geometric analysis framework, measuring relative displacement and applying orthogonal Procrustes analysis to characterize the transformations of the residual stream across layers in transformer models.
Results
The analysis reveals that relative displacement is larger in the early and late layers of the transformer, with a quieter middle third. The rotation magnitude is consistent across depths, while the Procrustes residuals and angles show depth modulation, peaking at the final transition. Additionally, the study finds that conditions primarily rescale late-layer amplitudes rather than altering the overall depth curve.
Implications
The findings provide insights into the geometric properties of transformer residual streams, which can inform the design and interpretation of transformer models in NLP tasks. Understanding these geometric regularities may lead to improved model architectures and training strategies.
Counterfactual Shapley Credit Assignment
Reinforcement Learning
Theory
Interpretability
- Introduction of Counterfactual Shapley Credit Assignment framework for RL.
- Demonstration of ϕ-values that preserve optimal policy while redistributing rewards.
- Development of a consistent estimator for ϕ-values with amortized constant time complexity.
- Introduction of ϕ-PPO with Prioritized Trajectory Replay, showing improved sample efficiency.
Read more
Counterfactual Shapley Credit Assignment
Summary
The paper addresses the Credit Assignment Problem (CAP) in Reinforcement Learning (RL), which is crucial for developing efficient and interpretable agents. Traditional methods often struggle to accurately attribute credit between an agent's skill and environmental randomness. The authors propose a novel framework called Counterfactual Shapley Credit Assignment, which utilizes the Counterfactual Shapley Value (ϕ-value) to redistribute rewards in a way that enhances credit assignment while preserving the optimal policy. This approach effectively tackles issues of sparse causality, high stochasticity, and delayed rewards. The authors derive a consistent estimator for ϕ-values that operates in constant time, leading to the development of a new policy gradient method, ϕ-PPO, combined with Prioritized Trajectory Replay (PTR). Empirical evaluations show that the ϕ-values align closely with the true causal contributions to task rewards, demonstrating superior sample efficiency in complex environments where existing methods fail to converge.
Methodology
The authors developed a framework based on the Counterfactual Shapley Value (ϕ-value) to address the CAP in RL. They derived a consistent estimator for ϕ-values that operates in amortized O(1) time per action, allowing for efficient computation. The new policy gradient method, ϕ-PPO, was introduced alongside Prioritized Trajectory Replay (PTR) to enhance learning efficiency.
Results
The empirical results indicate that the ϕ-values accurately reflect the ground truth causes of task rewards, achieving superior sample efficiency in environments characterized by sparse rewards and high stochasticity. The proposed methods outperformed existing state-of-the-art techniques, demonstrating faster convergence and better credit assignment.
Implications
This research has significant implications for the development of more efficient and interpretable RL agents, particularly in complex environments where traditional methods struggle. The framework can be applied to various RL applications, enhancing the understanding of causal relationships in decision-making processes.
HantaWatch: Federated Learning for Hantavirus Genomic Surveillance
Federated Learning
- HantaWatch enables decentralized federated learning for hantavirus genomic surveillance without centralizing raw data.
- The framework produces comprehensive outputs including predicted labels, risk scores, and expert-review priorities.
- Adaptive optimization techniques are employed to manage client drift in non-IID environments.
- Surveillance-specific model selection is based on multiple performance metrics beyond accuracy.
Read more
HantaWatch: Federated Learning for Hantavirus Genomic Surveillance
Summary
HantaWatch is a novel federated learning framework designed to enhance hantavirus genomic surveillance by enabling laboratories and surveillance sites to collaboratively train sequence-based models without sharing raw data. The framework addresses challenges such as distributed sequence data, non-IID source heterogeneity, and limited expert-review capacity. HantaWatch incorporates k-mer feature extraction, source-aware federated client construction, and an adaptive DU-FedProx optimization strategy. It also includes a surveillance-specific model selection process and a prediction-only triage system. The framework produces risk scores, confidence estimates, uncertainty flags, and ranked expert-review priorities, facilitating effective decision support for public health without replacing laboratory interpretations. Experiments demonstrate that HantaWatch can effectively perform high-risk screening, outbreak-associated prediction, clade classification, and clinical-syndrome categorization while balancing predictive performance and false-negative risks.
Methodology
The authors developed HantaWatch by formulating hantavirus genomic surveillance as a decentralized federated learning problem. They designed a source-aware non-IID client construction method to capture heterogeneity across various factors. An adaptive Deep-Unfolded FedProx controller was introduced to optimize local learning rates and reduce client drift. The framework also includes a model selection layer that evaluates models based on recall, false-negative rates, and other metrics.
Results
HantaWatch was evaluated on binary and multiclass tasks related to hantavirus surveillance. The results indicated that the framework effectively supports high-risk screening, outbreak prediction, and clade classification while maintaining a balance between predictive performance and false-negative risks. The outputs from HantaWatch facilitated prioritization for expert review, demonstrating its practical utility in genomic surveillance.
Implications
HantaWatch has significant implications for public health surveillance, particularly in managing low-frequency but high-impact zoonotic diseases. By enabling collaborative model training without data sharing, it addresses privacy and operational constraints, potentially improving the speed and accuracy of outbreak responses and risk assessments.
Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion
Theory
Efficient ML
- Introduction of target-aligned input reparameterization (TAIR) to improve neural network predictions of thermodynamic properties.
- TAIR replaces raw enthalpy inputs with thermodynamically informed coordinates, enhancing model learnability.
- Significant reductions in root-mean-squared error (RMSE) for temperature, density, and compressibility compared to baseline methods.
- Demonstrated effectiveness on supercritical methane-oxygen counterflow flame data, showing improved accuracy and efficiency.
Read more
Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion
Summary
This paper addresses the computational challenges associated with evaluating real-fluid thermodynamic properties in supercritical combustion simulations. The authors propose a novel approach called target-aligned input reparameterization (TAIR), which transforms the input coordinates for neural networks to improve the prediction of temperature, density, and compressibility coefficient from the solver-available state variables (enthalpy, pressure, and species mass fractions). By replacing the raw enthalpy coordinate with thermodynamically aligned coordinates, the networks can focus on learning the deviations from ideal-gas behavior rather than reconstructing the entire closure from raw inputs. The effectiveness of TAIR is demonstrated through assessments using supercritical methane-oxygen counterflow flame data, showing significant reductions in prediction errors compared to baseline methods. This approach not only enhances the accuracy of thermodynamic property evaluations but also reduces the computational burden associated with real-fluid simulations, making it a promising strategy for future applications in high-pressure propulsion systems.
Methodology
The authors developed a thermodynamics-informed input reparameterization strategy (TAIR) that transforms the input variables for neural networks. This involves using algebraic transformations to replace the raw enthalpy coordinate with target-matched thermodynamic coordinates, specifically a temperature estimate from an ideal-gas mixture and an ideal-gas density estimate. The neural networks are then trained to learn the deviations from these ideal-gas baselines, rather than the full closure from raw inputs.
Results
The implementation of TAIR resulted in a reduction of RMSE by approximately 1.5, 2.0, and 7.5 times for temperature, density, and compressibility, respectively, on held-out test sets. For an unseen strain-rate flame, the factors improved to 3.6, 14.5, and 6.0. The target-inconsistent controls performed significantly worse, confirming the benefits of the thermodynamically matched input design.
Implications
The findings suggest that the proposed TAIR method can significantly enhance the efficiency and accuracy of thermodynamic property evaluations in supercritical combustion simulations. This has potential applications in the design and optimization of high-pressure propulsion systems, such as liquid-propellant rocket engines, where accurate real-fluid thermodynamics are critical.
Beyond Output-Space Calibration: Spectral Evidence Bundling for Selective Reliability Estimation in Time-Series Classification
Time Series
- SEB-Cal introduces a novel approach to reliability estimation in time-series classification that goes beyond traditional output-space calibration.
- The method leverages spectral evidence from the entire sample to assess the trustworthiness of predictions.
- Validation-gated policy ensures that spectral conditioning is applied only when it improves correctness ranking without violating error constraints.
- Empirical results show significant improvements in selective reliability metrics across multiple datasets and backbone architectures.
Read more
Beyond Output-Space Calibration: Spectral Evidence Bundling for Selective Reliability Estimation in Time-Series Classification
Summary
This paper addresses the limitations of traditional post-hoc calibration methods in time-series classification, which often rely solely on output scores without considering the temporal evidence supporting those scores. The authors introduce SEB-Cal, a validation-gated reliability policy that estimates the trustworthiness of predictions based on spectral evidence from the input time series. SEB-Cal utilizes deterministic whole-sample spectral descriptors such as band energy, entropy, peak dominance, and phase stability to provide a scalar reliability estimate and band-level diagnostics. The method is designed to improve the reliability of predictions by ensuring that high-confidence scores are supported by organized temporal evidence. The authors evaluate SEB-Cal across eight diverse datasets and various backbone architectures, demonstrating that it significantly enhances selective reliability metrics while maintaining the original predicted labels. The results indicate that SEB-Cal not only improves calibration quality but also addresses the specific challenges of time-series reliability by focusing on the global organization of evidence in the input data.
Methodology
The authors developed SEB-Cal, which employs a validation-gated reliability policy that estimates the trustworthiness of predictions based on spectral descriptors derived from a whole-sample Fourier transform. These descriptors include band energy, entropy, peak dominance, and phase stability, which collectively assess the organization of evidence in the time series. The method maintains the original predicted labels while providing a reliability estimate and diagnostics.
Results
SEB-Cal demonstrated improvements in fixed-label selective-reliability metrics, increasing the correlation area under the receiver operating characteristic (Corr-AUROC) from 0.693 to 0.779, and further to 0.786 with the validation-gated policy. The method also reduced the false high-confidence error rate (FalseConf@0.9) to 0.094, indicating enhanced reliability in predictions across the evaluated datasets.
Implications
The findings suggest that incorporating spectral evidence into reliability estimation can significantly enhance the trustworthiness of time-series classification systems. This approach may be applicable in various domains where time-series data is prevalent, such as finance, healthcare, and environmental monitoring, potentially improving decision-making processes based on model predictions.
A Better Start for Language Models: Domain-Conditional Position Offsets
NLP
Large Language Models
Efficient ML
- Introduces domain-conditional position offsets to mitigate cold-start penalties in language models.
- Offsets are efficient, requiring minimal training data and parameters, while maintaining model weights frozen.
- Demonstrates significant reductions in in-domain perplexity across multiple language model architectures.
- Offsets improve retrieval reranking and domain classification without affecting few-shot reasoning tasks.
Read more
A Better Start for Language Models: Domain-Conditional Position Offsets
Summary
This paper addresses the cold-start penalty faced by autoregressive language models, particularly at the beginning of a sequence where context is minimal. The author introduces a novel approach called domain-conditional position offsets, which involves adding a learned vector to the embedding activation at the initial sequence positions while keeping all model weights frozen. This method is efficient, requiring only a small number of parameters and minimal training time on a limited dataset. The offsets significantly reduce in-domain perplexity across various models, demonstrating that they effectively propagate through the model state rather than merely recalibrating outputs. The paper also compares the proposed method with other adaptation techniques, highlighting its efficiency and effectiveness in improving retrieval reranking and domain classification tasks without negatively impacting few-shot reasoning accuracy. Overall, the study provides a comprehensive analysis of the cold-start penalty and offers a lightweight solution that enhances language model performance in domain-specific applications.
Methodology
The methodology involves training a single learned vector (the domain-conditional position offset) that is added to the token embedding at the first few positions of a sequence. This vector is trained using ordinary language modeling loss on a small sample of domain-specific text, with all model weights remaining unchanged. The offsets are applied during inference through a forward hook, allowing for quick domain switching without added latency.
Results
The introduction of domain-conditional position offsets led to a reduction in held-out in-domain perplexity by up to 27% across various models ranging from 410M to 8B parameters, with effects persisting even at 70B parameters. The method outperformed a matched direct logit-bias correction and showed improvements in retrieval and classification tasks that depend on early tokens.
Implications
The findings suggest that domain-conditional position offsets can serve as a lightweight and efficient tool for enhancing language model performance in specific domains, particularly in applications requiring rapid adaptation without extensive retraining. This approach may be beneficial in retrieval systems, classification services, and multi-agent systems where quick responses are critical.
Compressing What Matters: Neuron Importance Meets Data-Aware Low Rank Approximation for Language Model Compression
NLP
Large Language Models
Efficient ML
- Combines neuron importance estimation with data-aware low-rank approximation for improved model compression.
- Introduces a dynamic compression rate allocation algorithm for better distribution of compression across layers.
- Achieves state-of-the-art performance in compressing various foundational LLMs.
- Demonstrates that grouping weight matrices by layer index is more effective than previous methods.
Read more
Compressing What Matters: Neuron Importance Meets Data-Aware Low Rank Approximation for Language Model Compression
Summary
This paper addresses the challenge of compressing large language models (LLMs) to make them more applicable in resource-constrained environments. The authors propose a novel approach that combines neuron importance estimation with data-aware low-rank approximation to enhance the compression of neural networks. Previous methods focused either on parameter importance or functional equivalence in isolation, but this work integrates both perspectives into a unified framework. Additionally, the paper introduces a dynamic compression rate allocation algorithm that optimally distributes the compression rate across different layers and parameters, improving upon the traditional uniform allocation methods. Experimental results demonstrate that the proposed method achieves state-of-the-art performance in compressing models like BERT, DistilBERT, MobileBERT, and TinyBERT, particularly under high compression ratios, thus validating the effectiveness of the combined approach and the new allocation strategy.
Methodology
The authors utilize Singular Value Decomposition (SVD) for low-rank approximation of weight matrices in neural networks. They propose a hybrid approach that integrates neuron importance estimation and data-aware techniques, along with a novel algorithm for dynamic compression rate allocation based on layer indices. Experimental evaluations are conducted on several LLMs to assess performance improvements.
Results
The proposed method shows significant improvements in compression performance compared to previous state-of-the-art techniques, particularly under high compression ratios. The experimental results indicate that the new dynamic allocation strategy outperforms traditional uniform allocation methods, leading to better model performance on downstream tasks.
Implications
This research has important implications for deploying large language models in environments with limited computational resources, such as mobile devices and edge computing. The findings suggest that effective model compression techniques can maintain performance while significantly reducing memory and computational requirements.
Spatio-Temporal Prediction of Unsteady Airfoil Aerodynamics Using Augmented Graph Neural Ordinary Differential Equations with Exogenous Controls
Graph Learning
Time Series
Optimization
- Introduction of the GNODE framework for predicting unsteady aerodynamics.
- Augmentation of GNNs with Neural Ordinary Differential Equations improves prediction stability and accuracy.
- The model outperforms traditional autoregressive GNNs in handling complex aerodynamic phenomena.
- Demonstrated effectiveness on a dataset simulating a pitching airfoil with transonic shocks.
Read more
Spatio-Temporal Prediction of Unsteady Airfoil Aerodynamics Using Augmented Graph Neural Ordinary Differential Equations with Exogenous Controls
Summary
This paper presents a novel approach for predicting unsteady aerodynamic phenomena in transonic airfoil applications using a framework called Graph Neural Ordinary Differential Equations (GNODE). Traditional computational fluid dynamics (CFD) methods, while accurate, are computationally expensive and often limited by linear assumptions. The authors propose augmenting Graph Neural Networks (GNNs) with Neural Ordinary Differential Equations to enhance the temporal stability and accuracy of predictions. The GNODE framework effectively captures historical effects and reduces phase lag in aerodynamic responses. The study demonstrates that this method outperforms a time-discrete autoregressive GNN baseline in terms of accuracy and stability, particularly in scenarios involving complex aerodynamic behaviors such as transonic shocks and dynamic non-linearities. The results indicate that the augmented model can effectively handle the complexities of unsteady aerodynamics, making it a promising tool for future aircraft design and optimization.
Methodology
The authors developed the GNODE framework by combining Graph Neural Networks with Neural Ordinary Differential Equations. This approach allows for continuous-time predictions of aerodynamic forces while capturing historical effects through augmentation with additional latent dimensions. The model was trained and tested on a dataset derived from simulations of a pitching airfoil, focusing on transonic flow conditions.
Results
The GNODE framework provided temporally stable and spatially smoother predictions compared to the autoregressive GNN baseline. It demonstrated superior accuracy in predicting surface forces on the airfoil, effectively addressing the challenges posed by unsteady aerodynamic phenomena.
Implications
The findings suggest that the GNODE framework can serve as an efficient surrogate model for real-time predictions of unsteady aerodynamics, potentially reducing the reliance on computationally intensive CFD methods. This could significantly enhance the design and optimization processes in aerospace engineering, particularly for future aircraft configurations that require accurate modeling of dynamic non-linear behaviors.
A Reinforcement-Learning-Augmented Liquid-Fueled Reactor Network Model for Predicting Lean Blowout in Gas Turbine Combustors
Reinforcement Learning
- Introduction of a reinforcement learning framework for optimal reactor network generation.
- Goal-oriented clustering that improves lean blowout prediction accuracy.
- Demonstration of improved predictive fidelity and computational efficiency over traditional methods.
- First liquid-fueled reactor network capable of reproducing parametric LBO trends.
Read more
A Reinforcement-Learning-Augmented Liquid-Fueled Reactor Network Model for Predicting Lean Blowout in Gas Turbine Combustors
Summary
This paper presents a novel reinforcement learning (RL) framework designed to enhance the prediction of lean blowout (LBO) in gas turbine combustors using a liquid-fueled reactor network model. Traditional methods for determining reactor partitions often rely on manual heuristics or distance-based metrics, which can lead to suboptimal configurations, particularly for complex phenomena like LBO. The proposed RL framework introduces a goal-oriented approach that directly optimizes reactor configurations based on LBO prediction accuracy. The methodology involves a multi-stage process where initial clustering (e.g., k-means) generates micro-clusters, which are then refined by an actor-critic RL agent to create optimal reactor zones. Validation using a detailed Jet-A mechanism demonstrates that the RL framework significantly improves predictive accuracy and computational efficiency compared to conventional methods. This approach not only captures the correct trends in LBO behavior but also offers substantial speedups over high-fidelity computational fluid dynamics (CFD) models, making it a promising tool for rapid design-space exploration in gas turbine combustors operating near LBO limits.
Methodology
The methodology involves a multi-stage clustering-classification strategy where an initial k-means clustering generates micro-clusters, followed by the application of an actor-critic RL agent that merges these clusters into optimal reactor zones based on LBO prediction accuracy.
Results
The RL framework demonstrated improved predictive fidelity compared to k-means clustering, accurately capturing LBO trends and achieving significant computational speedups relative to high-fidelity CFD models.
Implications
The proposed RL-driven approach offers a computationally efficient reduced-order modeling technique that can significantly enhance the design and optimization processes for gas turbine combustors, particularly in scenarios where LBO is a critical concern.
SechKAN: Kolmogorov-Arnold Networks with Hyperbolic Secant Functions
Efficient ML
Theory
Computer Vision
- Introduction of SechKAN, a KAN architecture based on hyperbolic secant functions.
- SechKAN achieves competitive performance while using fewer parameters than existing KAN variants.
- The architecture is effective in function fitting, PDE problems, and image classification tasks.
- Comprehensive analysis of design choices provides practical guidelines for model configuration.
Read more
SechKAN: Kolmogorov-Arnold Networks with Hyperbolic Secant Functions
Summary
This paper introduces SechKAN, a novel architecture of Kolmogorov-Arnold Networks (KANs) that utilizes hyperbolic secant (sech) functions as basis functions. The motivation behind SechKAN is to address the limitations of traditional KANs, particularly their high parameter count and longer training times compared to multilayer perceptrons (MLPs). By employing 1D linear transformations, SechKAN reduces the number of parameters while maintaining competitive performance in various tasks such as function fitting, solving partial differential equations (PDEs), and image classification. The architecture is evaluated on benchmark datasets including MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100, demonstrating superior performance relative to MLPs and other KAN variants. The paper also provides a detailed analysis of design choices in SechKAN, including grid size, activation functions, and normalization strategies, offering practical guidelines for model configuration. Overall, SechKAN represents a significant advancement in the design of neural networks, combining efficiency with high performance.
Methodology
The authors propose SechKAN by utilizing hyperbolic secant functions as basis functions and incorporating 1D linear transformations to reduce the parameter count. The architecture is evaluated through experiments on function fitting, PDE problems, and image classification, with a focus on comparing performance metrics against MLPs and other KAN variants.
Results
SechKAN demonstrates superior performance in function fitting, PDE problems (specifically Navier–Stokes and Shallow Water equations), and image classification tasks on benchmark datasets. While its running time is slightly longer than that of MLPs, it outperforms other KAN variants in terms of accuracy and efficiency with a comparable number of parameters.
Implications
The development of SechKAN has potential implications for various applications in machine learning and scientific computing, particularly in scenarios where parameter efficiency and model performance are critical. Its design could inspire further research into alternative basis functions and architectures that enhance the capabilities of neural networks.
Quantifying Ranking Uncertainty in LLM Benchmarks
NLP
Large Language Models
- Introduces rank confidence intervals (CIs) to quantify uncertainty in model rankings.
- Analyzes sources of ranking uncertainty in the MMLU benchmark.
- Demonstrates substantial variability in rankings across MMLU subjects.
- Proposes modifications to hypothesis tests for better uncertainty estimation.
Read more
Quantifying Ranking Uncertainty in LLM Benchmarks
Summary
This paper addresses the uncertainty in rankings of pretrained models on multi-task leaderboards, particularly focusing on the Massive Multitask Language Understanding (MMLU) benchmark. The authors introduce rank confidence intervals (CIs) as a method to quantify this uncertainty, which has been overlooked in existing evaluations. They analyze the sources of ranking uncertainty and propose modifications to hypothesis tests to better account for these effects. The study reveals significant variability in rankings across different subjects within the MMLU benchmark, suggesting that such variability should be considered when comparing large language models (LLMs). By employing directional pairwise tests, the authors construct rank CIs that provide a more nuanced understanding of model performance, highlighting the importance of subject-level analysis in identifying top-performing models. The findings underscore the need for uncertainty quantification in model rankings to enhance the reliability of comparisons in the field of NLP.
Methodology
The authors construct rank confidence intervals using directional pairwise tests based on model performance scores. They analyze the impact of different statistical choices on the interpretation of rank intervals and control for family-wise error rates in hypothesis testing.
Results
The study finds that ranking variability is significant across different subjects in the MMLU benchmark, exceeding variability from prompt variants. The proposed rank CIs effectively communicate the uncertainty in model rankings, allowing for a more accurate comparison of LLMs.
Implications
The findings suggest that incorporating uncertainty quantification in model rankings can lead to more reliable assessments of model performance, guiding researchers and practitioners in selecting and comparing LLMs more effectively.
GUIDED Network-Agnostic Feature Initialization for Spatial Transferability in GNN-based Models
Graph Learning
Optimization
- Introduction of GUIDED, a network-agnostic feature initialization layer for GNNs.
- Enhances spatial transferability of GNN models for traffic assignment tasks.
- Maintains high predictive accuracy and robustness under data scarcity.
- Enables efficient domain adaptation without structural modifications.
Read more
GUIDED Network-Agnostic Feature Initialization for Spatial Transferability in GNN-based Models
Summary
This paper addresses the limitations of Graph Neural Networks (GNNs) in solving the Traffic Assignment Problem (TAP), which is crucial for transportation planning but computationally intensive. Traditional GNN approaches are constrained by a spatial generalization gap, as they rely on transductive feature initializations tied to specific network topologies. To overcome this, the authors propose a novel initialization layer called Geometrically Unconstrained Inductive Demand EmbeDding (GUIDED). This method injects travel demand as a scalar attribute on auxiliary virtual links rather than as fixed node features, allowing for a standardized input space across different urban environments. The integration of GUIDED with a Heterogeneous Graph Attention Network (HetGAT) model demonstrates that it maintains high predictive accuracy on single-network tasks while exhibiting robustness to out-of-distribution demand patterns. The proposed framework also facilitates efficient domain adaptation for inter-network transfer learning without requiring artificial input homogenization. Additionally, the optimized operations of the GUIDED layer lead to a significant reduction in training time, making it a versatile solution for various spatial interaction problems beyond vehicular traffic.
Methodology
The authors developed the GUIDED initialization layer, which reframes demand initialization as a localized graph aggregation problem using virtual links in a heterogeneous graph. This approach standardizes input dimensions across different networks, allowing for seamless transfer of models without structural changes. The GUIDED layer was integrated into a state-of-the-art HetGAT model and evaluated across multiple real-world urban networks.
Results
The integration of the GUIDED layer with the HetGAT model resulted in state-of-the-art predictive accuracy on single-network tasks and demonstrated superior robustness to out-of-distribution demand patterns. The model maintained a performance advantage even under severe data scarcity, and the initialization layer led to a 50% reduction in training time per epoch compared to baseline approaches.
Implications
The GUIDED framework provides a robust foundation for inductive models in traffic assignment and can be applied to various spatial interaction problems, such as freight logistics and multimodal network optimization. Its network-agnostic nature allows for broader applicability in transportation planning and decision-making processes.
Program Synthesis for Simulation-Based Inference: Joint Model Selection and Parameter Estimation
NLP
Large Language Models
Theory
- Introduces a framework for joint model selection and parameter estimation using LLMs and SBI.
- LLMs generate candidate simulator programs from natural language descriptions, enhancing model exploration.
- The method iteratively refines models and evaluates them using neural density estimation.
- Validated on benchmarks with known ground truth, showing effective model identification.
Read more
Program Synthesis for Simulation-Based Inference: Joint Model Selection and Parameter Estimation
Summary
This paper presents a novel framework that integrates large language models (LLMs) for program synthesis with neural simulation-based inference (SBI) to facilitate joint model selection and parameter estimation. Traditional SBI methods require a fixed model structure, but this approach allows for the exploration of a diverse pool of candidate models derived from natural language descriptions of the system under investigation. The framework utilizes LLMs to generate executable simulator programs, which are then refined through feedback-driven mutation. The candidate models are evaluated using neural density estimation, enabling the identification of plausible model families based on the data's information content. The method is validated on various benchmarks, including deterministic dynamics, stochastic epidemic models, and gravitational-lensing images, demonstrating its ability to accurately reflect model identifiability and data relevance. This work addresses the challenge of navigating the vast space of scientific models by automating model discovery and inference processes.
Methodology
The proposed framework synthesizes multiple candidate simulator programs using an LLM based on a natural language description and dataset. It generates simulations from these models, trains neural classifiers to estimate log Bayes factors for model comparison, and fits neural posterior estimators for parameter inference. The process iterates to refine models and produce a joint posterior over both model and parameter spaces.
Results
The framework successfully identifies plausible model families across various benchmarks, demonstrating accuracy that correlates with the information content of the data and the identifiability of candidate models. The iterative refinement process enhances the robustness of model selection and parameter estimation.
Implications
This work has significant implications for scientific modeling, allowing researchers to automate the discovery of complex models and improve the efficiency of parameter estimation in scenarios where traditional methods are limited. It opens new avenues for applying machine learning in fields requiring sophisticated model inference.
Automated Data Engineering and Feature Selection for the Case Study of Warpage Detection in Fused Deposition Modeling
Reinforcement Learning
Interpretability
Efficient ML
- Development of an Automated Data Processing framework for FDM warpage detection.
- Utilization of reinforcement learning for dynamic model-feature selection.
- Implementation of SHAP XAI for effective feature subset generation.
- Significant improvements in model accuracy and stability demonstrated.
Read more
Automated Data Engineering and Feature Selection for the Case Study of Warpage Detection in Fused Deposition Modeling
Summary
This paper presents an Automated Data Processing (ADP) framework aimed at optimizing machine learning model-feature combinations for predictive tasks in fused deposition modeling (FDM) datasets, specifically for warpage detection. The methodology employs a reinforcement learning-inspired policy updating mechanism, where multiple machine learning models are trained on both complete feature sets and subsets selected using Shapley-based Explainable AI (SHAP XAI) across 217 datasets. The framework evaluates predictive accuracy and F1-scores at each episode, computes scalar rewards, and updates Q values to refine future model selections. The use of SHAP XAI allows for the generation of reduced yet informative feature subsets, facilitating exploration of model performance with varying dimensionality. Results demonstrate that the ADP framework effectively converges towards optimal model-feature configurations, achieving significant improvements in accuracy and stability. The test-set AUC improved from 0.9248 to 0.9731, and the mean reward value increased by over fifty percent compared to the baseline full-feature configuration, indicating the framework's efficacy in enhancing predictive performance while managing feature complexity.
Methodology
The study employs a reinforcement learning-inspired approach to update policies for selecting machine learning models and features. It trains models on both full and SHAP XAI-selected feature subsets across 217 datasets, assessing predictive accuracy and F1-scores to compute rewards that guide future selections.
Results
The ADP framework led to an increase in test-set AUC from 0.9248 to 0.9731 and a more than fifty percent increase in the mean reward value compared to the baseline configuration, indicating enhanced model performance and stability.
Implications
The findings suggest that the ADP framework can significantly improve machine learning applications in manufacturing, particularly in real-time monitoring and defect detection, by automating feature selection and model optimization processes.
Unsupervised Multi-kernel Learning for Automated Algorithm Selection
Optimization
- Introduces an unsupervised multi-kernel learning framework for algorithm selection, avoiding the pitfalls of supervised models.
- Utilizes a multi-kernel k-means approach to cluster problem instances based on heterogeneous landscape representations.
- Demonstrates superior performance on Differential Evolution tasks and competitive results on Particle Swarm Optimization tasks.
- Provides interpretable insights into the relevance of different landscape representations for algorithm selection.
Read more
Unsupervised Multi-kernel Learning for Automated Algorithm Selection
Summary
This paper addresses the challenge of automated algorithm selection (AAS) in black-box optimization, which traditionally relies on supervised models that map landscape features to algorithm performance labels. The authors propose an unsupervised approach using multi-kernel clustering over heterogeneous landscape representations, allowing for the grouping of problem instances without performance labels during the clustering phase. The methodology employs a multi-kernel k-means formulation that learns cluster assignments and kernel weights from four landscape views: Exploratory Landscape Analysis (ELA), DeepELA, DoE2Vec, and TransOptAS. The clusters are then mapped to solver recommendations in a post hoc manner. The study evaluates the proposed method on portfolios for Differential Evolution (DE) and Particle Swarm Optimization (PSO) using a strict three-stage evaluation protocol. The results indicate that the multi-kernel clustering approach outperforms existing methods on the DE portfolio and remains competitive on the PSO portfolio, demonstrating the effectiveness of the unsupervised learning strategy. Additionally, the learned kernel weights provide insights into which landscape representations are most relevant for algorithm selection.
Methodology
The authors employ a multi-kernel k-means clustering approach that integrates multiple heterogeneous landscape representations without using performance labels during the clustering process. The clusters are formed based on landscape characteristics, and performance data is utilized only after clustering for mapping clusters to solver recommendations.
Results
The multi-kernel clustering method achieved the strongest mean selector profile on the Differential Evolution portfolio and showed competitive performance on the Particle Swarm Optimization portfolio. The learned kernel weights indicated a preference for ELA and TransOptAS representations, while assigning zero weight to DeepELA and DoE2Vec, highlighting task-specific relevance.
Implications
This research could significantly impact the field of automated algorithm selection by providing a more generalizable and cost-effective approach to selecting optimization algorithms, potentially enhancing the performance of AutoML systems in diverse problem settings.
Neural Controlled Differential Equations for EMT-Level Surrogate Modeling of Grid-Forming Inverters
Time Series
- Introduction of Neural CDE framework for continuous-time surrogate modeling of grid-forming inverters.
- Affine-control decomposition to separate system dynamics from control effects.
- Dual-path control embedding for capturing multi-time-scale behaviors.
- Physics-informed regularization to enhance stability and coherence in simulations.
Read more
Neural Controlled Differential Equations for EMT-Level Surrogate Modeling of Grid-Forming Inverters
Summary
This paper introduces a Neural Controlled Differential Equation (Neural CDE) framework aimed at enhancing the surrogate modeling of grid-forming inverters for electromagnetic transient (EMT) simulations. The authors identify challenges in existing AI-based modeling approaches, particularly in handling multi-time-scale dynamics and ensuring physics-aware evaluations. The proposed Neural CDE framework allows for continuous-time modeling, accommodating discretely sampled control signals and capturing the hierarchical behavior of inverter dynamics. Key innovations include an affine-control formulation that separates intrinsic system dynamics from control-driven effects, and a dual-path control embedding that enhances model expressiveness and stability. The model is evaluated against EMT-generated trajectories, demonstrating its ability to accurately reproduce transient responses, maintain effective damping, and ensure stability over long-horizon rollouts. The findings suggest that Neural CDEs provide a robust and physically consistent approach for EMT-level simulations, addressing the limitations of traditional methods and enabling more complex system analyses.
Methodology
The authors developed a Neural CDE framework that models grid-forming inverter dynamics by parameterizing control paths and utilizing an affine-control formulation. This approach allows for continuous-time learning from discretely sampled control signals and incorporates a dual-path control mechanism to enhance expressiveness and stability. The model is evaluated using EMT-generated trajectories to assess its performance in reproducing transient behaviors.
Results
The Neural CDE framework successfully reproduced the transient responses of grid-forming inverters, preserving effective damping and dominant oscillatory characteristics. The model maintained stability over long-horizon rollouts, demonstrating its capability to handle complex dynamics in large-scale power systems.
Implications
The proposed Neural CDE framework has significant implications for the modeling of grid-forming inverters in power systems, enabling more efficient and accurate simulations of large-scale energy storage systems. This approach can facilitate the integration of renewable energy sources into the grid by improving the analysis of multi-machine interactions and enhancing the reliability of power electronic converters.
L1 Augmented Attention as an Improved Vector Similarity Metric
NLP
Large Language Models
- L1-augmented attention improves upon traditional dot-product attention by incorporating L1 distance metrics.
- The method captures both directional alignment and coordinate-wise deviations for better similarity assessment.
- Evaluation on WikiText-2 shows a reduction in perplexity by up to 14.5% compared to the baseline.
- The approach allows for computational efficiency through low-dimensional projections of queries and keys.
Read more
L1 Augmented Attention as an Improved Vector Similarity Metric
Summary
This paper addresses the limitations of the scaled dot-product attention mechanism used in Transformer models, which conflates directional alignment and vector magnitude, thereby reducing its effectiveness as a similarity metric. The author proposes a novel approach called L1-augmented attention, which modifies the traditional attention mechanism by subtracting a learned, head-specific L1 distance between queries and keys from the dot-product score. This hybrid similarity metric captures complementary geometric information, where the dot-product rewards directional alignment and the L1 norm penalizes coordinate-wise deviations. To enhance computational efficiency, the method projects queries and keys into low-dimensional subspaces that specialize in preserving informative L1 structures. The proposed L1-augmented attention was evaluated on the WikiText-2 dataset using a compact Transformer model, achieving a notable reduction in perplexity by up to 14.5% compared to the baseline model. The analysis of norm variance and learned L1 weights indicates distinct geometric roles across layers and strong specialization at the head level. Overall, the findings suggest that incorporating L1 geometry into attention mechanisms significantly improves similarity computation in language models, offering practical benefits in terms of both accuracy and parallel efficiency.
Methodology
The paper introduces L1-augmented attention by modifying the standard attention mechanism to include a learned L1 distance term. This involves projecting queries and keys into low-dimensional subspaces to maintain L1 structure while ensuring computational efficiency. The performance of this method is evaluated on the WikiText-2 dataset using a compact Transformer architecture.
Results
The L1-augmented attention mechanism achieved up to a 14.5% reduction in perplexity on the WikiText-2 dataset compared to the traditional Vaswani baseline. The analysis revealed distinct geometric roles across layers and strong specialization at the head level, indicating the effectiveness of the proposed method.
Implications
The findings suggest that enhancing attention mechanisms with L1 geometry can lead to more accurate similarity computations in language models, potentially improving performance in various NLP tasks. This approach may also inform future research on attention mechanisms and their applications in Transformer-based architectures.
From Outcomes to Actions: Leveraging Hindsight for Long-Horizon Language Agent Training
NLP
Large Language Models
Reinforcement Learning
- Introduction of Hindsight Policy Optimization (HPO) to improve long-horizon RL training.
- HPO utilizes an intent space to reduce optimization variance by aggregating semantically similar actions.
- The method employs Wasserstein distance to measure discrepancies between policy distributions.
- Empirical results show HPO significantly reduces training noise and stabilizes optimization.
Read more
From Outcomes to Actions: Leveraging Hindsight for Long-Horizon Language Agent Training
Summary
This paper addresses the challenges faced by reinforcement learning (RL) methods in training language agents for long-horizon interactions. The authors introduce a novel policy gradient method called Hindsight Policy Optimization (HPO), which aims to reduce optimization variance by leveraging hindsight distributions in an intent space. Traditional RL methods struggle with long-horizon tasks due to difficulties in attributing outcomes to sequences of interdependent actions, often leading to high optimization variance. HPO projects both the current policy and hindsight distributions into a lower-dimensional intent space, measuring their discrepancy using Wasserstein distance. This approach aggregates semantically similar states and actions, resulting in a bounded-variance estimator that enhances policy performance. The authors provide both theoretical and empirical evidence supporting the effectiveness of HPO, demonstrating its ability to stabilize training and improve sample efficiency across multiple long-horizon tasks.
Methodology
The authors reformulate existing Monte Carlo policy gradient methods by introducing HPO, which constructs hindsight distributions in an intent space. The method uses Wasserstein distance to measure discrepancies between the current policy and hindsight distributions, allowing for a more stable and efficient optimization process.
Results
HPO was evaluated on various long-horizon tasks, showing substantial reductions in training noise and improved stability in optimization. The results indicate that HPO provides interpretable learning signals that facilitate effective policy learning, even in the absence of outcome rewards.
Implications
The findings suggest that HPO can enhance the training of language agents in complex decision-making environments, potentially leading to more robust applications in areas such as tool use, smart device operation, and interactive environments.
MXSens: Sensitivity-Aware Mixed-Precision Quantization for Efficient LLM Inference
Large Language Models
Efficient ML
- MXSens introduces a sensitivity-guided approach to mixed-precision quantization, optimizing bitwidth allocation based on layer and column sensitivity.
- The method is training-free and utilizes the MXINT format for efficient low-bit inference.
- MXSens achieves state-of-the-art performance on LLaMA models, significantly improving perplexity metrics.
- The approach addresses the challenges posed by outliers in activations, which are common in low-bit quantization.
Read more
MXSens: Sensitivity-Aware Mixed-Precision Quantization for Efficient LLM Inference
Summary
The paper presents MXSens, a novel method for mixed-precision quantization aimed at improving the efficiency of large language model (LLM) inference. Traditional 4-bit quantization techniques often lead to significant accuracy degradation due to the presence of outliers in activations. Existing solutions either rely on software-managed scaling or mixed-precision integer quantization, which can incur substantial overhead. MXSens addresses these issues by introducing a sensitivity-aware approach that assigns varying mantissa bitwidths (4, 6, or 8 bits) based on the sensitivity of different model layers and columns. This method is training-free and leverages the block-wise structure of microscaling formats like MXINT, allowing for efficient implementation without the need for model retraining. The authors demonstrate that MXSens outperforms state-of-the-art quantization methods across various models and tasks, achieving significant improvements in perplexity metrics on the WikiText-2 dataset. This work establishes a new balance between accuracy and resource efficiency in LLM quantization.
Methodology
MXSens employs a sensitivity analysis to categorize outliers in model activations and assigns mixed mantissa bitwidths based on this sensitivity. The method leverages the block-wise structure of the MXINT format to facilitate efficient mixed-precision implementation, avoiding the overhead associated with traditional software-managed scaling.
Results
Under the W4A4KV4 setting, MXSens achieves perplexities of 3.77 and 7.63 on the LLaMA-2-70B and LLaMA-3-8B models, respectively, on the WikiText-2 dataset, outperforming existing quantization baselines.
Implications
The findings suggest that MXSens can significantly enhance the efficiency of LLM inference in practical applications, making it a valuable tool for deploying large language models in resource-constrained environments. This method could lead to broader adoption of low-bit quantization techniques in various AI applications.
Spectral Higher-Order Neural Networks Have Sharp Expressivity Bounds
Theory
Efficient ML
Interpretability
- SHONNs reduce parameter scaling to O(N^2) through spectral parameterization and weight sharing.
- They can operate without enforcing traditional non-linearities, enhancing model expressiveness.
- The framework was benchmarked on N-bit parity tasks, demonstrating improved performance.
- SHONNs offer a highly tunable hypothesis space, making them versatile for various applications.
Read more
Spectral Higher-Order Neural Networks Have Sharp Expressivity Bounds
Summary
This paper introduces Spectral Higher-Order Neural Networks (SHONNs), a novel approach to neural hypergraphs that addresses the parameter explosion issue associated with traditional higher-order neural networks (HONNs). By leveraging spectral properties and a weight-sharing scheme, SHONNs achieve a significant reduction in computational cost while maintaining a compact parameter scaling of O(N^2). The authors demonstrate the versatility and tunability of SHONNs through rigorous benchmarking on N-bit parity tasks, a challenging testbed for neural architectures. The findings suggest that SHONNs can operate without the need for ad hoc non-linearities, allowing for enhanced expressiveness and interpretability compared to standard neural networks. The paper concludes with a discussion on future research directions and the potential for SHONNs to advance the field of machine learning.
Methodology
The authors propose a new parameterization for triadic HONNs that utilizes spectral attributes to reduce the number of parameters needed. They conduct a mathematical analysis of the forward propagation in SHONNs and benchmark their performance on N-bit parity tasks to evaluate expressivity and interpretability.
Results
The results indicate that SHONNs outperform traditional HONNs in terms of computational efficiency and expressiveness. The ability to fine-tune the model's expressiveness without the need for additional non-linearities is a significant advantage, as evidenced by their performance on the N-bit parity tasks.
Implications
The findings suggest that SHONNs could be applied in various domains requiring efficient and interpretable models, potentially impacting areas such as complex system modeling, data analysis, and advanced AI applications.
Elicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent Posterior
NLP
Large Language Models
Optimization
- Introduces Posterior Prefix Tuning (PPT) for eliciting transformer behavior without backpropagation.
- Utilizes latent posterior models to optimize prompts based on expected utility.
- Demonstrates efficiency by using prior samples for multiple utility functions without additional transformer calls.
- Validates the method on Beta–Bernoulli and reinforced urn BFTs across different utility families.
Read more
Elicitation without Backpropagation: Steering Model Behavior by Optimizing the Latent Posterior
Summary
This paper introduces a novel method called Posterior Prefix Tuning (PPT) for eliciting desired behavior from transformer models without the need for backpropagation. The authors leverage the latent posterior model of transformer behavior, where the next-token distribution is derived from a posterior over latent predictive models conditioned on the context. By focusing on Bayes-filtered transformers (BFTs) that are meta-learned on sequences from a hierarchical prior, PPT optimizes prompts to maximize expected utility based on a given utility function. The method operates entirely in latent-posterior space, allowing for efficient gradient estimation through importance sampling from prior samples. This approach eliminates the need for transformer forward passes and backpropagation, enabling the same set of prior samples to be used for multiple utility functions at minimal additional cost. The authors validate PPT on various BFTs and utility families, demonstrating its effectiveness in steering model behavior while avoiding the computational overhead associated with traditional methods.
Methodology
The methodology involves optimizing a distribution over hard prompts using a gradient estimator derived from the latent posterior. The authors perform predictive Monte Carlo sampling to gather prior samples, which are then used to estimate gradients through importance sampling, allowing for the optimization of prompts without direct interaction with the transformer model.
Results
The experimental results indicate that PPT and its Rao–Blackwellized variant (PPT-RB) effectively optimize prompts across various utility families and BFTs. The method significantly outperformed the Greedy Coordinate Gradient (GCG) baseline in certain scenarios, particularly in the reinforced urn task and reverse cross-entropy, while showing competitive performance in others.
Implications
The findings suggest that PPT can provide a more efficient and scalable approach to controlling transformer behavior, which is crucial for applications in AI safety and alignment. This method could facilitate the development of more reliable AI systems by allowing for better elicitation of desired behaviors without the computational burden of traditional training methods.
WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning
Reinforcement Learning
Optimization
Efficient ML
- WAR optimizes rollout generation in synchronous agentic RL by adapting to workload conditions.
- Under low load, it employs SuffixDecoding for efficient speculative decoding without additional model overhead.
- Under high load, it utilizes cache-aware scheduling to enhance resource utilization and reduce computation redundancy.
- WAR achieves throughput improvements of 1.4× under low load and up to 1.6× under high load.
Read more
WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning
Summary
The paper introduces WAR, a workload-aware rollout system designed to enhance the efficiency of synchronous agentic reinforcement learning (RL) by optimizing both decoding and scheduling based on runtime load. The authors identify that long-horizon rollout generation is a significant bottleneck in agentic RL due to the increasing complexity and length of trajectories, which can grow to tens of thousands of tokens. WAR employs two main strategies: under low load conditions, it utilizes a model-free speculative decoding method called SuffixDecoding, which reuses suffix patterns from previous trajectories to create speculative drafts for future rollouts, thereby avoiding GPU contention. Under high load conditions, WAR focuses on cache-aware scheduling, where a global scheduler allocates requests to rollout replicas based on factors like cache locality and server load, reducing redundant computations and improving resource utilization. The combination of these strategies allows WAR to improve rollout throughput significantly without altering the underlying RL algorithm.
Methodology
The methodology involves a dual approach: using SuffixDecoding for speculative decoding in low-load scenarios and implementing cache-aware scheduling in high-load scenarios. This allows the system to adaptively optimize rollout generation based on the current workload, improving efficiency and throughput.
Results
WAR demonstrated a 1.4× improvement in long-context agentic rollout throughput under low load conditions and up to 1.6× under high load conditions, effectively alleviating the rollout bottleneck in synchronous RL training.
Implications
The findings suggest that WAR can significantly enhance the scalability and efficiency of training long-context agents in reinforcement learning, making it a practical solution for real-world applications that require extensive interaction with complex environments.
ROMS-IMLE: A Minimalist Approach to Competitive Single-Step Generative Modelling
Generative Models
- Introduces a minimalist approach to generative modeling, focusing on simplicity and efficiency.
- Utilizes Implicit Maximum Likelihood Estimation (IMLE) for a stable and effective training objective.
- Achieves an FID of 2.56 on ImageNet 256 with a single forward pass, demonstrating high sample quality.
- Reduces the number of parameters and function evaluations compared to state-of-the-art models.
Read more
ROMS-IMLE: A Minimalist Approach to Competitive Single-Step Generative Modelling
Summary
The paper introduces ROMS-IMLE, a minimalist single-step generative model that challenges the prevailing belief that complex iterative transformations are necessary for high-quality image generation. The authors argue that many recent advancements in generative modeling, particularly in diffusion and flow-matching models, have led to unnecessary complexity. Instead, they propose a simpler approach by utilizing Implicit Maximum Likelihood Estimation (IMLE) as the training objective and a convolutional network architecture. The model incorporates essential elements such as per-stage supervision and spectral specialization, which are typically associated with more complex models, while avoiding iterative denoising. The results demonstrate that ROMS-IMLE achieves competitive performance, producing high-quality samples with significantly fewer parameters and function evaluations compared to existing models. The findings suggest that a straightforward single-step generative model can effectively retain the quality and diversity of samples without the computational burden of iterative sampling.
Methodology
The authors developed ROMS-IMLE by starting with a simple training objective (IMLE) and a convolutional network architecture. They incorporated essential components from modern generative models, such as per-stage supervision and spectral specialization, while avoiding the complexity of iterative denoising. The model was trained on diverse datasets to evaluate its performance.
Results
ROMS-IMLE achieved an FID score of 2.56 on ImageNet 256, demonstrating high-quality image generation. It also showed superior precision and recall on CIFAR-10 and CelebA-HQ datasets compared to baseline models, while using 54% fewer parameters and requiring 250 times fewer function evaluations than more complex models like DiT-XL/2 and SiT-XL/2.
Implications
The findings suggest that simpler generative models can achieve competitive performance, potentially leading to more efficient and accessible generative modeling techniques. This could have applications in various fields such as computer vision, art generation, and real-time image synthesis.
Agentic Calibration of Grey-Box Simulation Models: An LLM-Driven Alternative
Optimization
Large Language Models
Interpretability
- Introduces an LLM-driven optimization method for calibrating grey-box simulation models in CEA.
- Achieves lower median best error with fewer model evaluations compared to traditional methods.
- Incorporates constraints through natural language prompts, eliminating the need for complex surrogate models.
- Provides an auditable optimization trace, enhancing transparency and justification of modeling choices.
Read more
Agentic Calibration of Grey-Box Simulation Models: An LLM-Driven Alternative
Summary
This paper presents a novel approach to calibrating grey-box simulation models used in cost-effectiveness analysis (CEA) by employing a large language model (LLM) as an optimizer. Calibration is a challenging optimization problem due to the high-dimensional parameter space, expensive model evaluations, and the need to respect clinical plausibility constraints. Traditional methods like Nelder-Mead (NM) and Bayesian Optimization (BO) have limitations in efficiency and constraint handling. The proposed agentic calibration method leverages the LLM's ability to process natural language descriptions of the model, including parameters and constraints, allowing it to propose parameter vectors with rationales at each iteration. This method not only improves sample efficiency but also provides an auditable trace of the optimization process, making it suitable for regulated environments. The evaluation on a 10-dimensional anal cancer simulation model shows that the agentic method outperforms NM and achieves competitive results compared to BO, requiring significantly fewer model evaluations while maintaining comparable error levels under constraints.
Methodology
The agentic calibration method utilizes an LLM to optimize parameter vectors by providing it with historical data on parameter guesses, simulated outputs, calibration targets, and a structured breakdown of residuals. Constraints are integrated into the optimization process via natural language prompts, allowing the LLM to reason about the calibration effectively without additional modeling machinery.
Results
The agentic method achieved substantially lower median best error than BAxUS and NM in unconstrained calibration, requiring only 16 model evaluations compared to 110 for BAxUS and hundreds to thousands for NM. In constrained calibration, it maintained comparable error levels while outperforming NM, with no extra implementation needed for constraints.
Implications
This approach offers a promising alternative for calibrating simulation models in health-economic decision-making, particularly in contexts where simulation time is a limiting factor. The ability to provide rationales for optimization decisions enhances transparency, making it suitable for regulated assessments in health technology.
AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters
NLP
Large Language Models
Efficient ML
- ADAFLASH addresses high variance in draft quality from diffusion drafters.
- The framework includes an on-policy distillation algorithm for stable convergence.
- An adaptive length head reduces verification costs by dynamically adjusting sequence lengths.
- Experimental results show significant performance improvements over previous methods.
Read more
AdaFlash: Adaptive Speculative Decoding via On-Policy Distilled Diffusion Drafters
Summary
The paper introduces ADAFLASH, a novel framework aimed at enhancing the efficiency of speculative decoding in large language models (LLMs) by addressing the high variance issues associated with diffusion drafters. Speculative decoding accelerates inference by using a lightweight draft model to generate sequences that are verified by a larger target model. While diffusion drafters have shown promise due to their parallel generation capabilities, they suffer from significant variance in acceptance rates across different domains and token positions. The authors propose two main components in ADAFLASH: an on-policy distillation (OPD) algorithm that stabilizes convergence and reduces domain-level variance, and an adaptive length head that dynamically adjusts the candidate sequence length to optimize verification costs. The framework is designed to operate efficiently in high-concurrency scenarios, where traditional methods often falter. Experimental results demonstrate that ADAFLASH outperforms existing state-of-the-art methods, achieving up to 5.3× speedup over standard autoregressive decoding and up to 66% higher throughput in high-concurrency environments.
Methodology
The ADAFLASH framework consists of two main components: (1) an on-policy distillation (OPD) algorithm utilizing reverse-KL divergence to stabilize convergence and reduce domain-level variance, and (2) an adaptive length head that adjusts the candidate sequence length dynamically during inference. The authors also implement infrastructure improvements for efficient online adaptation.
Results
The experiments conducted across eight benchmarks and three different foundation models indicate that ADAFLASH consistently outperforms previous speculative decoding methods, achieving up to 5.3× speedup over standard autoregressive decoding and up to 66% higher throughput in high-concurrency scenarios.
Implications
The findings suggest that ADAFLASH can significantly enhance the efficiency of large language model inference, making it more feasible for real-time applications and scenarios requiring high throughput. This could lead to broader adoption of diffusion-based models in practical applications.
HindsightBench: A Black-Box Behavioral Audit Protocol for Parametric Hindsight in Time-Indexed LLM Decision Tasks
Large Language Models
NLP
Time Series
- HindsightBench allows for cost-effective auditing of LLMs without requiring access to training data or backtesting.
- The protocol integrates multiple components into a single causal design for comprehensive behavioral analysis.
- Key findings include the influence of training generation on model behavior and the variability of effective cutoffs across models.
- Audit results are sensitive to serving configurations, necessitating specific operational requirements.
Read more
HindsightBench: A Black-Box Behavioral Audit Protocol for Parametric Hindsight in Time-Indexed LLM Decision Tasks
Summary
The paper introduces HindsightBench, a novel black-box behavioral audit protocol designed to assess parametric hindsight in time-indexed decision tasks involving large language models (LLMs). The protocol enables users to audit models without requiring access to training data, log probabilities, or backtesting infrastructure, thus providing a cost-effective solution for practitioners. HindsightBench integrates various components into a single causal design, including a four-arm date-manipulation matrix and dual memory probes, allowing for a comprehensive analysis of model behavior. The study evaluates 15 models from seven vendors, revealing key patterns in model behavior, such as the influence of training generation on hindsight reflexes and the variability of effective cutoffs across different models. The findings highlight the importance of serving configurations on audit results and provide operational requirements to ensure accurate assessments. The authors release the necessary resources for replication and further research, emphasizing the protocol's potential to enhance understanding of LLM decision-making processes.
Methodology
The methodology involves a black-box behavioral audit protocol that utilizes a four-arm date-manipulation matrix and dual memory probes to assess model behavior. The protocol measures various metrics related to hindsight and effective cutoffs, enabling a detailed analysis of model performance across different configurations.
Results
The application of HindsightBench to 15 models revealed that the date-trigger reflex is linked to training generation rather than model scale. Additionally, behaviorally effective cutoffs were found to span 22 months and could precede vendor-reported dates by up to eight months. The results also indicated that audit outcomes are not invariant to serving details, highlighting the need for specific operational requirements.
Implications
The findings from HindsightBench have significant implications for practitioners deploying LLMs in financial decision-making tasks. The protocol provides a framework for understanding and mitigating the effects of hindsight bias, ultimately improving the reliability of model outputs in time-sensitive contexts.
Hybrid Latent-Structural Fusion (HLSF) for Cyber Anomaly Detection
Theory
- Introduction of Hybrid Latent-Structural Fusion (HLSF) framework for anomaly detection.
- Integration of CP-APR and normalizing flows improves detection performance.
- HLSF demonstrated superior results on real-world cyber anomaly data from LANL.
- The methodology effectively captures both structural and distributional aspects of cyber behavior.
Read more
Hybrid Latent-Structural Fusion (HLSF) for Cyber Anomaly Detection
Summary
The paper addresses the challenge of detecting malicious anomalous activities in cybersecurity systems by proposing a novel framework called Hybrid Latent-Structural Fusion (HLSF). This framework integrates two powerful unsupervised machine learning techniques: CANDECOMP-PARAFAC alternating Poisson regression (CP-APR) for tensor decomposition and normalizing flows for density estimation. The authors demonstrate that HLSF effectively combines the structural anomaly scores from CP-APR with the latent-space density scores from normalizing flows, leading to improved anomaly detection performance. The study utilizes a dataset of compromised user credentials from the Los Alamos National Laboratory (LANL) network, collected during a red-teaming exercise. The results indicate that HLSF outperforms both CP-APR and normalizing flows when used independently, showcasing its ability to capture both the structure and distribution of cyber behavior effectively. The paper highlights the significance of leveraging hybrid approaches in enhancing anomaly detection capabilities in complex cyber environments.
Methodology
The methodology involves using CP-APR for tensor decomposition to derive structural anomaly scores and normalizing flows to model the density of latent behavior. The HLSF framework combines these scores through a weighted sum to enhance anomaly detection performance.
Results
The experimental results show that the HLSF framework significantly improves anomaly detection performance compared to using CP-APR or normalizing flows alone. The hybrid approach effectively captures complex patterns in the data, leading to better identification of compromised user credentials.
Implications
The findings suggest that hybrid machine learning approaches can enhance the effectiveness of cybersecurity systems in detecting subtle anomalies, potentially leading to more robust defenses against cyber threats. This methodology could be applied to various cybersecurity scenarios, including intrusion detection and monitoring of user behavior.
TurboVec: A Case Study in Cost-Efficient Private Retrieval for Enterprise RAG via Codebook-Oblivious Quantization
NLP
Large Language Models
Efficient ML
- TurboVec addresses codebook leakage and recall degradation in multi-tenant RAG systems.
- TurboQuant outperforms trained FAISS Product Quantization in recall while using less memory.
- The implementation achieves significantly lower query latency compared to traditional methods.
- Kernel-level filtering maintains high recall rates for tenant-isolated retrieval.
Read more
TurboVec: A Case Study in Cost-Efficient Private Retrieval for Enterprise RAG via Codebook-Oblivious Quantization
Summary
The paper presents TurboVec, an open-source vector index designed to address challenges in Retrieval-Augmented Generation (RAG) systems, particularly in multi-tenant enterprise environments. It highlights two main issues: the potential leakage of corpus statistics during index construction due to trained codebook quantizers and the degradation of recall in tenant-isolated retrieval due to post-hoc filtering. TurboVec utilizes TurboQuant, a codebook-oblivious scalar quantizer that analytically derives quantization boundaries without requiring corpus-dependent training. The authors demonstrate that TurboQuant outperforms traditional trained quantization methods, achieving higher recall rates while using significantly less memory. In a case study, TurboVec shows a median query latency of 11ms for 100K vectors, compared to 707ms for brute-force scans. Additionally, the paper evaluates the effectiveness of kernel-level filtering for tenant isolation, achieving high recall rates across varying workloads. The authors also assess the privacy implications of their design, finding that TurboVec's approach significantly reduces the risk of membership inference attacks compared to traditional codebook methods. However, they acknowledge limitations, including the scope of evaluation on a single dataset and the need for further exploration of privacy concerns.
Methodology
The methodology involves the use of TurboQuant, which employs an analytically derived scalar quantization process that does not rely on training data. The process includes normalization, random rotation, and optimal boundary computation based on known distributional properties. The paper also implements kernel-level filtering for efficient tenant isolation and evaluates the privacy of the system against membership inference attacks.
Results
TurboQuant consistently outperformed trained FAISS Product Quantization by 8.5–8.9 percentage points in Recall@5 across various scales. It achieved a median query latency of 11ms for 100K vectors, significantly faster than the 707ms for brute-force scans. The kernel-level filtering maintained a Recall@10 of 0.86–0.93 across different tenant workloads, compared to 0.09–0.19 for a baseline method. The privacy evaluation indicated a reduction in membership inference attack accuracy to near-random levels.
Implications
The findings suggest that TurboVec can enhance the efficiency and privacy of RAG systems in enterprise applications, making it a viable option for organizations looking to deploy multi-tenant architectures without compromising on performance or security. The approach could be applied to various domains where retrieval systems are critical, such as customer service, knowledge management, and data retrieval in large-scale applications.
Building a Neural Network from Scratch: Implementation, Evaluation, and Optimization
Theory
Optimization
Efficient ML
- The framework is implemented from scratch, enhancing understanding of neural network mechanics.
- It includes diverse modules such as multi-layer architectures, activation functions, and optimization techniques.
- The framework successfully validates its performance on a multi-class classification task.
- It serves as both an educational tool and a baseline for future research.
Read more
Building a Neural Network from Scratch: Implementation, Evaluation, and Optimization
Summary
This paper addresses the gap between practical usage and fundamental understanding of neural networks by presenting a self-contained framework implemented entirely from scratch. The framework includes essential components such as multi-layer architectures, various activation functions, regularization techniques, and state-of-the-art optimizers, all without relying on high-level deep learning libraries. The primary objectives are to construct a neural network architecture based on fundamental principles and to apply it to a multi-class classification problem. The implementation serves as a pedagogical tool, providing insights into forward and backward propagation, gradient dynamics, and optimization landscapes. The framework demonstrates robust performance across different configurations, validating its correctness, numerical stability, and generalization capabilities. The clean modularity of the design positions it as a reliable baseline for educational purposes and future research, with the relevant code publicly available on GitHub.
Methodology
The methodology involves constructing a neural network architecture from first principles, implementing various modules including multiple hidden layers, Kaiming initialization, weight decay, batch normalization, and dropout. Each module is described with its mathematical formulation and functional rationale, followed by systematic training and empirical analysis to identify optimal configurations.
Results
The framework demonstrates robust performance in a multi-class classification task, validating its correctness and numerical stability. The empirical analysis reveals optimal architectures and configurations, showcasing the effectiveness of the implemented techniques.
Implications
This work has significant implications for education in deep learning, providing a deeper understanding of neural network operations. It can also serve as a foundation for researchers looking to explore novel architectures or diagnose training issues without relying on high-level abstractions.
Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer
Time Series
Optimization
Efficient ML
- Introduction of RoSIP-Batt framework for joint SOH and RUL prediction.
- Dynamic loss balancing through a Bayesian homoscedastic uncertainty weighting mechanism.
- Utilization of Rotary Position Embedding for capturing degradation patterns.
- Significant performance improvements over state-of-the-art methods on multiple datasets.
Read more
Dynamic Loss Balancing for Joint SOH and RUL Prediction of Lithium-Ion Batteries via a Rotary SOH-Injected Prior Battery Transformer
Summary
The paper addresses the challenge of joint prognosis of State of Health (SOH) and Remaining Useful Life (RUL) in lithium-ion batteries, which is hindered by task heteroscedasticity. Traditional multi-task learning frameworks struggle to balance the low-variance noise of SOH estimation with the high-variance uncertainty of RUL predictions. The authors propose the Rotary SOH-Injected Prior Battery Transformer (RoSIP-Batt), a unified co-estimation framework that formulates the joint prediction as a Bayesian multi-task objective. This framework introduces a homoscedastic uncertainty weighting mechanism that dynamically adjusts task-specific gradients based on learned residual noise levels. Key architectural innovations include decoupled dual classification tokens, a per-dimension gated fusion mechanism, and a gradient-detachment operator to protect the SOH representation from high-variance RUL updates. Additionally, Rotary Position Embedding (RoPE) is utilized to model translation-invariant relative temporal profiles. The intermediate SOH estimate is directly injected into the RUL regression head as a physical degradation prior. Evaluations on multiple datasets demonstrate that RoSIP-Batt significantly outperforms existing methods, achieving a mean absolute error (MAE) of 1.994% for SOH on the NASA dataset and a RUL prediction error of 62.85 cycles on the Stanford dataset, showcasing its generalizability and efficiency for real-time battery management system deployment.
Methodology
The RoSIP-Batt framework employs a modular, physics-constrained multi-task learning architecture. It processes multi-channel charging profiles using a one-dimensional convolutional frontend, followed by a shared Transformer with Rotary Position Embedding to capture long-term degradation dependencies. The architecture includes dual classification tokens for SOH and RUL, a gated fusion layer for dynamic representation combination, and a gradient-detachment operator to shield SOH features from RUL gradient noise.
Results
The RoSIP-Batt framework achieved a mean absolute error of 1.994% for SOH estimation on the NASA dataset and a RUL prediction error of 62.85 cycles on the Stanford dataset, outperforming existing methods and demonstrating robust performance across different datasets.
Implications
The proposed framework offers a highly generalizable and computationally efficient solution for real-time battery management systems, enhancing the reliability of lithium-ion battery health monitoring and prognosis, which is critical for the electrification of transportation and energy storage systems.
FALCON-Discover: Discovering Concentrated False-Confidence Regions for Calibration
Theory
Interpretability
- Introduces false-confidence concentration as a critical aspect of calibration analysis.
- Presents FALCON-Discover, a post-hoc framework for discovering dangerous confident errors.
- Demonstrates that different ranking strategies are effective in different regimes.
- Establishes the importance of localizing high-risk error regions in predictive models.
Read more
FALCON-Discover: Discovering Concentrated False-Confidence Regions for Calibration
Summary
This paper addresses the limitations of traditional calibration methods in machine learning, which often evaluate confidence at an aggregate level. The authors introduce the concept of 'false-confidence concentration,' focusing on the identification of regions in the prediction space where high-confidence errors occur. They propose FALCON-Discover, a model-agnostic framework that ranks predictions based on various signals, including confidence, local support, neighborhood agreement, and perturbation stability. The framework aims to discover and localize dangerous errors that are often overlooked in aggregate calibration metrics. The empirical analysis across seven binary tabular datasets reveals that false-confidence concentration is a regime-dependent phenomenon, with different ranking strategies performing better under varying conditions. The findings suggest that a sample-level view of calibration can complement traditional aggregate methods, providing insights into where confident predictions may be unreliable.
Methodology
The authors developed FALCON-Discover, a post-hoc discrepancy-discovery framework that utilizes signals from confidence, local support, neighborhood agreement, and perturbation stability to rank predictions. The framework is model-agnostic and does not modify the underlying classifier. The evaluation involved seven heterogeneous binary tabular datasets, employing four seeds and five-fold cross-fitting to assess performance against validation-selected calibration and trust-scoring baselines.
Results
The results indicate that false-confidence concentration is a recurring phenomenon that varies by regime. The discrepancy family outperformed traditional validation-selected baselines in identifying dangerous errors, with learned discrepancy ranking being most effective when multiple cues are combined, while stability-centered ranking excelled in cases of local decisional fragility. The study also found that the strongest detector varies depending on the regime, highlighting the complexity of the problem.
Implications
The findings suggest that calibration strategies should explicitly target regions where confidence, support, and stability diverge. This approach can enhance the reliability of predictive models, particularly in high-stakes applications where trusting incorrect predictions can lead to significant consequences.
Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes
Graph Learning
Optimization
- Introduction of GNNAS-TSP, a GNN-based framework for TSP algorithm selection.
- Elimination of manual feature engineering by learning directly from raw graph data.
- Evaluation of various cost-based and rank-based loss functions for algorithm selection.
- Significant performance improvements over the Single Best Solver under fixed computational budgets.
Read more
Graph Neural Network-based Algorithm Selection for the Traveling Salesman Problem: A Systematic Study of Cost and Rank Losses under Distinct Budget Regimes
Summary
This paper presents GNNAS-TSP, a Graph Neural Network (GNN)-based framework for Automated Algorithm Selection (AS) specifically tailored for the Traveling Salesman Problem (TSP). The authors argue that traditional AS methods often rely on manually engineered features, which can be limiting. GNNAS-TSP learns instance representations directly from raw graph data, thereby eliminating the need for manual feature engineering. The framework formulates AS as a joint cost-prediction and ranking task, evaluating various cost-based and rank-based loss functions across a portfolio of TSP solvers. The experiments conducted under fixed computational budgets of 10 and 60 seconds demonstrate that GNNAS-TSP significantly outperforms the Single Best Solver (SBS) in terms of normalized solution cost. The results indicate that GNNAS-TSP effectively leverages the variation in solver performance across different instances, making it a promising meta-solving strategy for TSP.
Methodology
The GNNAS-TSP framework utilizes Graph Neural Networks to learn representations of TSP instances from raw graph data. It formulates the algorithm selection task as a joint cost-prediction and ranking problem, employing various loss functions including mean squared error, mean absolute error, and rank-based methods like RankNet and ListNet. The framework is evaluated using a portfolio of TSP solvers under two distinct computational budgets.
Results
The experiments reveal that GNNAS-TSP achieves substantial and statistically significant improvements in normalized solution cost compared to the Single Best Solver across both 10-second and 60-second computational budgets. The results are based on evaluations of 300 TSP test instances, showcasing the framework's ability to select the most suitable algorithm for varying problem instances effectively.
Implications
The findings suggest that GNNAS-TSP can serve as a valuable tool for automated algorithm selection in combinatorial optimization problems, particularly in scenarios where solver performance varies significantly across different instances. This could lead to more efficient problem-solving strategies in various applications of the Traveling Salesman Problem and similar optimization challenges.
Selectivity Matters: Source Node Influence Pruning for Unsupervised Graph Domain Adaptation
Graph Learning
- SNIP shifts the focus from feature-level alignment to data-level refinement in UGDA.
- The method quantifies structural discrepancies using centrality measures to assign influence scores to source nodes.
- A rank-based normalization mechanism is introduced to standardize influence scores across different centrality measures.
- SNIP effectively filters out structurally incompatible nodes, resulting in a refined sub-source graph.
Read more
Selectivity Matters: Source Node Influence Pruning for Unsupervised Graph Domain Adaptation
Summary
This paper addresses the challenge of Unsupervised Graph Domain Adaptation (UGDA), which seeks to transfer knowledge from a labeled source graph to an unlabeled target graph while mitigating distribution shifts. Traditional methods focus on feature-level alignment, assuming all source nodes contribute positively to this process. However, this assumption is flawed as nodes with significant structural deviations can introduce noise and lead to negative transfer. The authors propose a novel approach called Source Node Influence Pruning (SNIP), which emphasizes selective utilization of source nodes based on their structural compatibility with the target domain. SNIP quantifies the structural discrepancies of individual source nodes using various centrality measures, assigning influence scores to each node. A rank-based normalization mechanism is employed to standardize these scores, allowing for effective filtering of structurally incompatible nodes. The refined 'sub-source' graph constructed through SNIP enhances the alignment process, leading to improved adaptation performance. Experimental results across multiple datasets demonstrate that SNIP consistently outperforms existing methods, validating the importance of selective node utilization in UGDA.
Methodology
The methodology involves quantifying the structural discrepancies of source nodes relative to the target domain using multiple centrality measures (e.g., degree and betweenness centrality). A rank-based normalization process is applied to these measures to create a uniform scale for influence scores. Nodes with low influence scores, indicating structural incompatibility, are pruned to create a refined sub-source graph that enhances subsequent alignment efforts.
Results
The experiments conducted across eight transfer scenarios on five real-world datasets demonstrate that SNIP consistently outperforms competitive baselines, significantly improving adaptation performance and validating the effectiveness of selective source node utilization.
Implications
The findings suggest that selective node utilization can be a critical factor in improving the performance of graph-based models in domain adaptation tasks. This approach can be applied in various fields where graph-structured data is prevalent, such as social network analysis, bioinformatics, and recommendation systems.
Neural Kolmogorov Equations: Parallelizable Learning of Stochastic Dynamics under General Noise
Time Series
Generative Models
Theory
- NKEs provide a deterministic reformulation of Neural SDEs, focusing on probability density evolution.
- The framework allows for noise-agnostic learning, accommodating diverse stochastic drivers.
- NKEs enable parallel-in-time training, improving computational efficiency and scalability.
- The method accurately models both deterministic and stochastic dynamics with competitive predictive accuracy.
Read more
Neural Kolmogorov Equations: Parallelizable Learning of Stochastic Dynamics under General Noise
Summary
This paper introduces Neural Kolmogorov Equations (NKEs), a novel framework for learning stochastic dynamics that overcomes limitations of existing Neural Stochastic Differential Equations (SDEs). Traditional Neural SDEs often assume uncoupled and continuous noise, which restricts their applicability to real-world scenarios characterized by complex stochastic drivers. NKEs reformulate the learning problem by focusing on the evolution of probability densities through the Kolmogorov Forward Equation (KFE), allowing for a more flexible and accurate representation of stochastic processes. This framework enables the modeling of general Lévy-type stochastic forcing and facilitates parallel-in-time training using a Lagrangian Galerkin projection and operator splitting techniques. The authors validate NKEs on various stochastic benchmarks, demonstrating their ability to accurately recover both deterministic and stochastic dynamics while improving training efficiency compared to existing methods. The proposed approach not only enhances scalability but also allows for the learning of noise structures without predefined assumptions, making it a significant advancement in the field of stochastic modeling.
Methodology
The authors leverage the Kolmogorov Forward Equation to reformulate the learning of stochastic dynamics, employing neural networks to model the advection and diffusion dynamics. They introduce a novel training scheme based on a deterministic representation of the KFE, which avoids the computationally expensive autoregressive simulation typically required in Neural SDE training. The approach utilizes operator splitting techniques to facilitate parallelization and improve accuracy.
Results
NKEs were evaluated on multiple stochastic benchmarks, including systems with coupled noise and jump processes. The results indicated that NKEs achieved high predictive accuracy while significantly enhancing training efficiency compared to traditional Neural SDEs and other simulation-free frameworks.
Implications
The proposed NKE framework has the potential to advance the modeling of complex stochastic systems in various domains, including finance, engineering, and scientific research. Its ability to learn diverse noise structures and improve computational efficiency could lead to broader applications in real-time data-driven modeling and simulation.
Cost Accounting for Reactive Computational Graphs: Exhaustive Sweeps, Sequential Mutation, and the Backward-Locality Gap
Theory
Efficient ML
Interpretability
- The speedup of exhaustive sweeps is not constant and depends on the depth profile of the network.
- Sequential mutations incur an additional cost due to overcounting, while batched applications are order-independent and sub-additive.
- The backward-locality gap restricts speedup during backpropagation in architectures lacking long skip connections.
- Empirical results validate theoretical predictions across different cost profiles and mutation scenarios.
Read more
Cost Accounting for Reactive Computational Graphs: Exhaustive Sweeps, Sequential Mutation, and the Backward-Locality Gap
Summary
This paper addresses the computational costs associated with exhaustive interventions on neural networks' computational graphs, specifically focusing on activation-patching sweeps, circuit-discovery searches, and systematic ablation studies. The author presents a comprehensive cost accounting framework for reactive graph engines, which efficiently manage invalidation of downstream nodes during mutations. Key findings include the variable speedup of exhaustive sweeps over independent recomputations, which depends on the computational weight distribution across layers, and the exact cost of sequential mutations, which reveals an overcounting phenomenon. The paper also establishes a backward-locality gap that limits speedup during backpropagation in certain architectures. Empirical validation using the NeuroDSL engine confirms the theoretical predictions, demonstrating that the proposed models accurately reflect real-world performance in various scenarios.
Methodology
The paper employs a theoretical framework to analyze the costs associated with various mutation strategies in reactive computational graphs. It derives closed-form expressions for the costs of exhaustive sweeps and sequential mutations, and validates these findings through empirical measurements using the NeuroDSL reactive graph engine.
Results
The study finds that the speedup ratio for exhaustive sweeps converges to (q + 2)/(q + 1) or q + 2 based on the Karamata index of the network's cost profile. The empirical results show that measured sweep ratios align with theoretical predictions, with training-mode ratios collapsing to 1 under specific conditions. The paper also confirms that all sequential costs and batched totals match the derived closed forms exactly.
Implications
The findings have significant implications for optimizing the efficiency of neural network interventions, particularly in mechanistic interpretability tasks. Understanding the cost dynamics of mutations can lead to more efficient designs of reactive graph engines and improve the performance of various machine learning applications.
HyBDM: Multi-Scale Hybrid Experts for Time Series Forecasting with Bidirectional Dependency Modeling
Time Series
- HyBDM introduces a multi-scale hybrid framework with specialized experts for long- and short-range dependency modeling.
- The Global Patterns Expert employs BiConv-Mamba with bidirectional convolutions and a forgetting mechanism for enhanced temporal modeling.
- The Local Variations Expert uses a Local Window Transformer for efficient locality-aware attention.
- HyBDM achieves superior accuracy and efficiency compared to state-of-the-art forecasting models.
Read more
HyBDM: Multi-Scale Hybrid Experts for Time Series Forecasting with Bidirectional Dependency Modeling
Summary
The paper presents HyBDM, a novel multi-scale hybrid model designed for time series forecasting (TSF) that effectively captures both long-range global patterns and short-range local variations. Traditional models often struggle to balance these two types of dependencies, leading to suboptimal forecasting performance. HyBDM addresses this issue by employing two specialized experts: the Global Patterns Expert, which utilizes an enhanced BiConv-Mamba module for capturing long-range trends, and the Local Variations Expert, which implements a Local Window Transformer (LWT) for short-term fluctuations. The model also features a Multi-Scale Patcher and Long-Short Router that facilitate adaptive fusion of the two experts based on input characteristics. Extensive experiments on six benchmark datasets demonstrate that HyBDM outperforms existing state-of-the-art methods in terms of both accuracy and computational efficiency, showcasing its capability to bridge global and local dependencies in multivariate time series forecasting.
Methodology
HyBDM consists of two main components: the Global Patterns Expert, which uses a BiConv-Mamba module with bidirectional convolutions and an M-SSM layer to capture long-range dependencies, and the Local Variations Expert, which employs a Local Window Transformer for short-range variations. A Long-Short Router dynamically fuses the outputs of both experts based on the characteristics of the input data, allowing for adaptive representation at multiple scales.
Results
The experimental results indicate that HyBDM significantly outperforms several state-of-the-art time series forecasting models across six benchmark datasets, demonstrating both improved accuracy and computational efficiency. The model's design allows it to effectively capture and model the intertwined dependencies of global patterns and local variations.
Implications
HyBDM has potential applications in various domains requiring accurate time series forecasting, such as finance, healthcare, and environmental science. Its ability to efficiently model complex dependencies can enhance decision-making processes in real-time applications.
Fully-sensorized smart-eyewear platform for on-device Machine Learning
Computer Vision
Efficient ML
Multimodal
- ARGO leverages on-device machine learning for real-time processing, enhancing user privacy.
- Introduces Head-wise Parallel Attention (HPA) for efficient model execution on the NPU.
- Achieves a mean Average Precision (mAP50–95) of 24% with a memory footprint of 2.483 MB.
- Integrates a multimodal sensor suite for comprehensive environmental monitoring.
Read more
Fully-sensorized smart-eyewear platform for on-device Machine Learning
Summary
This paper introduces ARGO, a smart eyewear platform that integrates ergonomic design with high computational performance and energy efficiency. Unlike traditional cloud-based solutions, ARGO utilizes the STM32N6 microcontroller equipped with a Neural Processing Unit (NPU) to facilitate on-device machine learning, thereby reducing latency and enhancing user privacy through local data processing. The main contribution of the work is the comprehensive co-design of hardware, firmware, and AI, particularly through the deployment of an optimized YOLOv11 model for real-time urban obstacle recognition. To adapt the model for the NPU, the authors propose a novel Head-wise Parallel Attention (HPA) architecture, which allows efficient execution while maintaining the computational integrity of the original model. The YOLOv11 model is trained on the Walking On The Road (WOTR) dataset, achieving a mean Average Precision (mAP50–95) of 24% under stringent memory constraints, with a compact memory footprint of 2.483 MB. The ARGO platform is equipped with a multimodal sensor suite, including RGB cameras, Time-of-Flight sensors, microphones, and ambient sensors, enabling it to operate at 10 frames per second (FPS) with a continuous autonomy of 113 minutes on a 200 mAh battery. The findings underscore the potential of ARGO as a high-performance, privacy-preserving assistive device, emphasizing the need for integrated, multidisciplinary approaches in developing competitive AI solutions.
Methodology
The authors designed the ARGO platform by co-developing hardware and firmware tailored for the STM32N6 microcontroller and its NPU. They implemented the YOLOv11 model with the HPA architecture for efficient processing. The model was trained on the WOTR dataset, and the system was tested for performance metrics including mAP and battery life.
Results
The ARGO platform achieved a mean Average Precision (mAP50–95) of 24% while maintaining a memory footprint of only 2.483 MB. It operates at 10 FPS and provides continuous usage for 113 minutes on a 200 mAh battery, demonstrating its effectiveness as a smart eyewear solution.
Implications
The ARGO platform has significant implications for the development of assistive technologies, particularly in urban environments. Its design prioritizes user comfort and privacy, making it suitable for widespread adoption in healthcare and safety applications. Additionally, it highlights the importance of integrated design in advancing wearable technology.
Exposure-Based Reinforcement Learning to Rank
Reinforcement Learning
Optimization
Efficient ML
- Introduces an exposure-based RL approach that avoids custom gradient computations.
- Achieves higher sample efficiency and faster convergence compared to existing methods.
- Integrates seamlessly with auto-differentiation, simplifying implementation for practitioners.
- Demonstrates no additional computational costs when using GPUs.
Read more
Exposure-Based Reinforcement Learning to Rank
Summary
This paper addresses the challenges of applying reinforcement learning (RL) to learning-to-rank (LTR) tasks, particularly the inefficiencies and complexities associated with existing methods that rely on custom gradient computations. The authors propose a novel approach that emphasizes variance reduction and GPU computation, allowing for high sample efficiency through baseline corrections and partial marginalization. They introduce an abstraction that integrates gradient estimation with a document-exposure distribution, facilitating seamless compatibility with auto-differentiation frameworks. This approach simplifies the implementation process for practitioners, as it only requires the definition of a differentiable loss function based on exposure. Experimental results demonstrate that the proposed exposure-based RL method converges faster and achieves higher ranking performance compared to traditional custom gradient methods, without incurring additional computational costs on GPUs. The findings suggest that this new methodology significantly enhances the effectiveness, efficiency, and accessibility of RL for LTR applications.
Methodology
The authors focus on variance reduction techniques, such as baseline corrections and partial marginalization, to enhance sample efficiency. They propose an estimator based on document-exposure distributions that allows for gradient estimation to be integrated with auto-differentiation, thus simplifying the implementation process for various exposure-based loss functions.
Results
The experimental results indicate that the exposure-based RL method converges significantly faster and achieves better ranking performance than existing custom gradient methods. Additionally, it maintains computational efficiency on GPUs, with no significant increase in computation time.
Implications
This work has the potential to make RL for LTR more accessible to practitioners, enabling broader application of exposure-based ranking metrics in information retrieval and recommendation systems. The improved stability and efficiency could lead to better user experiences in systems that rely on ranking algorithms.
Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting
Time Series
- Introduces a lightweight adaptation framework for TSFMs in drought forecasting.
- Utilizes two wrappers (SMR2 and MBB) to enhance model performance without fine-tuning.
- Achieves up to 26% reduction in MSE for SPEI predictions.
- Addresses practical constraints of proprietary models and limited local data.
Read more
Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting
Summary
This paper addresses the challenges of applying large Time Series Foundation Models (TSFMs) to regional drought forecasting, particularly in resource-constrained environments. Traditional fine-tuning methods are often infeasible due to proprietary model weights, limited local training data, and computational constraints. To overcome these issues, the authors propose a lightweight adaptation framework that enhances frozen TSFMs at inference time using two innovative wrappers: SMR2 (Stationarity aware multi-resolution Residual) and MBB (Moving Block Bootstrap). SMR2 decomposes the input into multi-resolution temporal views, learning specific residual corrections that capture regional dynamics, while MBB preserves temporal dependencies through block resampling. The framework allows for diverse input views and aggregates forecasts without requiring weight updates. Evaluated on one-month-ahead Standardized Precipitation Evapotranspiration Index (SPEI) predictions across multiple sites in South Australia, the proposed method demonstrates significant improvements in forecasting performance, achieving up to a 26% reduction in mean squared error (MSE) compared to the frozen backbone models. This approach enables practical deployment in regional forecasting systems, making it a valuable contribution to climate forecasting methodologies.
Methodology
The authors developed two wrappers: SMR2, which decomposes input data into multi-resolution views and learns stride-specific residual corrections, and MBB, which uses block resampling to maintain temporal dependencies. Both wrappers operate at inference time, allowing for adaptation without access to model weights or extensive retraining.
Results
The proposed framework consistently improved forecasting performance across various backbone models, achieving up to a 26% reduction in mean squared error (MSE) for one-month-ahead SPEI predictions in South Australia.
Implications
This work provides a practical solution for enhancing drought forecasting capabilities using existing TSFMs, making it applicable in regions with limited data and computational resources. It opens avenues for further research in lightweight model adaptation techniques for other climate-related forecasting tasks.
ALAS: Additive Learnable Alpha-Stable Kernels for Flexible Bayesian Optimization
Optimization
Theory
- Introduction of ALAS, a learnable α-stable kernel family for Bayesian Optimization.
- ALAS adapts its smoothness based on data, capturing diverse objective structures.
- ALAS-Sep variant improves robustness in high-dimensional optimization tasks.
- Theoretical connections between spectral tails and information gain are established.
Read more
ALAS: Additive Learnable Alpha-Stable Kernels for Flexible Bayesian Optimization
Summary
The paper introduces ALAS, a novel family of Gaussian Process (GP) kernels designed for Bayesian Optimization (BO) that enhances flexibility by utilizing symmetric α-stable spectral components. The key innovation is the ability to learn the stability parameter α from data, allowing the kernel to adapt its smoothness and effectively capture both smooth trends and sharp irregularities in the objective function. The authors present two variants: ALAS, a single stationary component with joint spectral modulation, and ALAS-Sep, a separable variant that learns dimension-wise tail behavior for improved robustness in high-dimensional settings. Theoretical insights link spectral tail behavior to Mercer eigenvalue decay, providing explicit information-gain and regret bounds. Empirical evaluations demonstrate that ALAS performs robustly across various benchmarks, with ALAS-Sep excelling in higher dimensions and approximately decomposable objectives.
Methodology
The authors develop a Gaussian Process kernel family based on symmetric α-stable spectral components, allowing for the learning of the stability parameter α through GP marginal likelihood. They introduce two parameterizations: ALAS and ALAS-Sep, the latter being a separable construction that sums one-dimensional ALAS components across dimensions. The paper also derives theoretical results linking spectral tail behavior to Mercer eigenvalue decay.
Results
ALAS demonstrates strong and robust performance across a range of synthetic and real-world benchmarks, particularly excelling in low-to-moderate dimensional tasks. ALAS-Sep shows superior effectiveness in high-dimensional settings by allowing each dimension to learn its own tail parameter, thus improving adaptability and performance.
Implications
The proposed ALAS kernel family can significantly enhance the efficiency and effectiveness of Bayesian Optimization in various applications, particularly in fields requiring optimization of expensive black-box functions, such as engineering, finance, and scientific research.
Physics-Guided Masked Multi-Task Network for Edge-Friendly Battery Health Diagnostics from Stochastically Fragmented Charging Profiles
Time Series
Efficient ML
Theory
- Introduction of PG-M2TN, a compact architecture for battery health diagnostics.
- Utilizes a combination of BiLSTM–Attention, Masked Autoencoder, and dual-stream prediction for enhanced SOH estimation.
- Achieves a global RMSE of 0.0781 across multiple datasets, indicating high accuracy.
- Demonstrates robustness to data fragmentation and improved predictive sensitivity near late-life capacity collapse.
Read more
Physics-Guided Masked Multi-Task Network for Edge-Friendly Battery Health Diagnostics from Stochastically Fragmented Charging Profiles
Summary
This paper addresses the critical challenge of accurately estimating the State of Health (SOH) of lithium-ion batteries, which is essential for ensuring operational safety in various applications such as electric vehicles and renewable energy systems. The authors propose a novel architecture called the Physics-Guided Masked Multi-Task Network (PG-M2TN) that integrates physical principles with deep learning techniques to overcome issues related to data incompleteness and the complexities of multi-task learning. The PG-M2TN employs a BiLSTM–Attention backbone for capturing temporal aging dynamics, a Masked Autoencoder (MAE) for reconstructing voltage-current profiles from fragmented data, and a dual-stream prediction head that combines macroscopic SOH estimation with a microscopic indicator, the Voltage Distortion Ratio (VDR). The model is evaluated on five cross-chemistry lithium-ion datasets, demonstrating superior performance with a global RMSE of 0.0781, showcasing its robustness under severe data fragmentation and its ability to enhance predictive sensitivity near critical late-life capacity collapse points. This architecture is particularly suited for real-world Battery Management Systems (BMSs), offering a compact and efficient solution for battery health diagnostics.
Methodology
The PG-M2TN architecture integrates three key mechanisms: a BiLSTM–Attention backbone for temporal dynamics, a Masked Autoencoder for reconstructing incomplete data, and a dual-stream prediction head that combines SOH estimation with VDR. This design allows for effective learning from fragmented charging profiles while mitigating negative transfer in multi-task learning.
Results
The PG-M2TN model achieved a global RMSE of 0.0781 across five cross-chemistry lithium-ion datasets, demonstrating its superior performance in estimating SOH under conditions of data fragmentation. The model also showed enhanced predictive sensitivity near critical late-life capacity collapse points, making it a reliable tool for battery health diagnostics.
Implications
The findings suggest that the PG-M2TN can significantly improve the accuracy and reliability of battery health diagnostics in real-world applications, particularly in edge computing scenarios where data may be incomplete or fragmented. This has implications for the development of more effective Battery Management Systems (BMSs) in electric vehicles and renewable energy systems.
Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs
Large Language Models
NLP
- Token Inoculation allows LLMs to retain hazardous knowledge while controlling its expression using a special token.
- The method achieves a significant reduction in hazardous query accuracy while preserving benign domain performance.
- Refusal selectivity is influenced by the quality of the training signal during fine-tuning.
- The approach demonstrates better safety-utility trade-offs compared to traditional methods like unlearning and refusal training.
Read more
Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMs
Summary
The paper presents a novel approach called Token Inoculation aimed at addressing the challenges of managing dual-use knowledge in large language models (LLMs). Traditional safety interventions either destroy hazardous content or suppress it at the output layer, leading to a trade-off in model performance. The authors argue for a conditioning approach that retains hazardous knowledge while controlling its expression through a special token. During continued pre-training, hazardous content is marked with a token, allowing the model to bind this marker to the underlying semantics. In supervised fine-tuning, the model learns to respond correctly to hazardous queries when the token is present and to refuse them when absent. This method significantly reduces the model's accuracy on hazardous queries while maintaining high performance on benign queries. The authors demonstrate that Token Inoculation achieves a superior safety-utility trade-off compared to existing methods, suggesting that retaining sensitive knowledge under controlled access is more effective than attempting to forget it.
Methodology
The authors introduce Token Inoculation, which involves marking hazardous content with a special token during continued pre-training and teaching the model to respond to hazardous queries based on the presence of this token during supervised fine-tuning. The methodology includes a binding-and-branching approach to condition the model's responses.
Results
Token Inoculation reduces accuracy on hazardous queries from 79% to 18% while retaining 93% of the base model's performance on benign queries. The method outperforms existing safety interventions across various model scales, achieving the best trade-off in safety and utility.
Implications
The findings suggest that conditioning hazardous knowledge rather than erasing it can lead to more effective safety mechanisms in LLMs. This approach could be applied in sensitive domains such as biosecurity and cybersecurity, where dual-use knowledge is prevalent.
After the Euclidean Highway: Hyperbolic Expert AI as the Next Innovation
NLP
Large Language Models
Theory
- HySAT introduces hyperbolic geometry at the loss layer to enhance training stability and performance in domain-expert AI.
- The paper identifies a critical placement law for integrating non-Euclidean structures into existing models without destabilizing training.
- Extensive experiments with six SLMs show significant improvements in training efficiency and model performance.
- The methodology allows for operational deployment of models with hierarchical domain knowledge, addressing limitations of traditional Euclidean approaches.
Read more
After the Euclidean Highway: Hyperbolic Expert AI as the Next Innovation
Summary
This paper introduces a novel approach to training small language models (SLMs) using hyperbolic geometry, termed Hyperbolic Structure-Aware Training (HySAT). The authors argue that traditional Euclidean transformers struggle with hierarchical domain expertise due to their inability to maintain structural integrity at depth. Instead of embedding hyperbolic geometry throughout the network, HySAT proposes to apply curvature only at the loss layer, which stabilizes training and enhances performance. The authors conducted extensive experiments with six SLMs, demonstrating that this method allows for effective training without the instability associated with previous approaches that integrated curvature into trainable weights. The paper also discusses the implications of this method for the future of domain-expert AI, emphasizing the importance of placement in model architecture and training strategy.
Methodology
The authors developed a training framework called HySAT, which applies hyperbolic losses at the loss layer of the model while keeping the rest of the architecture Euclidean. They conducted a series of experiments across six SLMs, utilizing a large corpus of 18 million samples and implementing a four-arm ablation study to isolate the effects of their proposed method. The training process was monitored for stability, with a detailed ledger of incidents recorded.
Results
The implementation of HySAT resulted in stable training across all models, with no NaN occurrences over approximately 317,000 optimizer steps. The models demonstrated effective adaptation to hierarchical domain knowledge, with four models operationally deployed, including one consumer-facing model. The research also provided a comprehensive failure ledger documenting 17 training incidents, contributing to transparency in the methodology.
Implications
The findings suggest that integrating hyperbolic geometry at the loss layer can significantly improve the training of domain-expert AI systems, making them more robust and efficient. This approach could lead to advancements in various applications requiring hierarchical knowledge representation, such as specialized language models in healthcare, law, and other fields.
GEqTrain: A Configuration-Driven Framework for Retargeting Equivariant Graph Neural Networks Across 3D Scientific Tasks
Graph Learning
Generative Models
- GEqTrain provides a modular framework for equivariant graph learning, enhancing reusability across different tasks.
- The framework separates dataset semantics, model composition, and training objectives, allowing for flexible configuration.
- GEqDiff extends the framework to generative modeling, enabling joint treatment of various geometric representations.
- The framework demonstrates competitive accuracy in three distinct scientific applications without requiring separate software ecosystems.
Read more
GEqTrain: A Configuration-Driven Framework for Retargeting Equivariant Graph Neural Networks Across 3D Scientific Tasks
Summary
The paper introduces GEqTrain, a novel configuration-driven framework designed to enhance the reusability of equivariant graph neural networks (GNNs) across various three-dimensional scientific tasks. Traditional implementations of equivariant GNNs are often tightly coupled to specific tasks, limiting their adaptability. GEqTrain addresses this issue by decoupling dataset semantics, model composition, and training objectives, allowing for a more flexible and modular approach. The framework utilizes a declarative configuration system (Hydra) to assemble model stacks, losses, and training workflows, enabling researchers to retarget models to new tasks primarily through configuration changes. The authors validate GEqTrain on three distinct scientific problems: coarse-grained-to-atomistic backmapping of biomolecular systems, prediction of NMR chemical shifts in molecular solids, and equivariant generative modeling. Additionally, they introduce GEqDiff, a generative extension that treats user-defined equivariant fields as first-class targets, demonstrating the capability to jointly transport various geometric representations. The results indicate that GEqTrain can achieve competitive accuracy across diverse tasks while reducing software overhead, thereby promoting reproducibility and extensibility in equivariant modeling.
Methodology
GEqTrain employs a configuration-driven approach using Hydra to assemble model stacks and training workflows. It maps raw dataset fields to internal typed representations and constructs geometry-derived quantities explicitly. The framework emphasizes local equivariant stacks for efficient inference and supports both predictive and generative workflows.
Results
The authors demonstrate that GEqTrain can effectively handle three different scientific tasks with competitive accuracy, showcasing its versatility and the ability to switch between predictive and generative settings without significant software overhead.
Implications
GEqTrain has the potential to streamline the development of equivariant models in computational chemistry and materials science, making it easier for researchers to adapt models to new tasks and improve reproducibility in scientific computing.
An unsupervised clustering analysis of breast cancer data derived from electronic health records enhanced through UMAP dimensionality reduction
Theory
Efficient ML
Interpretability
- The study applies unsupervised clustering to breast cancer data from electronic health records.
- DBSCAN is used for clustering, enhanced by UMAP for dimensionality reduction.
- Three datasets are analyzed, revealing significant clustering patterns among patients.
- Statistical indices confirm the effectiveness of the UMAP-DBSCAN combination.
Read more
An unsupervised clustering analysis of breast cancer data derived from electronic health records enhanced through UMAP dimensionality reduction
Summary
This study investigates the application of unsupervised clustering techniques to breast cancer data sourced from electronic health records (EHRs). The authors utilize the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm to identify patient groups with significant medical features across three independent datasets: Seoul2016, Qingdao2016, and Heidelberg2019. To enhance clustering performance, they incorporate UMAP (Uniform Manifold Approximation and Projection) for dimensionality reduction prior to applying DBSCAN. The effectiveness of this combined approach is evaluated using three statistical indices: DBCV, DCSI, and DISCO. The findings indicate that the integration of UMAP with DBSCAN improves clustering outcomes, facilitating better identification of patient groups and trends that may be clinically relevant. This research underscores the potential of computational methods in extracting insights from EHRs, which could lead to improved understanding and treatment of breast cancer.
Methodology
The authors applied the DBSCAN clustering method to three datasets derived from EHRs of breast cancer patients. They first performed clustering on the original datasets and then on their low-dimensional representations obtained through UMAP. The performance of the clustering was evaluated using three internal clustering metrics: DBCV, DCSI, and DISCO.
Results
The results demonstrated that the combination of UMAP and DBSCAN significantly improved clustering performance compared to using DBSCAN alone on the original datasets. The statistical indices indicated that the patient groups identified through this method were more distinct and medically relevant.
Implications
This research highlights the potential for using advanced computational techniques to analyze EHRs, which can lead to new insights into breast cancer pathology and potentially inform clinical decision-making and personalized treatment strategies.
Causal Discovery on Irregular Time Series
Time Series
- Proposes a time-aware extension of PCMCI+ for irregular time series.
- Aggregates causal influence over predefined temporal windows instead of fixed lags.
- Demonstrates superior performance in recovering causal structures from synthetic irregular event streams.
- Maintains interpretability and scalability of the original PCMCI+ framework.
Read more
Causal Discovery on Irregular Time Series
Summary
This paper addresses the limitations of existing causal discovery methods, particularly PCMCI+, which are designed for regularly sampled time series. The authors propose an extension of PCMCI+ that accommodates irregularly sampled time series, which are common in real-world applications such as healthcare and financial transactions. The new method aggregates causal influences over predefined temporal windows rather than relying on fixed-lag dependencies. The authors evaluate their approach on synthetic irregular event streams with known causal structures, demonstrating that it consistently recovers the underlying causal graph and significantly outperforms the standard PCMCI+ method in scenarios with irregular sampling. This work bridges the gap between traditional causal discovery frameworks and the challenges posed by irregular temporal data, enhancing the applicability of causal discovery techniques in practical settings.
Methodology
The proposed method reformulates the construction of causal triplets in PCMCI+ to accommodate irregular time intervals. Instead of using fixed discrete lags, it pairs observations based on their timestamps, allowing for a more flexible approach to causal inference that aligns with the actual timing of events.
Results
The evaluation on synthetic datasets shows that the proposed method consistently recovers the true causal graph across various signal-to-noise ratios, outperforming the standard PCMCI+ method, which struggles with irregularly sampled data.
Implications
This work has significant implications for fields that rely on causal inference from irregular time series data, such as finance, healthcare, and sensor networks. By improving causal discovery methods, it enhances decision-making processes and risk assessments in these domains.
TRACE: Trajectory-Based Safety Patch Learning for LLM Post-Training Realignment
NLP
Large Language Models
Optimization
- TRACE shifts the focus from online merging to offline safety patch learning.
- The framework simulates harmful tuning trajectories to optimize safety patches.
- TRACE achieves nearly 100% safety while preserving task utility.
- The method effectively disentangles safety recovery from user task updates.
Read more
TRACE: Trajectory-Based Safety Patch Learning for LLM Post-Training Realignment
Summary
The paper introduces TRACE, a novel framework designed to enhance the safety of large language models (LLMs) post fine-tuning, particularly in Fine-Tuning-as-a-Service (FTaaS) environments. Existing methods for safety recovery often face challenges due to the entanglement of task updates and safety patches, leading to difficulties in balancing safety and utility. TRACE addresses this issue by focusing on offline patch learning rather than online merging operations. It simulates harmful tuning trajectories to create progressively corrupted states and optimizes a safety patch that minimizes interference with task-relevant directions while effectively controlling unsafe behaviors. The framework was evaluated across six benchmarks and two models, demonstrating its ability to achieve nearly 100% safety while maintaining comparable utility to undefended fine-tuned models. This approach represents a significant advancement in post-training safety recovery, allowing service providers to restore safety without compromising the utility gained from user-customized tasks.
Methodology
TRACE employs a trajectory-based optimization approach, simulating harmful tuning trajectories to generate corrupted states. It then learns a safety patch that can effectively recover safety while minimizing interference with the utility of the model. This offline learning process allows for a more decisive and disentangled safety patch compared to traditional merging methods.
Results
The evaluation of TRACE across six benchmarks and two models showed that it consistently outperformed existing safety-utility trade-off methods, achieving nearly 100% safety in all settings while maintaining utility levels comparable to those of undefended fine-tuned models.
Implications
TRACE has significant implications for FTaaS platforms, enabling service providers to enhance the safety of LLMs without the need for extensive retraining. This could lead to more robust and reliable LLM applications in various domains, particularly where user-generated data may contain harmful content.
ConceptCF: Concept-based Counterfactuals for the Explainability of Time Series
Time Series
Interpretability
Optimization
- ConceptCF generates counterfactuals based on high-level human-interpretable concepts rather than individual data points.
- The method utilizes time series decomposition to extract meaningful concepts like scale and periodicity.
- A genetic algorithm is employed to optimize concept mutations for counterfactual generation.
- ConceptCF outperforms existing counterfactual generation methods in terms of validity, confidence, proximity, sparsity, and plausibility.
Read more
ConceptCF: Concept-based Counterfactuals for the Explainability of Time Series
Summary
The paper introduces ConceptCF, a novel method for generating counterfactual explanations in time series data by focusing on human-interpretable concepts. In high-stakes domains like healthcare, explainability is crucial to ensure AI models rely on causal relationships rather than spurious correlations. Traditional counterfactual methods often modify individual data points or subsequences, which can obscure interpretability. ConceptCF addresses this by altering meaningful concepts derived from time series decomposition, such as scale and frequency. The method employs a genetic algorithm to optimize these concept mutations, allowing for clear explanations of how changes in concepts affect model predictions. The authors validate ConceptCF against five state-of-the-art approaches, demonstrating superior performance across various metrics, including validity, confidence, proximity, sparsity, and plausibility. This work contributes to the field of explainable AI by formalizing the counterfactual problem for time series and providing a framework for generating interpretable counterfactuals that enhance understanding of model behavior.
Methodology
ConceptCF operates in three stages: first, it extracts concepts from time series data using decomposition techniques; second, it generates counterfactuals by perturbing these concepts rather than the raw signal; and third, it maps the perturbed concepts back to produce the final counterfactuals. The optimization of concept mutations is performed using a genetic algorithm.
Results
The evaluation of ConceptCF against five state-of-the-art counterfactual generation methods shows that it consistently achieves top-tier performance across multiple metrics, including validity, confidence, proximity, sparsity, and plausibility, indicating its effectiveness in generating interpretable counterfactuals.
Implications
ConceptCF has significant implications for the explainability of AI models in critical applications such as healthcare and predictive maintenance, where understanding model decisions is essential. By providing clear and interpretable counterfactuals, it can help stakeholders make informed decisions based on AI predictions.
Breaking Feedback-Blindness: Utility-Augmented Transformer for Sequential Decision Making
Reinforcement Learning
Theory
Interpretability
- Introduces the concept of feedback-blind retrieval, highlighting a structural limitation in existing Transformer models.
- Proposes the Utility-Augmented Transformer (UAT) that integrates feedback into the attention mechanism.
- Theoretically proves that UAT can uniformly approximate feedback-dependent decision maps.
- Demonstrates superior performance of UAT on various benchmarks compared to observation-only models.
Read more
Breaking Feedback-Blindness: Utility-Augmented Transformer for Sequential Decision Making
Summary
This paper addresses the limitations of existing Transformer-based models in sequential decision-making tasks, particularly in non-stationary and partially observable environments. The authors identify a phenomenon termed 'feedback-blind retrieval,' where the attention mechanism fails to distinguish between histories with identical observations but different action-reward outcomes. To overcome this challenge, they propose the Utility-Augmented Transformer (UAT), which incorporates a utility state that modulates the attention mechanism based on action-reward history. The UAT architecture allows for direct conditioning of the attention kernel on feedback during the forward pass, enhancing the model's ability to adapt to changes in the environment. Theoretical proofs demonstrate that UAT generalizes the observation-only Transformer class and can approximate feedback-dependent decision maps. Empirical evaluations across four benchmarks show that UAT consistently outperforms existing models, particularly in noisy conditions requiring rapid adaptation.
Methodology
The authors developed the Utility-Augmented Transformer (UAT) by introducing a utility encoder that maps action-reward pairs to a compact utility state. This state modulates the query, key, and value projections in the attention mechanism, allowing the model to adaptively retrieve context based on feedback during inference. Theoretical analysis was conducted to establish the expressiveness of UAT, and extensive empirical evaluations were performed on four diverse benchmarks.
Results
UAT consistently outperformed observation-only Transformers, test-time adaptation methods, and input-level feedback baselines across four non-stationary benchmarks. The model showed particularly significant improvements in noisier environments, demonstrating its effectiveness in adapting to feedback changes. Boundary analysis indicated performance degradation as feedback quality diminished, and ablation studies confirmed the importance of each component in the UAT architecture.
Implications
The findings suggest that incorporating feedback directly into the attention mechanism can significantly enhance decision-making in dynamic environments. This approach could be applied to various fields such as clinical treatment planning, financial portfolio management, and robotics, where rapid adaptation to changing conditions is crucial.
Incomplete Observations Boost Evolutionary Performance in Ocean Modeling
Generative Models
Optimization
Time Series
- Introduces a generative model that learns ocean dynamics from sparse observations.
- Develops an optimization framework based on the expectation-maximization algorithm.
- Demonstrates improved reconstruction and prediction performance using real-world data.
- Addresses limitations of current models that rely on complete reanalysis datasets.
Read more
Incomplete Observations Boost Evolutionary Performance in Ocean Modeling
Summary
This paper presents a novel generative state-space model designed to enhance ocean modeling by learning directly from sparse and noisy observations, rather than relying on complete reanalysis datasets. The authors propose an optimization framework that integrates physical states and observations, allowing the model to reconstruct high-fidelity ocean fields and predict their evolution over time. The model employs a hidden Markov structure with continuous state spaces, where oceanic physical quantities are treated as hidden states and measurements as observations. Neural networks are utilized for both initial-state and state-transition modules to capture the complexity of ocean dynamics, while a masked Gaussian distribution is used for the emission module. The optimization framework is based on the expectation-maximization (EM) algorithm, which alternates between reconstructing ocean fields and optimizing neural networks. Experimental results demonstrate that the proposed approach significantly improves reconstruction and prediction accuracy using CMIP6 simulation data and satellite data, suggesting a scalable pathway for future Earth system models to learn from real-world observations.
Methodology
The methodology involves a generative state-space model structured as a hidden Markov model with continuous states. The model incorporates neural networks for capturing ocean dynamics and uses a masked Gaussian distribution for observations. An optimization framework based on the EM algorithm is employed to iteratively reconstruct ocean fields and optimize the model parameters.
Results
The experiments conducted using CMIP6 simulation data and FY-3D satellite data show that the proposed model achieves high-fidelity reconstruction of ocean fields and accurate predictions, demonstrating that learning from sparse observations can enhance the representation of ocean-state dynamics.
Implications
This work suggests a significant advancement in ocean modeling by enabling AI models to learn directly from incomplete and noisy observations, potentially leading to more accurate and efficient Earth system models that can operate without the constraints of complete reanalysis datasets.
Federated Lightweight Fine-Tuning
Federated Learning
Efficient ML
- FLITE reduces communication overhead in federated learning by transmitting a small latent vector instead of full model weights.
- The method achieves an 8718× reduction in communication size while maintaining competitive accuracy.
- A low-rank factorization of the projection matrix significantly decreases memory requirements.
- FLITE demonstrates robustness under non-IID data distributions and small client counts.
Read more
Federated Lightweight Fine-Tuning
Summary
The paper presents Federated Lightweight Fine-Tuning (FLITE), a novel approach to federated learning that addresses the significant communication overhead associated with traditional methods like FedAvg. By utilizing mapping networks, FLITE generates a model's weights from a small, trainable latent vector through a frozen affine projection. This allows for the transmission of a compact latent instead of the full model weights, drastically reducing the communication payload. The authors introduce a low-rank, seed-regenerable factorization of the projection matrix, which reduces memory requirements from approximately 80 GB to about 10 MB. Additionally, they propose a delta formulation that learns an additive correction around a shared, centrally-pretrained base model, enabling effective federated fine-tuning. The method achieves a communication reduction of 8718 times, transmitting only 1,280 floats (approximately 5 KB) per client per round while maintaining competitive accuracy levels. The results demonstrate that FLITE matches or exceeds the performance of full-weight FedAvg under various conditions, including strong non-IID data distributions, and shows robustness to quantization, achieving further reductions in communication size without sacrificing accuracy.
Methodology
FLITE employs mapping networks to generate model weights from a small latent vector through a frozen affine projection. The method incorporates a low-rank, seed-regenerable factorization to minimize memory usage and utilizes a delta formulation for federated fine-tuning, allowing for effective model adaptation without the need for full model transmission.
Results
On the CIFAR-100 dataset with ResNet-18+GroupNorm, FLITE communicates only 1,280 floats (≈5 KB) per client per round, achieving an accuracy of 74.67%, which is within 0.5 percentage points of full-weight FedAvg. The method also shows stability across varying client counts and outperforms traditional methods in terms of bandwidth and accuracy trade-offs.
Implications
FLITE's approach can significantly enhance the efficiency of federated learning systems, making them more feasible for applications with limited bandwidth or resource constraints. This could lead to broader adoption of federated learning in real-world scenarios, particularly in privacy-sensitive domains where data cannot be centralized.
Intelligence from Learnable Novelty
Theory
Reinforcement Learning
Optimization
- Introduces 'learnable novelty' as a key concept linking different definitions of intelligence.
- Develops a closed-form estimator for learnable novelty using a differentiable reservoir computer.
- Demonstrates the estimator's effectiveness in complexity classification and unsupervised learning tasks.
- Shows that using learnable novelty as an intrinsic reward improves exploration in reinforcement learning.
Read more
Intelligence from Learnable Novelty
Summary
This paper introduces the concept of 'learnable novelty' as a unifying principle across various fields that define intelligence differently, such as data compression in machine learning and adaptive behavior in agents. The authors argue that both novelty search and the free-energy principle fail because they conflate learnable and unlearnable surprise. They propose a closed-form estimator for learnable novelty using a differentiable reservoir computer, which serves as both a measure and an objective for systems. The estimator successfully classifies complexity, ranking Turing-complete rule 110 highest among cellular automata, and facilitates the development of solitons in neural cellular automata. Additionally, when used as an intrinsic reward in reinforcement learning, it enhances exploration and improves performance in nine out of ten environments without supervision. The findings suggest that complexity generation, abstraction, and exploration can be achieved through the maximization of a single differentiable quantity, providing a common quantitative basis for various projections of intelligence.
Methodology
The authors developed a closed-form estimator for learnable novelty based on a differentiable reservoir computer. They applied this estimator in various contexts, including complexity classification of cellular automata, unsupervised learning with neural cellular automata, and as an intrinsic reward in reinforcement learning environments.
Results
The estimator successfully ranked Turing-complete rule 110 highest among elementary cellular automata and facilitated the emergence of solitons in neural cellular automata. In reinforcement learning, it improved task performance in nine out of ten environments, demonstrating its effectiveness as an intrinsic reward.
Implications
The findings suggest that learnable novelty can serve as a foundational principle for developing intelligent systems across various domains, potentially leading to advancements in unsupervised learning, reinforcement learning, and the understanding of complex systems.
Apeliotes: A Diffusion-Based Modeling Framework for km-scale Multi-Level Atmospheric Fields
Generative Models
Time Series
Efficient ML
- Apeliotes leverages a global weather foundation model and a region-specific generative diffusion model for efficient km-scale weather forecasting.
- The framework produces multi-level atmospheric fields, enhancing the detail and accuracy of weather predictions.
- Apeliotes achieves competitive performance metrics, including less than 3% error in vertical wind profile predictions.
- The model significantly reduces computational time, enabling rapid forecasting compared to traditional methods.
Read more
Apeliotes: A Diffusion-Based Modeling Framework for km-scale Multi-Level Atmospheric Fields
Summary
The paper introduces Apeliotes, a novel framework designed for high-resolution weather forecasting at the kilometer scale. Traditional weather forecasting methods, particularly dynamical downscaling, are computationally intensive and often limited in their applicability across various regions and scenarios. Apeliotes addresses these challenges by integrating a global weather foundation model with a regionally trained generative diffusion model. This combination allows for the generation of multiple weather variables stochastically and the production of high-resolution atmospheric fields that are not directly available from existing global datasets. The framework is trained on nine years of three-hourly reanalysis data, demonstrating its ability to accurately predict vertical wind profiles and other atmospheric variables with impressive performance metrics. The results indicate that Apeliotes can produce forecasts in a fraction of the time required by traditional methods, making it a promising tool for applications in aviation safety, wind energy forecasting, and air quality modeling.
Methodology
Apeliotes combines a pre-trained global weather foundation model with a regionally trained generative diffusion model. This approach allows for the mapping of global weather dynamics to localized km-scale forecasts, utilizing a data-driven learning process to capture complex nonlinear relationships in atmospheric fields.
Results
The evaluation of Apeliotes shows that it predicts vertical wind profiles with less than 3% error, achieving correlations of 0.91 for 10-m wind speed and 0.99 for 2-m temperature. The normalized root mean square error (NRMSE) values are 0.42 for wind speed and 0.17 for temperature, indicating high accuracy and reliability in the forecasts.
Implications
The development of Apeliotes has significant implications for improving local weather forecasting, enhancing public safety, and supporting economic activities that depend on accurate weather data. Its ability to produce high-resolution atmospheric fields can benefit various sectors, including aviation, renewable energy, and environmental management.
BearingNAS: Obtaining In-Sensor Intelligent Fault Diagnosis Systems for Bearings Using a Laptop
Efficient ML
Optimization
Time Series
- Introduction of BearingNAS, a framework for in-sensor fault diagnosis.
- Elimination of reliance on GPUs by conducting searches on a laptop CPU.
- Achieves 99.50% diagnostic accuracy on the CWRU benchmark.
- Optimized for resource-constrained microcontrollers and ISPUs.
Read more
BearingNAS: Obtaining In-Sensor Intelligent Fault Diagnosis Systems for Bearings Using a Laptop
Summary
The paper presents BearingNAS, a Hardware-Aware Neural Architecture Search (HW-NAS) framework aimed at enabling in-sensor intelligent fault diagnosis systems for bearings. The framework is designed to operate on resource-constrained devices, specifically targeting microcontrollers and Intelligent Sensor Processing Units (ISPUs) with limited memory and processing capabilities. By framing the architecture search as a constrained optimization problem, BearingNAS eliminates the need for expensive GPU resources, allowing the search to be conducted entirely on a laptop CPU within an hour. The framework utilizes a lightweight, derivative-free search strategy and a single data-flow search space that incorporates a decaying kernel growth formulation to manage parameter complexity. Evaluations on the Case Western Reserve University (CWRU) bearing benchmark demonstrate that the best architecture derived from BearingNAS achieves a diagnostic accuracy of 99.50% on the ISPU. This work highlights the potential for deploying machine learning workloads directly within sensor packages, facilitating low-cost, scalable solutions for bearing fault diagnosis across various industrial applications.
Methodology
BearingNAS employs a Hardware-Aware Neural Architecture Search (HW-NAS) approach, treating the architecture optimization as a constrained problem. It utilizes a lightweight search strategy that is derivative-free and operates within a single data-flow search space, incorporating a decaying kernel growth formulation to control parameter growth. The framework is designed to run on laptops without discrete GPUs, targeting microcontrollers and ISPUs with limited RAM and Flash memory.
Results
The best architecture identified by BearingNAS achieved a diagnostic accuracy of 99.50% on the CWRU bearing benchmark, demonstrating its effectiveness in generating high-performance models suitable for in-sensor deployment.
Implications
The findings suggest that it is feasible to integrate advanced machine learning capabilities directly into sensor systems, enabling cost-effective and scalable fault diagnosis solutions in various industrial applications, thereby enhancing operational reliability and reducing maintenance costs.
Robust Multi-View Classification under Noisy Supervision via Global Anchor Consensus
Multimodal
- GALA is the first method to audit noisy labels in multi-view classification using global class anchors and cross-view consensus.
- An adaptive label correction strategy combines soft sample weighting with a conservative rewriting rule.
- GALA shows superior performance, ranking first in 33 out of 36 experimental settings.
- The method improves label quality and representation learning iteratively during training.
Read more
Robust Multi-View Classification under Noisy Supervision via Global Anchor Consensus
Summary
This paper addresses the challenge of multi-view classification in the presence of noisy labels, which are common due to imperfect annotations. The authors propose a novel method called Global Anchor-based Label Auditing (GALA) that constructs a global anchor for each class in every view. This anchor aggregates samples from the entire class, providing a stable reference that is less sensitive to individual predictions. The method evaluates how close an instance is to the anchor of its observed label compared to the nearest competing anchor, resulting in a cross-view audit score. Suspicious samples are assigned lower weights, and labels are corrected only when the anchor-based candidate aligns with the classifier's prediction. This iterative process refines both the labels and the anchors, leading to improved noise-robust representation learning. Extensive experiments on six datasets demonstrate that GALA significantly outperforms eight state-of-the-art methods, particularly under high noise conditions.
Methodology
GALA constructs global anchors for each class in every view, measuring the proximity of instances to these anchors. It computes a cross-view audit score by averaging per-view evaluations and classifier confidence. Suspicious samples receive reduced weights, and labels are corrected based on agreement with anchor-based candidates, refining both labels and anchors throughout training.
Results
GALA outperformed eight state-of-the-art methods across six datasets, achieving an average improvement of 7.34 percentage points under 50% noise. It consistently ranked first in the majority of experimental settings, demonstrating its effectiveness in handling noisy labels in multi-view classification.
Implications
The proposed method has significant implications for applications involving multi-view data, such as image classification, social media analysis, and any domain where data is represented through multiple heterogeneous views. GALA's ability to mitigate the effects of noisy labels can enhance the reliability and accuracy of machine learning models in real-world scenarios.
Visual Semantic Decoding of Electrocorticography from Video Stimuli using End-to-End Deep Learning
Time Series
Multimodal
Interpretability
- Demonstrates the potential of end-to-end deep learning for visual semantic decoding from ECoG data.
- Identifies the Transformer architecture and high-gamma frequency inputs as optimal for decoding performance.
- Highlights the importance of specific cortical regions in visual processing and decoding.
- Utilizes data augmentation techniques to enhance model performance despite limited training samples.
Read more
Visual Semantic Decoding of Electrocorticography from Video Stimuli using End-to-End Deep Learning
Summary
This study explores the feasibility of visual semantic decoding from electrocorticography (ECoG) data using an end-to-end deep learning framework. The primary objective is to predict visual categories from video stimuli based on time-series neural inputs collected from 17 participants with drug-resistant epilepsy. The research evaluates various deep learning approaches, including few-shot learning, data augmentation, and contrastive representation learning, to address the challenge of limited training samples. The best-performing model employs a Transformer-based encoder with high-gamma frequency inputs (80-150 Hz) and utilizes mixup augmentation. The analysis reveals that the early visual cortex and other specific brain regions significantly contribute to decoding performance. The findings indicate that the end-to-end deep learning approach can effectively decode visual semantics from dynamic stimuli while maintaining interpretability consistent with established neuroscience knowledge.
Methodology
The study employs an end-to-end deep learning framework to decode visual categories from ECoG signals. It evaluates multiple architectures, including Transformer-based models, and applies techniques such as mixup augmentation and frequency-band filtering. The analysis focuses on the spectral, temporal, and cortical dimensions of the neural data to understand the model's performance.
Results
The research finds that the Transformer-based model, using high-gamma frequency inputs and a 900 ms post-stimulus window, achieves the best decoding performance. The analysis indicates that specific brain regions, particularly the early visual cortex and the ventral stream, play crucial roles in the decoding process. The model's behavior aligns with established neuroscience knowledge, confirming its interpretability.
Implications
This work has significant implications for brain-computer interfaces and neuroprosthetics, as it demonstrates a reliable method for interpreting complex neural signals in real-time. It also opens avenues for further research into dynamic visual perception and its neural representations, potentially aiding in the development of assistive technologies for individuals with neurological disorders.
Decafs: Disentangled Conditional adversarial Flows
Generative Models
- Introduction of a novel adversarial conditional flow framework that disentangles latent spaces for controlled generation.
- Utilization of a Lie algebra-based generator to learn both disentangled and coupled components of the latent space.
- Empirical validation of DECAFS in generating high-quality conditional outputs in both image and molecular domains.
- Demonstration of improved interpretability and control over generative factors in tasks such as drug discovery.
Read more
Decafs: Disentangled Conditional adversarial Flows
Summary
The paper introduces DECAFS, a novel framework for conditional generation that addresses the challenges of entangled latent spaces in flow-based generative models. Traditional flow models excel in generative tasks but struggle with interpretability due to complex latent embeddings. DECAFS utilizes a conditional generator based on Lie groups to create a disentangled latent space that aligns with the flow latent space through adversarial loss. This approach allows for controlled generation without the need to expand the dimensionality of the flow space, thus avoiding issues like the curse of dimensionality. The authors demonstrate the effectiveness of DECAFS in generating conditional images and molecular structures, outperforming existing models like StyleGAN on datasets such as MNIST and dSprites, and successfully manipulating molecular properties in tasks involving drug design. The paper emphasizes the importance of disentangled representations for better generalization and interpretability in generative modeling.
Methodology
The authors propose a conditional generator based on Lie groups that creates a disentangled latent space aligned with the flow latent space. This is achieved through an adversarial loss that ensures precise control over generated outputs based on user-defined conditions. The model is designed to learn both disentangled and coupled components of the latent space, enhancing flexibility and interpretability.
Results
DECAFS demonstrates superior performance in conditional image generation, outperforming StyleGAN on benchmark datasets like MNIST and dSprites. Additionally, it effectively generates molecular graphs by manipulating distinct properties such as QED, logP, and SA scores, showcasing its versatility across different domains.
Implications
The findings suggest that DECAFS can significantly enhance the interpretability and control of generative models, making it applicable in various fields such as image synthesis, drug design, and other areas requiring precise manipulation of generative factors.
Scalable Causal Imitation Learning
Reinforcement Learning
Robotics
Theory
- Introduction of Causal SQIL and Causal IQ-Learn for continuous control tasks with unobserved confounders.
- Development of an efficient approximation of the sequential π-backdoor criterion for causal adjustment.
- Empirical results show significant performance improvements over prior CIL algorithms in long-horizon tasks.
- Causal SQIL and Causal IQ-Learn can achieve competitive success rates, sometimes exceeding expert performance.
Read more
Scalable Causal Imitation Learning
Summary
This paper addresses the challenges of imitation learning (IL) in environments where the observations of the imitator and the expert are mismatched and where unobserved confounders exist. Traditional causal imitation learning (CIL) methods, such as Causal Behavioral Cloning (Causal BC) and Causal Generative Adversarial Imitation Learning (Causal GAIL), are limited to short-horizon, low-dimensional settings and struggle with continuous control tasks that have long horizons and high-dimensional state-action spaces. The authors propose two new algorithms: Causal Soft Q Imitation Learning (SQIL) and Causal Inverse soft-Q Learning (IQ-Learn). These algorithms integrate causal adjustment frameworks with advanced inverse reinforcement learning objectives, allowing them to operate on causally-adjusted state representations. A significant innovation is the development of an efficient approximation of the sequential π-backdoor criterion, which reduces the computational complexity of causal adjustments from a full horizon to a manageable sliding window. The empirical evaluation demonstrates that Causal SQIL and Causal IQ-Learn outperform existing CIL methods in confounded high-dimensional long-horizon control tasks, achieving competitive success rates and sometimes surpassing expert performance.
Methodology
The paper introduces two off-policy causal imitation learning algorithms, Causal SQIL and Causal IQ-Learn, which leverage state-of-the-art inverse reinforcement learning objectives. These algorithms utilize causally-adjusted state representations derived from an efficient approximation of the sequential π-backdoor criterion, allowing for effective learning in environments with long horizons and high-dimensional state-action spaces.
Results
The evaluation of Causal SQIL and Causal IQ-Learn across various confounded environments reveals that these methods significantly outperform prior CIL algorithms, particularly in long-horizon tasks. In many cases, they achieve success rates comparable to or better than the expert, while traditional causally unaware imitation methods fail to learn meaningful behaviors.
Implications
The findings suggest that Causal SQIL and Causal IQ-Learn can be effectively applied in real-world scenarios where unobserved confounders are present, such as robotics and autonomous systems, enhancing the ability to learn from expert demonstrations in complex environments.
Effects of width-dependent model hyperparameters and ℓ2-regularization on the loss landscape of two-layer ReLU networks
Theory
Optimization
- Derivation of conditions for global minima collapse to zero in two-layer ReLU networks under ℓ2-regularization.
- Demonstration that AdamW optimizer prevents parameter collapse, unlike SGD.
- Analytical solutions for globally optimal parameters in one-dimensional input settings.
- Width-invariant connectivity effects of ℓ2-regularization, with stronger dimensionality reduction as width increases.
Read more
Effects of width-dependent model hyperparameters and ℓ2-regularization on the loss landscape of two-layer ReLU networks
Summary
This paper investigates the theoretical properties of two-layer ReLU networks, particularly focusing on the effects of width-dependent hyperparameters and ℓ2-regularization on the loss landscape. The authors derive a sufficient condition under which the global minima of the network collapse to the zero solution, highlighting the role of the optimizer in this phenomenon. Their experiments demonstrate that using AdamW as an optimizer prevents parameter collapse, unlike SGD. Additionally, the paper provides an analytical solution for the globally optimal parameter sets when the input dimension is restricted to one, revealing that ℓ2-regularization has a width-invariant effect on connectivity but a stronger dimensionality-reducing effect as network width increases. These findings contribute to a deeper understanding of how hyperparameter settings influence the geometry of regularized loss landscapes in neural networks.
Methodology
The authors analyze the loss landscape of two-layer ReLU networks by deriving theoretical conditions for the collapse of global minima under ℓ2-regularization. They conduct numerical experiments to validate their theoretical findings and derive analytical solutions for specific cases, particularly focusing on one-dimensional inputs. The effects of different optimizers on parameter behavior are also examined.
Results
The study finds that when the ℓ2-regularization coefficient scales faster than the network width, the zero solution becomes the unique global minimum for sufficiently overparameterized models. The use of AdamW optimizer is shown to prevent the collapse of learned parameters, while SGD leads to parameter collapse. The analysis also reveals that ℓ2-regularization maintains a width-invariant effect on connectivity but enhances dimensionality reduction as network width increases.
Implications
These findings have significant implications for the design and training of neural networks, particularly in understanding how hyperparameter settings can influence model behavior and performance. The insights gained may inform better practices in selecting optimizers and regularization techniques, ultimately improving training outcomes in deep learning applications.
Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation
Graph Learning
Interpretability
- The study highlights the importance of explainability and auditability in fraud detection systems.
- Graph-derived features and anomaly signals improve detection in specific cases but not overall performance.
- The investigation agent's reliance on model explanations does not consistently lead to better decision-making.
- An escalation rule based on disagreement between the agent and classifier was ineffective in identifying correct decisions.
Read more
Toward Auditable Fraud Detection: Combining Graph Features, Model Explanations, and Agentic Case Investigation
Summary
This paper addresses the challenges of scalable and explainable fraud detection systems in the context of increasing transaction volumes. The author proposes a layered pipeline utilizing the PaySim dataset, which integrates a gradient-boosted classifier, graph-derived structural features, an autoencoder-based anomaly detection signal, TreeSHAP for model explanations, and an investigation agent that employs a bounded large language model (LLM) for uncertain cases. The study first corrects for a simulator-specific balance shortcut that could inflate baseline performance. After this adjustment, it finds that neither graph features nor the anomaly signal improves Average Precision across the full test set, although they do enhance fraud detection within a subset of cases with intermediate baseline scores. In controlled experiments involving multi-account fraud rings, the engineered structural features successfully recover all injected fraudulent transactions, while the tabular baseline misses about 25%. The investigation agent, however, shows lower accuracy than direct thresholding of the classifier, achieving 65.0% accuracy compared to 71.7% on a balanced sample. The agent's decisions often replaced correct classifier outputs with errors, and while it provided coherent rationales, an escalation rule based on disagreement flagged errors without identifying correct decisions. The findings suggest that each component of the fraud detection system has specific conditions under which it contributes effectively, and that a plausible rationale from the investigation agent does not guarantee better decision-making.
Methodology
The methodology involves a layered pipeline combining various machine learning techniques, including a gradient-boosted classifier, graph-derived features, an autoencoder for anomaly detection, TreeSHAP for model explanations, and a bounded LLM investigation agent. The study also includes a correction for a performance inflation shortcut and controlled experiments with injected fraud cases.
Results
The results indicate that while graph features and anomaly signals do not enhance overall Average Precision, they are effective in ranking fraud cases with intermediate scores. The investigation agent performed worse than direct classifier thresholding, with a significant number of incorrect decisions despite providing rationales. The engineered structural features successfully identified all injected fraudulent transactions, contrasting with the tabular baseline's performance.
Implications
The findings suggest that fraud detection systems must be carefully designed to ensure that each component adds value under specific conditions. The study raises questions about the reliability of automated decision-making in fraud detection and the role of human oversight in reviewing flagged cases.
Probabilistic Physics-Aware Machine Learning Predictions of Electric Truck Energy Consumption with Field Data
Theory
Optimization
- Incorporation of physics principles improves energy consumption predictions for electric trucks.
- Bayesian linear regression outperforms standard linear regression in reliability.
- Complex models like neural networks and gradient boosted trees achieve higher accuracy with physics-aware inputs.
- Uncertainty estimation is integrated into the prediction framework, enhancing decision-making under uncertainty.
Read more
Probabilistic Physics-Aware Machine Learning Predictions of Electric Truck Energy Consumption with Field Data
Summary
This paper presents a novel approach to predicting the energy consumption of electric trucks by integrating first principle physics into machine learning models. The authors propose a physics-aware model that accounts for various energy loss sources during vehicle operations. They demonstrate that Bayesian linear regression based on this model enhances the reliability of energy consumption predictions compared to standard linear regression. Additionally, more complex models such as neural networks and gradient boosted regression trees, when informed by the same physical principles, significantly outperform traditional versions of these models in accuracy. The study also introduces a framework for estimating uncertainty in predictions, providing not only point estimates but also a measure of expected deviation, which is crucial for decision-making in logistics and fleet management. The experiments conducted on extensive real-world driving datasets highlight the effectiveness of the proposed methods in improving the accuracy and reliability of energy consumption forecasts for electric trucks.
Methodology
The authors developed a physics-aware model that incorporates first principles of vehicle dynamics and energy loss factors. They applied Bayesian linear regression and more complex machine learning techniques, including neural networks and gradient boosted regression trees, to predict energy consumption. The models were trained on extensive field data, and a framework for uncertainty estimation was also established to quantify prediction reliability.
Results
The results indicate that the physics-aware Bayesian linear regression model provides more reliable energy consumption predictions than standard linear regression. Furthermore, the advanced machine learning models demonstrated significantly improved accuracy in forecasting energy consumption. All models effectively estimated uncertainty, providing a standard deviation for predictions.
Implications
The findings have significant implications for the logistics and transportation sectors, particularly in optimizing route planning, battery management, and charging infrastructure utilization for electric trucks. The ability to quantify uncertainty in energy consumption predictions allows fleet operators to make more informed decisions, reducing risks associated with energy needs during operations.
FlashPDE: A Drop-in Fused Triton Operator Library for Neural PDE Solvers
Efficient ML
- FlashPDE provides a drop-in library of 14 differentiable PDE operators for efficient SciML applications.
- The library reduces memory usage from over 40 GB to 3.4 GB for complex simulations while maintaining performance.
- It achieves significant speedups (up to 2.30×) over traditional PyTorch implementations by minimizing kernel launches.
- FlashPDE supports various neural network architectures without requiring changes to the underlying models.
Read more
FlashPDE: A Drop-in Fused Triton Operator Library for Neural PDE Solvers
Summary
FlashPDE is a novel fused-kernel operator library designed to enhance the efficiency of training Physics-Informed Neural Networks (PINNs) for solving partial differential equations (PDEs). Traditional methods for computing high-order PDE residuals in PINNs often lead to significant memory overhead, particularly in 3D domains, due to the reliance on automatic differentiation and nested computation graphs. FlashPDE addresses these challenges by implementing 14 differentiable PDE operators using custom Triton kernels, which allow for a more efficient evaluation of PDEs without the memory constraints associated with standard PyTorch implementations. The library supports a variety of PDE configurations across 1D to 3D domains and is designed to be architecture-independent, enabling seamless integration with existing neural network models. By utilizing a fused approach that combines forward stencil evaluation, analytic-adjoint backward passes, and boundary gradient corrections, FlashPDE significantly reduces the number of kernel launches and memory usage during training. The results demonstrate that FlashPDE achieves median speedups of up to 2.30× compared to traditional PyTorch finite difference methods while maintaining comparable accuracy metrics. This advancement not only enhances the scalability of neural PDE solvers but also connects high-performance numerical computation with the flexibility of the PyTorch ecosystem.
Methodology
FlashPDE employs a fused-kernel approach using Triton GPU kernels to implement differentiable PDE operators. Each operator integrates forward finite-difference evaluation, analytic discrete adjoint backward passes, and boundary gradient corrections into a single execution layer. This design minimizes kernel launch fragmentation and memory overhead, allowing for efficient training of neural PDE solvers.
Results
The evaluation of FlashPDE across 17 benchmark configurations showed median speedups of up to 2.30× over the eager PyTorch finite difference baseline, with kernel-level acceleration reaching up to 19.2× on larger grids. The memory footprint for complex simulations was reduced from over 40 GB to 3.4 GB, demonstrating substantial improvements in both speed and resource efficiency.
Implications
FlashPDE's advancements could significantly impact the field of scientific machine learning by enabling more efficient training of neural PDE solvers, facilitating the application of PINNs in complex simulations across various domains, and improving the integration of deep learning with high-performance numerical methods.
Attractor Geometry Determines the Identifiability Limits of System Discovery
Theory
- The geometry of the attractor significantly influences the identifiability limits of system discovery.
- λmin(M), the smallest eigenvalue of the invariant-measure moment matrix, serves as a key metric for recovery potential.
- Recovery is hardest in fixed-point regimes, intermediate in limit cycles, and easiest in chaotic regimes.
- Soft F1 is introduced as a new metric for evaluating algorithm performance, capturing partial structural recovery.
Read more
Attractor Geometry Determines the Identifiability Limits of System Discovery
Summary
This paper investigates the limits of identifying governing equations from data in dynamical systems, emphasizing the role of attractor geometry. The authors utilize the Lorenz-84 model to demonstrate that the smallest eigenvalue of the invariant-measure moment matrix, λmin(M), serves as a critical measure of identifiability. This value indicates how well the attractor covers function space, influencing the performance of two prominent algorithms: Sparse Identification of Nonlinear Dynamics (SINDy) and evolutionary symbolic regression (PySR). The study reveals that recovery is most challenging in fixed-point regimes, intermediate in limit cycles, and easiest in chaotic regimes, with λmin(M) providing a quantifiable metric for recovery potential. The paper introduces Soft F1, a new metric that accounts for partial structural recovery, improving the evaluation of algorithm performance. The findings suggest that understanding the attractor's geometry is essential for predicting the success of system discovery methods, and the results are validated on a separate Lorenz-96 system, confirming the robustness of the proposed metrics.
Methodology
The authors employed a within-system design using the Lorenz-84 model, varying a forcing parameter to transition through different dynamical regimes (fixed-point, limit-cycle, chaotic). They analyzed the performance of SINDy and PySR under varying conditions of data volume, measurement noise, and structural prior quality, using λmin(M) as a central measure of identifiability. They also introduced Soft F1 to evaluate algorithm performance more effectively.
Results
The study found a consistent ordering of recovery difficulty across different regimes, with fixed points being the most challenging and chaos the least. The λmin(M) metric was shown to correlate with recovery success, where its positivity indicated the potential for recovery regardless of the algorithm used. The introduction of Soft F1 allowed for a more nuanced understanding of algorithm performance, revealing differences not captured by traditional metrics.
Implications
These findings have significant implications for data-driven science, particularly in fields requiring the discovery of governing equations from observed data. Understanding the limits imposed by attractor geometry can guide researchers in selecting appropriate algorithms and experimental conditions for system discovery tasks.
Relative Positions Generalize, Absolute Positions Memorize: An Implicit-Bias Account of Length Generalization in Attention
NLP
Theory
Optimization
- Relative positional encodings (RoPE) enable transformers to generalize to longer sequences, while absolute positional encodings (APE) do not.
- The authors provide an optimization explanation for this phenomenon, focusing on the implicit bias of attention heads.
- RoPE maintains high accuracy across varying lengths, whereas APE's accuracy drops significantly when encountering unseen lengths.
- The study characterizes the learned rotary rule as a low-rank carrier kernel aligned with the target offset.
Read more
Relative Positions Generalize, Absolute Positions Memorize: An Implicit-Bias Account of Length Generalization in Attention
Summary
This paper investigates the phenomenon of length generalization in transformers, specifically contrasting the performance of relative positional encodings (RoPE) with learned absolute positional encodings (APE). The authors present an optimization-based explanation for why RoPE allows for better extrapolation to longer sequences than APE. They conduct experiments on a minimal fixed-offset retrieval task to isolate the effects of positional selection. The findings reveal that RoPE maintains high accuracy even for lengths beyond those seen during training, while APE's performance collapses to near-chance levels when faced with unseen lengths. The authors prove that RoPE's attention logit is solely a function of relative offsets, ensuring that the learned selection rule is invariant to length. In contrast, APE is constrained to fixed absolute positions, leading to poor generalization. The paper also discusses the implications of these findings for understanding the implicit bias of attention mechanisms and their connection to recurrent models. Overall, the study provides a deeper understanding of how different positional encoding strategies affect the learning dynamics of transformers.
Methodology
The authors conducted experiments using a minimal fixed-offset retrieval task with a one-layer, single-head transformer model. They compared the performance of RoPE and APE by analyzing the attention logits and the selection rules learned during training. The study included empirical validation across multiple seeds and offsets to confirm the robustness of the results.
Results
The results showed that RoPE achieved perfect accuracy for lengths up to 48, while APE's accuracy dropped to near chance levels (0.07) for lengths beyond the training range. The study confirmed that RoPE's attention mechanism is length-invariant and that the learned selection rule aligns with the target offset, leading to graceful accuracy decay. In contrast, APE was found to pin to fixed absolute positions, resulting in poor performance for out-of-range lengths.
Implications
The findings suggest that the choice of positional encoding in transformers significantly impacts their ability to generalize to longer sequences. This has implications for model design in NLP tasks where length variation is common, and it may inform future research on optimizing transformer architectures for better performance.
Edge-Efficient Transformer for End-to-End RF Spectrum Monitoring
Efficient ML
- Introduction of E-SpecFormer, an edge-efficient Transformer for RF spectrum monitoring.
- LiTAN mechanism reduces complexity and improves accuracy without Softmax and LayerNorm.
- Four scalable model variants to accommodate diverse hardware constraints.
- Achieved high accuracy on modulation recognition and covert channel detection tasks.
Read more
Edge-Efficient Transformer for End-to-End RF Spectrum Monitoring
Summary
This paper introduces E-SpecFormer, an innovative edge-efficient Transformer designed for automatic modulation recognition (AMR) and covert channel (CC) detection in RF spectrum monitoring. The authors propose a novel attention mechanism called LiTAN (Linear Tanh Attention Network), which eliminates the need for Softmax and LayerNorm, thereby reducing computational complexity while enhancing accuracy. E-SpecFormer is available in four scalable variants (Nano, Small, Medium, Large) to cater to different hardware capabilities. The model is evaluated using the RadioML2018 dataset for modulation recognition, achieving an impressive 86.5% average accuracy at Signal-to-Noise Ratios (SNRs) greater than 0 dB with the Nano variant, and 94.2% accuracy on a hardware Trojan-based CC dataset, all while maintaining fewer than 10k parameters and achieving a processing speed of 92 µs per frame on FPGA/CPU co-execution. These results demonstrate E-SpecFormer’s superiority over existing edge models, making it a viable solution for real-time spectrum intelligence in IoT devices.
Methodology
The methodology involves the development of E-SpecFormer, which integrates a convolutional tokenizer with the LiTAN attention mechanism. The model processes raw I/Q samples directly for real-time monitoring, avoiding traditional preprocessing methods that introduce latency. The architecture is designed to be hardware-friendly, enabling efficient inference on edge devices without the need for resynthesis.
Results
The Nano variant of E-SpecFormer achieved 86.5% average accuracy for modulation recognition at SNRs > 0 dB and 94.2% accuracy for covert channel detection, all while maintaining fewer than 10k parameters and achieving a processing speed of 92 µs per frame on FPGA/CPU co-execution.
Implications
E-SpecFormer presents a significant advancement in edge computing for RF spectrum monitoring, enabling real-time applications in IoT devices. Its efficient architecture can facilitate enhanced spectrum intelligence, crucial for optimizing communication systems and improving security against covert channels.
ChemFusion: A Multimodal Cross-Attention Network for Reaction Yield Prediction
Multimodal
- ChemFusion integrates electronic features with 3D spatial data using a cross-attention mechanism.
- The model significantly outperforms traditional unimodal frameworks in predicting reaction yields.
- Attention matrices reveal the model's ability to identify and penalize steric clashes, enhancing interpretability.
- Achieved state-of-the-art results on the Buchwald-Hartwig benchmark dataset.
Read more
ChemFusion: A Multimodal Cross-Attention Network for Reaction Yield Prediction
Summary
The paper introduces ChemFusion, a novel hybrid neural network designed to predict the yields of transition-metal-catalyzed reactions by integrating electronic descriptors with 3D atomic coordinates. The challenge of accurately forecasting reaction outcomes is addressed through a cross-attention mechanism that allows the model to dynamically relate global electronic states to specific spatial constraints in molecular structures. By utilizing a pre-trained SchNet as a geometric feature extractor and a Multilayer Perceptron (MLP) for electronic properties, ChemFusion effectively overcomes the limitations of traditional unimodal models that often fail to capture the complexities of steric interactions. The model is benchmarked against a dataset of Buchwald-Hartwig C–N cross-coupling reactions, achieving state-of-the-art predictive performance with an RMSE of 6.22% and an R² of 0.9455. Furthermore, the attention matrices extracted from the model provide insights into how the network identifies and penalizes steric hindrances, enhancing interpretability and aligning with chemical intuition.
Methodology
The ChemFusion architecture employs a multimodal approach that combines a pre-trained SchNet for 3D geometric feature extraction with a Multilayer Perceptron (MLP) for processing tabular electronic descriptors. A Multi-Head Cross-Attention Fusion Layer is introduced to link these modalities without losing spatial resolution, allowing the model to dynamically attend to relevant spatial features during prediction.
Results
ChemFusion achieved an RMSE of 6.22% and an R² of 0.9455 on the Buchwald-Hartwig benchmark dataset, outperforming traditional models. The attention weights extracted post-training demonstrated the model's capability to effectively evaluate steric hindrances and catalytic interactions.
Implications
The findings suggest that ChemFusion could be applied to optimize catalytic reactions in chemical synthesis, providing a more accurate and interpretable framework for predicting reaction outcomes. This approach may also inspire further research into multimodal machine learning applications in molecular sciences.
What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach
NLP
Large Language Models
Interpretability
- Introduces the Bayes-filtered transformer (BFT) concept for understanding latent task beliefs.
- Utilizes predictive Monte Carlo (PMC) as a novel interpretability tool for BFTs.
- Demonstrates that BFTs can approximate their pretraining prior and posterior distributions.
- Applies PMC to various task families, revealing consistent latent phenomena.
Read more
What does a Bayes-filtered transformer believe? A predictive Monte Carlo approach
Summary
This paper introduces the concept of a Bayes-filtered transformer (BFT), a transformer model trained on sequences generated through a two-step process involving latent tasks and conditional observations. The authors explore how to interpret the beliefs internalized by a trained BFT regarding its prior and posterior distributions over latent tasks. They highlight that while the BFT aims to approximate the Bayesian posterior predictive distribution (PPD), it often only achieves an approximation in practice. To address the interpretive challenges, the authors employ a predictive Monte Carlo (PMC) approach, which allows for the direct sampling of the BFT's implicit prior and posterior distributions using next-token generation. The methodology is applied to three stylized task families, demonstrating that the BFT retains significant information about the pretraining prior and updates it according to Bayes' rule. The findings reveal that the PMC method effectively captures the latent beliefs of the BFT, providing insights into its internalized knowledge and decision-making processes.
Methodology
The authors employ predictive Monte Carlo (PMC) to sample from the BFT's implicit prior and posterior distributions. This involves generating sequences through next-token prediction, rolling out from both empty and non-empty prompts to derive empirical distributions that reflect the BFT's internalized beliefs about latent tasks.
Results
The application of PMC to a Beta-Bernoulli BFT shows that the implicit prior and posterior closely track their analytic counterparts, indicating that the model effectively encodes the pretraining prior and updates it according to Bayes' rule. The results confirm the presence of previously reported phenomena in latent space, demonstrating the utility of PMC in interpreting BFTs.
Implications
The findings suggest that PMC can serve as a robust interpretability tool for understanding complex transformer models, potentially aiding in the development of more transparent and reliable AI systems. This approach may also enhance the understanding of model behavior in various applications, including meta-learning and in-context learning.
From Trajectories to Instructions: Language-Conditioned Meta-Reinforcement Learning
Reinforcement Learning
NLP
Robotics
- LA-MAML leverages natural language task descriptions for efficient task adaptation in reinforcement learning.
- The framework replaces traditional trajectory collection and gradient updates with a learned embedding of task instructions.
- Empirical evaluations on the BabyAI benchmark show LA-MAML's improved performance and reduced training time compared to baselines.
- The study includes an ablation analysis to assess the role of language in task adaptation.
Read more
From Trajectories to Instructions: Language-Conditioned Meta-Reinforcement Learning
Summary
This paper introduces LA-MAML (Language Adapted Model-Agnostic Meta-Learning), a novel framework for reinforcement learning (RL) that utilizes natural language task descriptions to enhance the efficiency of meta-learning. Traditional MAML relies on trajectory collection and gradient updates for task adaptation, which can be computationally expensive. The authors propose that instead of this gradient-based adaptation, task-specific signals derived from language instructions can be employed. LA-MAML modifies the inner loop of MAML by using a learned embedding of the task instruction to adapt global policy parameters in a single step. This approach not only simplifies the adaptation process but also significantly reduces the per-iteration training time. The authors validate their method through experiments on the BabyAI benchmark, demonstrating that LA-MAML achieves competitive or superior performance compared to existing baselines while requiring less computational time. The findings suggest that leveraging language instructions can effectively replace trajectory-based adaptation in meta reinforcement learning, leading to faster convergence and improved learning efficiency.
Methodology
The authors propose LA-MAML, which modifies the inner loop of the MAML framework by using a learned language encoder to derive task-specific signals from natural language instructions. This allows for direct adaptation of global policy parameters without the need for trajectory collection or gradient updates, streamlining the adaptation process.
Results
Experiments on the BabyAI benchmark indicate that LA-MAML achieves competitive or improved performance relative to baseline methods, with significantly lower wall-clock training time per iteration. The results highlight the effectiveness of using language instructions for task adaptation in meta reinforcement learning.
Implications
The findings suggest that integrating natural language processing into reinforcement learning can enhance the efficiency of learning algorithms, making them more applicable in real-world scenarios where rapid adaptation is crucial. This approach could be beneficial in various domains, including robotics, autonomous systems, and interactive AI.
Predicting Activities in Aqueous Electrolyte Solutions with Hybrid Machine Learning
Theory
Optimization
Efficient ML
- Introduction of a hybrid model (Bromley-MCM) combining physics-based and machine learning approaches.
- Utilization of matrix completion to predict parameters for unstudied electrolytes.
- Training on a large dataset (478 electrolytes) to enhance predictive accuracy.
- Ability to predict activities for 9,296 electrolytes, significantly broadening the model's applicability.
Read more
Predicting Activities in Aqueous Electrolyte Solutions with Hybrid Machine Learning
Summary
This paper presents a novel hybrid model, termed Bromley-MCM, that integrates the physics-based Bromley model with a machine learning technique known as matrix completion method (MCM) to predict activities in aqueous electrolyte solutions. Activities, represented by ionic activity and osmotic coefficients, are crucial for modeling various industrial and natural processes. Traditional models like Pitzer and Bromley require specific fitting to experimental data for each electrolyte, limiting their applicability to unstudied systems. The Bromley-MCM model addresses this limitation by utilizing MCM to predict electrolyte-specific parameters of the Bromley model, which can be organized in a matrix format with cations and anions as rows and columns. The model was trained on a comprehensive dataset of 478 electrolytes at 298 K, resulting in a completed matrix of Bromley parameters for 83 cations and 112 anions. This allows for the prediction of concentration-dependent activities for 9,296 electrolytes, significantly extending the Bromley model's applicability while maintaining high predictive accuracy. The performance of Bromley-MCM was benchmarked against existing models, demonstrating its superior predictive capabilities.
Methodology
The Bromley-MCM model combines the Bromley model with a matrix completion method. The parameters of the Bromley model are arranged in a matrix format, where the rows represent cations and the columns represent anions. The MCM is trained on experimental data to predict the missing parameters, effectively solving a matrix completion problem due to sparse data for many electrolyte combinations.
Results
The Bromley-MCM model successfully completed the parameter matrix for 83 cations and 112 anions, enabling accurate predictions of activities for 9,296 electrolytes. The model showed high predictive accuracy when evaluated against electrolytes not included in the training dataset, outperforming existing models such as the ion-specific Bromley model and the Simoes model.
Implications
The hybrid Bromley-MCM model has significant implications for the fields of thermodynamics and electrochemistry, providing a robust tool for predicting electrolyte activities in various applications, including energy storage and desalination. Its ability to generalize across unstudied systems can facilitate advancements in industrial processes that rely on accurate electrolyte modeling.