AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
66
Papers today
8h
Update frequency
7
Days of history
MixQuant: Adaptive Mixed-Precision Quantization for Large Language Models
NLP
Large Language Models
Efficient ML
- MixQuant adapts mixed-precision quantization to varying deployment budgets without needing per-budget calibration.
- The framework considers the sensitivity of layers based on the quantization levels of upstream layers.
- MixQuant incorporates a tail regularization term to optimize bit allocation and improve model performance.
- It outperforms existing methods in terms of accuracy and perplexity across various large language models.
Read more
MixQuant: Adaptive Mixed-Precision Quantization for Large Language Models
Summary
The paper introduces MixQuant, an adaptive mixed-precision quantization framework designed for large language models (LLMs). Traditional mixed-precision quantization methods allocate bitwidths based on a fixed memory budget, which can vary across deployments and is often unknown at calibration time. MixQuant addresses this limitation by recognizing that a layer's sensitivity to quantization is influenced by the bitwidths of its upstream layers. The proposed framework operates in three stages: first, it builds a budget-agnostic distortion table by scoring each layer's output distortion under random quantized upstream configurations; second, it estimates technique parameters based on the actual allocations produced by a greedy solver; and third, it incorporates a tail regularization term to prevent low-bitwidth assignments from being disproportionately damaging. The framework is technique-agnostic and can wrap around existing quantizers like AWQ and GPTQ. Experimental results demonstrate that MixQuant outperforms existing adaptive and mixed-precision baselines, achieving accuracy improvements of up to 8 points and reducing perplexity significantly under tight memory budgets.
Methodology
MixQuant employs a three-stage process: (1) it creates a distortion table by evaluating each layer's output distortion across random quantized configurations of upstream layers, (2) it estimates quantization parameters based on actual allocations from a greedy solver, and (3) it applies a tail regularization to optimize bit allocation, ensuring that low-bitwidth layers are not neglected.
Results
MixQuant demonstrated significant improvements in model performance, achieving up to 8 points increase in average accuracy and reducing perplexity from 12.43 to 10.70 at the tightest budget across various large language models, including Llama-3.2-3B, Llama-2-7B, and Mistral-7B.
Implications
The MixQuant framework has the potential to enhance the deployment of large language models in resource-constrained environments by optimizing quantization strategies, leading to better performance without the need for extensive retraining or calibration.
A Physics-Informed Neural Operator for Thermal Ranking of Low-Cost Wall Materials in Hot-Dry Climates
Theory
- Introduces a two-stage computational framework for thermal ranking of indigenous wall materials.
- Utilizes a Physics-Informed Neural Operator (PINO) for efficient learning of thermal performance.
- Demonstrates that clay-straw adobe is the most effective low-cost material for thermal comfort.
- Identifies a regime boundary that affects material ranking based on outdoor temperature conditions.
Read more
A Physics-Informed Neural Operator for Thermal Ranking of Low-Cost Wall Materials in Hot-Dry Climates
Summary
This paper addresses the challenge of selecting cost-effective building materials that minimize heat penetration in low-income rural housing in hot-dry climates, particularly in Sindh, Pakistan, where summer temperatures can exceed 45 Β°C. The authors propose a two-stage computational framework for thermal ranking of five indigenous wall materials: mud brick, clay-straw adobe, lime-stabilized bamboo panel, fired clay brick, and lime-mud composite. The first stage employs a validated Crank-Nicolson finite difference method (FDM) to solve the one-dimensional transient heat equation under diurnal solar and outdoor air-temperature conditions, generating 1500 solutions through Latin Hypercube sampling. The second stage utilizes a Physics-Informed Neural Operator (PINO) with a Fourier Neural Operator (FNO) backbone to learn the parameter-to-solution mapping, ensuring data fidelity and PDE consistency. The trained PINO achieves a relative L2 field error of 5.14Γ10β4 and a mean absolute error of 0.201 K on peak inner surface temperature, accurately preserving the FDM material ranking. The study finds that clay-straw adobe offers the best cost-performance index among the materials evaluated, while a climate sweep reveals a regime boundary that alters material ranking under varying outdoor conditions. This framework supports evidence-based material selection for post-flood reconstruction in hot-dry regions.
Methodology
The methodology consists of two main stages: first, a Crank-Nicolson finite difference method (FDM) is used to solve the transient heat equation, generating a large dataset of thermal responses across a nine-dimensional parameter space. Second, a Physics-Informed Neural Operator (PINO) is trained to learn the mapping from parameters to thermal solutions, enforcing physical consistency through a PDE residual loss.
Results
The trained PINO achieves a relative L2 field error of 5.14Γ10β4 and a mean absolute error of 0.201 K on peak inner surface temperature, accurately reflecting the FDM material ranking. The study identifies clay-straw adobe as the best material for thermal performance, with a climate sweep revealing a regime boundary that changes the ranking of materials under different outdoor conditions.
Implications
The findings have significant implications for the selection of building materials in low-income rural housing, particularly in post-disaster reconstruction scenarios. The framework can guide the choice of materials that enhance indoor thermal comfort without relying on mechanical cooling, which is often unaffordable in these contexts.
AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction
Theory
- AMPBench-MT integrates multiple evaluation endpoints for antimicrobial peptides, including potency, spectrum, and safety.
- The benchmark employs a homology-controlled protocol to ensure robust evaluation across diverse peptide sequences.
- High binary classification performance does not guarantee effective assay outcomes, highlighting the need for endpoint-aware assessments.
- The study reveals that traditional performance metrics can be misleading, particularly in the context of scarce negative data.
Read more
AMPBench-MT: A Homology-Controlled Benchmark for Antimicrobial Peptide Potency, Spectrum, and Safety Prediction
Summary
The paper introduces AMPBench-MT, a novel benchmark designed to enhance the evaluation of antimicrobial peptides (AMPs) by integrating multiple assay-derived endpoints such as potency, spectrum, and safety. Existing benchmarks primarily focus on binary recognition or specific aspects of AMP evaluation, lacking a comprehensive approach that considers the interdependencies of various endpoints. AMPBench-MT standardizes peptide records and organizes them into a homology-controlled framework that includes binary recognition, species-conditioned potency regression, and endpoint-specific assessments. The authors highlight that high performance in binary classification does not necessarily correlate with effective assay outcomes, emphasizing the need for a more nuanced evaluation that considers multiple factors influencing AMP efficacy. The benchmark employs a 30% identity clustering method to control for sequence homology and provides a detailed analysis of model performance across 161 evaluations, revealing that traditional metrics can be misleading. The paper advocates for a shift in AMP evaluation towards a more endpoint-aware methodology, which is crucial for advancing antimicrobial discovery in the face of rising antibiotic resistance.
Methodology
The authors developed AMPBench-MT by standardizing peptide sequences and organizing them into a framework that includes binary recognition, species-conditioned potency regression, and endpoint-specific assessments. They utilized a 30% identity clustering method to control for sequence homology and conducted 161 model evaluations across various machine learning families, including classical models, small neural networks, and graph neural networks.
Results
The evaluation demonstrated that high binary performance does not reliably predict assay-endpoint behavior. The analysis identified that frozen protein-language-model embeddings were a significant source of prediction error, while classical regression models performed comparably. The study also found that traditional performance metrics could misrepresent the effectiveness of AMPs, particularly when negative data is scarce.
Implications
The findings suggest that AMP evaluation should transition from simple recognition metrics to a more comprehensive, endpoint-aware auditing process. This approach could significantly enhance the prioritization of antimicrobial candidates in drug discovery, particularly in addressing the challenges posed by antimicrobial resistance.
Behavior-Driven Explainability
Interpretability
- Introduces Behavior-Driven Explainability (BDX) as a method for deriving explanations from system specifications.
- Utilizes Behavior-Driven Development (BDD) to create structured scenarios that outline expected system behavior.
- Demonstrates the application of BDX through a case study on a RISC-V processor, highlighting its effectiveness in explaining system exceptions.
- Formalizes the steps for implementing BDX, providing actionable guidance for developers.
Read more
Behavior-Driven Explainability
Summary
This paper introduces Behavior-Driven Explainability (BDX), a novel approach to enhance the explainability of complex systems, particularly in safety-critical domains. As systems grow in complexity, understanding their behavior becomes increasingly challenging. The authors propose utilizing Behavior-Driven Development (BDD) to specify system behavior through structured scenarios, which can be directly translated into explanations. BDX aims to provide transparency and trust in system operations by deriving explanations from the specified behavior. The methodology is formalized, and a case study involving the explanation of exceptions in a RISC-V processor demonstrates the practical application of BDX. The results indicate that BDX can effectively support various development activities and enhance understanding of system behavior across different stages of the development life cycle.
Methodology
The authors leverage Behavior-Driven Development (BDD) to define system behavior through structured scenarios. BDX is formalized to derive explanations based on these scenarios, allowing for a clear and systematic approach to understanding system actions. The methodology includes identifying relevant scenarios and reformulating the action chains to generate explanations.
Results
The case study on the RISC-V processor illustrates that BDX can effectively explain exceptions and support various development activities. The formalization of BDX provides a clear pathway for implementation, demonstrating its applicability at different stages of the system development life cycle.
Implications
BDX has significant implications for the development of safety-critical systems, where understanding system behavior is crucial. By enhancing explainability, BDX can improve user trust, facilitate error identification, and support decision-making processes in both human and machine interactions.
ACRL: Adaptive Control of Training-Inference Discrepancy for Stable Reinforcement Learning
Reinforcement Learning
Large Language Models
Efficient ML
- ACRL framework stabilizes RL training by controlling training-inference discrepancies.
- The method enhances exploration and improves accuracy through adaptive token-level gradient adjustments.
- Empirical results show ACRL matches BF16 accuracy and outperforms importance sampling fixes under FP8 quantization.
- The approach addresses the challenges posed by architectural separations in training and inference engines.
Read more
ACRL: Adaptive Control of Training-Inference Discrepancy for Stable Reinforcement Learning
Summary
The paper addresses the instability issues in Reinforcement Learning (RL) training for Large Language Models (LLMs) caused by the training-inference discrepancy, which arises from architectural separations and low-precision quantization during inference. The authors propose a novel framework called Adaptive Control Reinforcement Learning (ACRL) that adaptively regulates this discrepancy to ensure stable training. ACRL utilizes measurements of sequence-level discrepancies to dynamically adjust token-level gradient ratios, preventing training collapse while enhancing exploration and accuracy. Experimental results demonstrate that ACRL effectively stabilizes RL training under aggressive FP8 quantization, achieving accuracy comparable to BF16 baselines and outperforming traditional importance sampling fixes. The paper contributes a new perspective on stable RL training and provides empirical validation of the ACRL framework's effectiveness in maintaining training-inference alignment.
Methodology
The ACRL framework employs a controller that adaptively maintains the training-inference discrepancy within a reasonable range. It uses sequence-level measurements to adjust token-level gradient ratios, ensuring that the discrepancy does not lead to training collapse or accuracy loss. This adaptive control mechanism inherently increases policy entropy, promoting better exploration during training.
Results
The experimental validation reveals that ACRL effectively manages the training-inference discrepancy under aggressive FP8 quantization, leading to stable RL training. Models trained with ACRL achieve accuracy levels comparable to those using BF16 precision and show significant improvements over previous importance sampling methods.
Implications
The findings suggest that ACRL can be a valuable approach for improving the stability and efficiency of RL training in LLMs, particularly in scenarios where low-precision quantization is necessary for computational efficiency. This could lead to broader applications in various RL tasks and enhance the deployment of LLMs in real-world applications.
Monitoring Post-Disaster Urban Recovery Using High-Resolution SAR Time Series and Unsupervised Learning: Evidence from the 2023 TΓΌrkiye-Syria Earthquake
Time Series
- Proposes an unsupervised framework for monitoring urban recovery using SAR time series.
- Identifies heterogeneous reconstruction dynamics in cities affected by the 2023 TΓΌrkiye-Syria earthquakes.
- Demonstrates the complementary nature of SAR and nighttime lights data in assessing recovery.
- Highlights the effectiveness of SAR-based anomaly detection in capturing early-stage reconstruction processes.
Read more
Monitoring Post-Disaster Urban Recovery Using High-Resolution SAR Time Series and Unsupervised Learning: Evidence from the 2023 TΓΌrkiye-Syria Earthquake
Summary
This paper addresses the challenge of monitoring urban recovery following natural disasters, specifically focusing on the 2023 TΓΌrkiye-Syria earthquakes. The authors propose an unsupervised learning framework that utilizes multi-temporal Synthetic Aperture Radar (SAR) observations to detect anomalies indicative of reconstruction activities. By leveraging COSMO-SkyMed (CSK) time series data, the framework generates spatially explicit recovery maps that reveal the dynamics of reconstruction across four affected cities. The study highlights the heterogeneous nature of recovery processes, identifying distinct patterns such as the establishment of temporary settlements and the rebuilding of residential areas. Additionally, the authors compare SAR-based findings with nighttime lights (NTL) data, demonstrating that while NTL reflects socioeconomic activity and electricity restoration, SAR anomalies provide insights into structural changes and earlier stages of recovery. This research underscores the potential of combining SAR observations with unsupervised learning for effective post-disaster monitoring, especially in the absence of labeled datasets, and contributes to advancing Earth Observation capabilities for long-term recovery assessments.
Methodology
The authors developed an unsupervised learning framework that analyzes multi-temporal SAR data from COSMO-SkyMed to detect persistent anomalies associated with reconstruction activities. This approach generates spatially explicit recovery maps and compares SAR findings with nighttime lights data to assess recovery dynamics.
Results
The application of the framework revealed distinct recovery patterns across the four cities, characterized by clusters of persistent anomalies linked to significant transformations such as reconstruction and temporary settlements. The comparison with NTL data illustrated the advantages of SAR in capturing structural changes at earlier recovery stages.
Implications
The findings suggest that the proposed SAR-based monitoring framework can significantly enhance the understanding of urban recovery processes post-disaster. It offers a scalable methodology for tracking reconstruction dynamics, which can inform disaster risk reduction strategies and improve response efforts in future events.
Multiclass Classification without Labels via Posterior Simplex Geometry
Theory
- Extension of CWoLa to multiclass classification without instance-level labels.
- Establishment of a (K-1)-simplex geometry for Bayes-optimal mixture classifiers.
- Development of prior-free recovery methods for latent class structures.
- Demonstrated effectiveness on multiple datasets, improving performance in label-scarce domains.
Read more
Multiclass Classification without Labels via Posterior Simplex Geometry
Summary
This paper addresses the challenge of multiclass classification in scenarios where reliable instance-level labels are unavailable. The authors extend the concept of Classification without Labels (CWoLa) from binary to multiclass settings, demonstrating that a classifier trained to distinguish between different mixtures of classes can recover the underlying class structure without knowing the mixture proportions. They establish that the Bayes-optimal mixture classifier maps data points into a (K-1)-simplex in mixture-posterior space, where K represents the number of latent classes. The vertices of this simplex correspond to the latent classes, allowing for the recovery of latent class structures through two proposed methods: post-hoc simplex fitting and a bottleneck architecture. Experiments conducted on datasets such as MNIST, CIFAR-10, and Galaxy10 DECaLS show that the proposed methods can effectively recover latent classes and their proportions from mixture identities alone, thereby bridging the gap between weakly supervised and fully supervised learning.
Methodology
The authors propose a framework that trains a classifier to predict mixture identities from unlabeled data. They leverage the geometry of the posterior simplex to recover latent class structures using two approaches: post-hoc simplex fitting and a bottleneck architecture, without requiring instance-level labels or known mixture proportions.
Results
The experiments on MNIST, CIFAR-10, and Galaxy10 DECaLS indicate that the proposed methods can successfully recover latent classes and their proportions based solely on mixture identities, achieving performance that narrows the gap between weakly supervised and fully supervised learning.
Implications
This work provides a scalable and mathematically grounded approach for multiclass discovery in domains where labeled data is scarce, potentially benefiting various scientific and clinical applications where expert annotation is challenging.
Spend Experts Where You Are Unsure: Confidence-Adaptive Routing for Mixture-of-Experts LoRA
NLP
Large Language Models
Efficient ML
- CARE adapts the number of active experts based on token uncertainty, improving computational efficiency.
- The method uses existing router output distributions as signals for confidence and ambiguity.
- CARE achieves better accuracy with fewer experts compared to traditional fixed top-k approaches.
- The approach is parameter-free and can be easily integrated into existing MoE-LoRA architectures.
Read more
Spend Experts Where You Are Unsure: Confidence-Adaptive Routing for Mixture-of-Experts LoRA
Summary
This paper introduces CARE (Confidence-Adaptive Routing of Experts), a novel method for optimizing the routing of tokens to experts in Mixture-of-Experts (MoE) architectures integrated with Low-Rank Adaptation (LoRA). Traditional MoE-LoRA approaches activate a fixed number of experts for each token, which can lead to inefficient resource allocation, particularly when tokens vary in complexity. The authors leverage the output distribution from the router as a per-token uncertainty signal, allowing for adaptive expert allocation based on the model's confidence in its predictions. CARE employs a nucleus sampling-like mechanism to activate experts based on their router weights until a cumulative mass threshold is reached, with additional experts admitted if there is disagreement among the initially selected ones. This approach is parameter-free, requires only a single forward pass, and can be integrated into existing MoE-LoRA frameworks without altering the training process. The authors validate CARE across various benchmarks, demonstrating improved accuracy and reduced expert usage compared to fixed top-k methods, while also enhancing out-of-distribution detection capabilities.
Methodology
The CARE method utilizes the router's output distribution to determine the confidence level for each token. It employs a nucleus sampling strategy to admit experts based on their router weights until a predefined cumulative mass threshold is reached. If the initially selected experts show disagreement in their outputs, additional experts are activated. This process is calibrated using a budget thermostat to ensure that the average number of active experts aligns with a target budget, all while maintaining a single forward pass and no additional parameters.
Results
CARE was tested on eight commonsense benchmarks using LLaMA-3.1-8B and Qwen2.5-7B models. The results showed that CARE improved accuracy over fixed top-k MoE-LoRA methods, achieving a notable increase of +0.5 on LLaMA commonsense tasks while activating 12% fewer experts. Additionally, CARE demonstrated improved out-of-distribution detection compared to existing methods such as MSP and entropy-based proxies.
Implications
The findings suggest that adaptive expert routing can significantly enhance the efficiency and effectiveness of large language models, particularly in scenarios where token complexity varies. This could lead to more resource-efficient deployment of MoE architectures in real-world applications, such as conversational agents, automated reasoning systems, and other NLP tasks requiring high accuracy and adaptability.
Human Preference aligned Tabular Similarity
Theory
- Current tabular embedding methods focus on prediction tasks, neglecting human preference alignment for similarity search.
- Standard evaluation metrics fail to capture the nuances of human-perceived similarity in industrial applications.
- The proposed workflow integrates human feedback into the evaluation of tabular embeddings for improved trustworthiness.
- Human preference alignment is crucial for ensuring that retrieved similar items are meaningful to domain experts.
Read more
Human Preference aligned Tabular Similarity
Summary
This paper addresses the limitations of existing tabular embedding methods that are primarily optimized for prediction tasks rather than for human preference aligned similarity rankings. The authors argue that current evaluation metrics do not adequately assess the trustworthiness of embeddings for similarity search, particularly in industrial applications such as Product Lifecycle Management (PLM). They propose a new evaluation procedure that incorporates human preferences into the assessment of task-agnostic tabular embeddings. The proposed workflow includes embedding tabular records, selecting anchor records based on user-defined criteria, and refining retrieval quality through continuous user feedback. The paper emphasizes the need for broader evaluation standards that align with trustworthy AI principles, advocating for a shift towards human-centered evaluation methods that consider diverse user preferences in similarity assessments.
Methodology
The authors propose a three-step workflow for human preference ranking of embedding similarity in tabular data. This includes embedding tabular records using a selected model, selecting anchor records based on user-defined criteria, and refining the retrieval process through continuous user engagement and feedback.
Results
The paper illustrates the necessity of human preference alignment in evaluating tabular embeddings through a PLM use case, highlighting that current metrics do not adequately reflect the semantic relevance of retrieved items to domain experts. The proposed evaluation procedure aims to enhance the trustworthiness of similarity search in industrial applications.
Implications
The findings suggest that incorporating human preferences into the evaluation of tabular embeddings can lead to more reliable and meaningful similarity searches in industrial contexts. This approach could improve user trust and satisfaction with AI systems in various applications, particularly in PLM and other domains reliant on tabular data.
Learning from the Descent Direction: Adaptive Gradient Descent under One-Sided HΓΆlder Regularity
Optimization
Theory
- Introduction of a one-sided HΓΆlder descent model that focuses on directional curvature.
- Development of an adaptive step-size rule based on local directional curvature and current gradient norm.
- Establishment of a best-iterate stationarity guarantee with a rate determined by the HΓΆlder exponent.
- Comparison of the proposed method against various baseline gradient descent strategies.
Read more
Learning from the Descent Direction: Adaptive Gradient Descent under One-Sided HΓΆlder Regularity
Summary
This paper investigates an adaptive gradient descent method tailored for continuously differentiable, potentially nonconvex optimization problems under one-sided HΓΆlder regularity. Traditional gradient descent methods often rely on full gradient variation, which can lead to conservative step sizes, especially when much of the gradient variation is orthogonal to the descent direction. The authors propose a novel adaptive gradient method that focuses on a one-sided HΓΆlder curvature estimate, which only considers the directional component relevant to the descent direction. This approach allows for more effective step size selection based on local curvature rather than a fixed decay schedule. The method is rigorously analyzed, providing a best-iterate stationarity guarantee with a rate dependent on the HΓΆlder exponent. The authors validate their method through experiments on two benchmarks: a binary classification task and a nonconvex HΓΆlder regression problem, where their approach outperformed traditional scalar gradient methods in terms of final objective values and gradient norms. The findings suggest that one-sided HΓΆlder curvature can serve as a robust adaptive signal for step size selection in gradient descent algorithms.
Methodology
The authors formulate an adaptive gradient descent algorithm that utilizes one-sided HΓΆlder curvature conditions to determine step sizes. The method adapts to the local geometry of the descent direction, allowing for more efficient optimization without relying on a pre-defined schedule. Theoretical guarantees are provided for the method's convergence properties, and it is evaluated against established gradient descent techniques in controlled experiments.
Results
The proposed one-sided HΓΆlder method achieved the lowest final cross-entropy, objective value, and gradient norm in a binary classification benchmark, as well as the lowest final objective gap in a nonconvex HΓΆlder regression benchmark. These results indicate that the method effectively utilizes directional curvature to enhance optimization performance.
Implications
The findings suggest that adaptive gradient descent methods can be significantly improved by focusing on one-sided curvature, leading to more efficient optimization in machine learning tasks. This approach could be particularly beneficial in scenarios where traditional methods struggle due to conservative step size choices.
Inverse RL Helps Align AI by Imitating Humans
NLP
Large Language Models
Reinforcement Learning
- PARED introduces a method for deriving explicit, inspectable rewards from expert demonstrations without needing task-specific annotations.
- The approach allows for both inference-time selection and on-policy reinforcement learning optimization.
- PARED supports audience-conditioned alignment, demonstrating that improvements can be achieved for multiple target audiences simultaneously.
Read more
Inverse RL Helps Align AI by Imitating Humans
Summary
This paper addresses the challenge of aligning language models to reflect desirable behaviors such as helpfulness and safety. Traditional methods rely on supervised fine-tuning or reinforcement learning with human feedback, which often underutilize expert demonstrations. The authors propose a novel approach called Projected Alignment Reward Estimated from Demonstrations (PARED), inspired by inverse reinforcement learning. PARED infers an implicit reward from expert demonstrations by mapping responses to a low-dimensional feature space and using a discriminator to separate expert outputs from policy samples. This method does not require task-specific preference annotations, allowing for a more flexible and interpretable reward structure. The authors demonstrate PARED's effectiveness in both inference-time reranking and on-policy reinforcement learning, showing that it can improve model performance without additional supervised training. Furthermore, PARED facilitates audience-conditioned contextual alignment, enabling a single policy to adapt to different audience preferences without compromising performance across groups.
Methodology
The methodology involves mapping response-level trajectories from expert demonstrations and policy outputs into a practitioner-chosen feature space. A logistic discriminator is employed to distinguish between expert and policy samples, generating an expert-likeness score that serves as the alignment reward. This reward can then be utilized for both best-of-N reranking and KL-regularized on-policy reinforcement learning, allowing the model to optimize based on the inferred reward without further training on the demonstrations.
Results
Experimental results indicate that the PARED approach significantly enhances the performance of a base policy in both reranking and reinforcement learning contexts. The inferred rewards lead to improvements beyond what is achieved through standard supervised fine-tuning, and the audience-conditioned alignment shows that the model can effectively cater to different user preferences without trade-offs.
Implications
The findings suggest that PARED can be a powerful tool for aligning AI systems with human values and preferences, making it applicable in various domains where user alignment is critical. This method could enhance the development of more responsive and responsible AI systems, particularly in areas like customer service, content generation, and interactive applications.
Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
Reinforcement Learning
- Introduction of the Reinformed Dreamer algorithm, which improves upon the Informed Dreamer.
- Identification of limitations in the learning objectives of existing asymmetric reinforcement learning models.
- Development of a new learning objective using a privileged variational encoder for better representation learning.
- Demonstration of improved performance and convergence speed across multiple benchmarks.
Read more
Reinformed Dreamer: An Asymmetric World Model Efficiently Trained through Latent Guidance
Summary
The paper introduces the Reinformed Dreamer, a novel algorithm in the realm of model-based reinforcement learning that enhances learning efficiency through asymmetric representation learning. The authors identify limitations in the Informed Dreamer algorithm, particularly in how it learns privileged information representations. They propose a new learning objective that utilizes a privileged variational encoder alongside an unprivileged encoder, allowing for improved representation learning without complicating the imagination process. The Reinformed Dreamer is tested across various benchmarks, demonstrating consistent improvements in convergence speed and overall performance compared to previous asymmetric approaches. This work emphasizes the importance of leveraging additional information during training to enhance representation learning in reinforcement learning settings, particularly under partial observability.
Methodology
The authors propose a new asymmetric representation learning objective that combines a privileged variational encoder with an unprivileged encoder. This approach allows the Reinformed Dreamer to maintain a single world model while effectively learning representations of both privileged and unprivileged information. The algorithm is evaluated through experiments on various benchmark environments to assess its performance against existing methods.
Results
The Reinformed Dreamer consistently outperformed or matched the performance of previous algorithms, such as the Informed Dreamer and Scaffolder, in terms of convergence speed and final performance across several benchmark environments.
Implications
The findings suggest that reinforcement learning algorithms can significantly benefit from additional supervisory information during training, leading to more efficient learning processes and better performance in complex decision-making tasks. This has potential applications in robotics, game playing, and other domains where reinforcement learning is applied.
Self-Boosting Vision-Language Models with Noisy Student On-Policy Self-Distillation
Multimodal
- NOPD enables self-improvement of VLMs without external supervision.
- The method utilizes prediction discrepancies between clean and corrupted inputs for self-supervision.
- NOPD outperforms traditional reinforcement learning and distillation methods in visual reasoning tasks.
- Significant performance gains were observed on multiple benchmarks, demonstrating generalization capabilities.
Read more
Self-Boosting Vision-Language Models with Noisy Student On-Policy Self-Distillation
Summary
This paper introduces NOPD (Noisy Student On-Policy Self-Distillation), a novel self-distillation method aimed at enhancing vision-language models (VLMs) without relying on external supervision or ground-truth answers. Traditional post-training methods for VLMs often depend on human-annotated data or external models, which can limit their scalability and adaptability. NOPD leverages the discrepancies in predictions between clean and corrupted inputs to create a self-supervision signal, allowing the model to learn from its own predictions. The authors demonstrate the effectiveness of NOPD across five visual reasoning tasks, showing that it can match or exceed the performance of existing reinforcement learning and knowledge distillation methods. Notably, NOPD achieved a significant improvement of 20 points on the validation set of Qwen2.5-VL-7B when trained with a limited dataset of 2.1K samples from Geometry3K and also showed strong generalization capabilities on out-of-distribution test sets, including a 7.4 point gain on MathVista. The results indicate that NOPD is a versatile approach that can enhance various VLMs across multiple benchmarks, achieving improvements without the need for external supervision.
Methodology
NOPD employs a self-distillation framework where a single model acts as both teacher and student. The teacher processes clean inputs while the student learns from corrupted inputs, using the teacher's predictions as token-level supervision. This method eliminates the need for high-quality responses or external models, allowing for effective learning through self-induced supervision.
Results
NOPD was validated through extensive experiments on five visual reasoning benchmarks, consistently showing improvements over existing methods. The model achieved a 20-point increase on the validation set of Qwen2.5-VL-7B and a 7.4 point gain on MathVista, demonstrating its effectiveness and generalization across different tasks and models.
Implications
The findings suggest that VLMs can be significantly improved without the constraints of external supervision, paving the way for more scalable and adaptable models in various applications, including visual question answering and reasoning tasks. This could lead to advancements in AI systems that require less human intervention and can learn more autonomously.
Mechanisms of Width Scaling in Normalized Residual Networks: The Effective Alignment Dimension
Theory
Optimization
- Introduction of the effective alignment dimension as a measure of signal-noise geometry in neural networks.
- Derivation of a finite-sample certificate for directional stability in model expansion, independent of previous assumptions.
- Empirical validation showing that wider models exhibit better alignment and lower misalignment.
- Direct interventions confirm that alignment statistics can predict changes in held-out loss.
Read more
Mechanisms of Width Scaling in Normalized Residual Networks: The Effective Alignment Dimension
Summary
This paper addresses the challenge of determining whether the expansion direction identified from finite training data in neural networks remains beneficial for unseen data. The authors introduce the concept of the effective alignment dimension, which quantifies the signal-noise geometry of activation gradients. By deriving the mean and variance of the inner product between training and test gradients, they establish a finite-sample upper bound on misalignment probability, which is dependent on the effective alignment dimension and effective sample size. This framework integrates a high-probability condition for test-risk improvement into the train-test residual-expansion methodology. Experiments conducted on width-controlled models, including LLaMA-style Transformers and ResNet-20, demonstrate that wider models tend to have larger effective alignment dimensions and lower empirical misalignment. The study also confirms that the alignment statistic can predict the sign and magnitude of changes in held-out loss, providing a more reliable framework for understanding the implications of model width on performance.
Methodology
The authors derive an exact second-moment characterization of the alignment between training and test gradients. They analyze the alignment statistic directly rather than separately controlling training and test deviations. This approach leads to an instance-specific finite-sample certificate that incorporates the effective alignment dimension and effective sample size, allowing for a high-probability guarantee of directional transfer in model expansion.
Results
The experiments reveal that the effective alignment dimension generally increases with model width, correlating with reduced empirical misalignment. The findings indicate that positive alignment is a reliable predictor of the sign and magnitude of changes in held-out loss, supporting the theoretical framework established in the paper.
Implications
The results suggest that understanding the effective alignment dimension can guide the design of neural networks, particularly in determining the benefits of width scaling. This could lead to more efficient model growth strategies and improved generalization on unseen data, which is crucial for applications in various domains such as computer vision and natural language processing.
DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment
Multimodal
- DynaBridge formulates mental health assessment as a structured ordinal cross-task problem, linking item responses to risk predictions.
- The framework integrates multimodal data (acoustic, visual, textual) with LLM-generated summaries to enhance predictive accuracy.
- It introduces DASS-aware item-to-risk reconstruction and confidence-aware refinement to improve consistency in predictions.
- DynaBridge outperforms existing methods in predicting depression, anxiety, and stress risk, demonstrating its effectiveness in mental health assessment.
Read more
DynaBridge: Dynamic Summary-Guided Cross-Task Multimodal Fusion for DASS-Structured Mental Health Assessment
Summary
The paper introduces DynaBridge, a novel framework designed for assessing mental health through multimodal behavioral analysis, specifically targeting the DASS-21 (Depression, Anxiety, Stress Scale). Traditional methods often overlook the psychometric structure of questionnaire labels, leading to inconsistent predictions. DynaBridge addresses this by integrating acoustic, visual, and textual cues from multiple sessions, enhanced by summaries generated from frozen large language models (LLMs). These summaries serve as participant-level semantic evidence, allowing for the prediction of ordinal item distributions and the reconstruction of risk evidence for depression, anxiety, and stress. The framework employs a confidence-aware refinement strategy to incorporate high-confidence semantic cues, ensuring consistency between item-level predictions and overall risk assessments. The results demonstrate that DynaBridge significantly outperforms existing baseline models and multimodal methods, achieving a mean F1 score of 0.5012 for risk prediction and a mean Quadratic Weighted Kappa (QWK) of 0.3216 for item prediction on the AdoDAS validation split, showcasing the effectiveness of combining multimodal cues with psychometric structures.
Methodology
DynaBridge employs a dynamic summary-guided approach that fuses multimodal behavioral cues with LLM-generated summaries. It encodes data from multiple sessions, reconstructs risk evidence from item-level soft scores, and refines predictions using a confidence-aware strategy. The framework specifically leverages the DASS-21 psychometric structure to ensure that predictions are consistent with the underlying questionnaire design.
Results
On the AdoDAS validation split, DynaBridge achieved a mean F1 score of 0.5012 for predicting depression, anxiety, and stress risks, and a mean QWK of 0.3216 for item-level predictions, outperforming both the official baseline and other representative multimodal methods.
Implications
The successful implementation of DynaBridge suggests significant potential for scalable, automated mental health assessments that leverage multimodal data. This could lead to improved mental health monitoring and intervention strategies, particularly in settings where traditional assessment methods are impractical.
Joint Flow Matching for Generator-Consistent Classification
Generative Models
Theory
Interpretability
- Introduction of Joint Flow Matching (JFM) for consistent generative and discriminative inference.
- JFM assigns opposite roles to noise and data distributions, enabling bidirectional inference.
- The framework guarantees consistency between generative and discriminative conditional distributions.
- Empirical validation shows competitive performance on generation and classification tasks.
Read more
Joint Flow Matching for Generator-Consistent Classification
Summary
This paper introduces Joint Flow Matching (JFM), a novel training framework for continuous normalising flows (CNFs) that addresses the limitations of standard flow matching in generating and classifying paired variables. Traditional flow matching methods typically treat the generation process as unidirectional, which complicates the inference of conditional distributions. JFM resolves this by assigning opposite roles to noise and data distributions at the temporal endpoints, allowing for simultaneous generation and classification. The authors prove that JFM produces a consistent joint distribution, ensuring that the forward and reverse conditional distributions are coherent. The framework is validated on conditional datasets, demonstrating competitive accuracy and well-calibrated confidence scores without the need for post-hoc calibration. JFM's ability to generate images and classify them consistently from a single model highlights its potential for interpretability in discriminative-generative models, providing insights into the relationships between features of the data.
Methodology
The authors develop a continuous normalising flow framework that learns a velocity field to transport probability mass from noise to data. JFM defines endpoint distributions that are factored by design, allowing for the generation of one variable given another and vice versa. The flow is trained to ensure that the learned joint distribution captures the dependencies between the variables, maintaining consistency across generative and discriminative tasks.
Results
JFM was empirically validated on small benchmark datasets, demonstrating that it achieves competitive accuracy in both classification and generation tasks. The results indicate that JFM provides well-calibrated confidence scores and maintains consistency between the generated samples and the features used for classification.
Implications
The development of JFM suggests a new approach to generative modeling that can enhance interpretability and consistency in machine learning applications. This framework could be particularly useful in scenarios where understanding the relationship between features is crucial, such as in medical imaging or other domains requiring reliable classification and generation.
SPARC Segmentation to Prediction via Affine Regression and Counterfactuals
Optimization
Theory
Interpretability
- Introduces a novel propensity modeling framework for B2B e-commerce that addresses unique challenges in transaction prediction.
- Replaces SMOTE with a counterfactual data generation approach (DiCE) for better representation of minority-class samples.
- Utilizes the PyPARC framework to provide calibrated propensity probabilities for customer segmentation.
- Achieves significant improvements in precision over traditional methods, validating the effectiveness of the proposed approach.
Read more
SPARC Segmentation to Prediction via Affine Regression and Counterfactuals
Summary
The paper addresses the challenges of transaction propensity prediction in B2B e-commerce, which differ significantly from B2C contexts due to the heterogeneous behaviors of organizational buyers. Traditional methods like SMOTE for data augmentation are inadequate as they assume within-class feature homogeneity, which does not hold true for B2B procurement cycles. The authors propose a new framework that replaces SMOTE with a synthetic data generation approach using Diverse Counterfactual Explanations (DiCE), which produces more representative minority-class samples. Additionally, they adapt the PyPARC piecewise-affine classification framework to generate calibrated propensity probabilities, allowing for the segmentation of customers into actionable risk tiers. The framework was evaluated on two years of longitudinal data from a large B2B e-commerce platform, achieving a precision of 93.1% at a decision threshold of 0.8, significantly outperforming SMOTE-based baselines. This demonstrates the framework's effectiveness in enhancing marketing campaigns and improving customer activation and ROI.
Methodology
The proposed framework employs Diverse Counterfactual Explanations (DiCE) for synthetic data generation, which is tailored to preserve the feasibility of B2B account states under class imbalance. It also modifies the PyPARC piecewise-affine classification framework to output calibrated propensity probabilities instead of discrete predictions, allowing for better customer segmentation.
Results
The framework achieved a precision of 93.1% at a decision threshold of 0.8, which is a 9.2 percentage-point improvement over SMOTE-based baselines at the same threshold (83.9%) and a 26.1-point improvement at threshold 0.7 (66.04%). These results indicate consistent superiority across various operating points.
Implications
The findings suggest that the proposed framework can significantly enhance the effectiveness of marketing campaigns in B2B e-commerce by enabling more accurate predictions of customer behavior, leading to improved resource allocation and higher returns on investment.
Label-free Industrial Fault Detection via Adversarial Inverse Reinforcement Learning: A System for Run-to-Failure Prognostics
Reinforcement Learning
Time Series
Optimization
- Introduces AIRL as a novel approach for label-free fault detection in industrial machinery.
- Demonstrates the ability to detect gradual degradation effectively without relying on labeled fault data.
- Achieves superior performance in terms of fault detection consistency across multiple datasets.
- Provides practical recommendations for automatic thresholding strategies in deployment.
Read more
Label-free Industrial Fault Detection via Adversarial Inverse Reinforcement Learning: A System for Run-to-Failure Prognostics
Summary
This paper addresses the challenge of machinery fault detection (MFD) in industrial settings, which traditionally relies on supervised learning methods that require extensive labeled data. The authors propose a novel framework using Adversarial Inverse Reinforcement Learning (AIRL) to treat MFD as an offline inverse reinforcement learning problem. This approach allows for the recovery of an intrinsic 'health' reward from observational state transitions without the need for manual reward engineering or fault labels. The AIRL framework is validated across three run-to-failure datasets (HUMS2023, IMS, XJTU-SY), demonstrating its ability to achieve non-saturated post-detection consistency and outperform existing contextual bandit and reconstruction-based methods. The paper also includes a comprehensive evaluation of various automatic thresholding strategies and a computational cost analysis, confirming AIRL's suitability for edge deployment in real-world applications.
Methodology
The authors developed an AIRL framework that learns a reward model from healthy operational trajectories, allowing for anomaly detection as deviations from learned dynamics. This method operates fully offline, avoiding the need for manual reward specification and leveraging a generative adversarial approach to differentiate between expert and policy-generated trajectories.
Results
The AIRL framework achieved earlier fault detection and more stable anomaly scoring compared to conventional methods across three distinct datasets. It was the only method to maintain non-saturated post-detection consistency, while other methods failed to detect gradual degradation or collapsed into always-anomalous states.
Implications
The proposed AIRL framework has significant implications for industrial applications, particularly in environments where labeled fault data is scarce. It offers a robust solution for real-time fault detection and prognosis, enhancing operational reliability and reducing maintenance costs.
Breaking the Periodicity Assumption: Robust Tensorial Multi-View Clustering via Graph-Spectral Low-Rank Learning
Graph Learning
Theory
Efficient ML
- Existing t-SVD-based TMC methods are sensitive to sample ordering due to an implicit periodicity assumption.
- The proposed graph-spectral low-rank tensor learning framework addresses this sensitivity by using a data-driven graph Fourier Transform.
- An anchor-based variant is introduced to improve scalability for large datasets.
- Extensive experiments show competitive or superior performance compared to existing TMC methods.
Read more
Breaking the Periodicity Assumption: Robust Tensorial Multi-View Clustering via Graph-Spectral Low-Rank Learning
Summary
This paper addresses a critical flaw in existing tensorial multi-view clustering (TMC) methods that rely on the Fast Fourier Transform (FFT) for low-rank tensor learning. The authors reveal that these methods depend on an implicit 'periodicity assumption' due to the arrangement of samples, which can lead to performance degradation when sample ordering is randomized. They systematically investigate the algebraic and spectral mechanisms behind this sensitivity and propose a novel graph-spectral low-rank tensor learning framework that utilizes a data-driven graph Fourier Transform (GFT) instead of a fixed Fourier basis. This approach captures the intrinsic manifold structure of the data without relying on specific sample ordering. Additionally, an anchor-based variant is developed to enhance efficiency for large-scale datasets. Extensive experiments demonstrate that the proposed methods outperform state-of-the-art TMC approaches, validating the effectiveness of the graph-spectral framework.
Methodology
The authors investigate the limitations of FFT-based low-rank tensor learning and propose a graph-spectral framework that employs a graph Fourier Transform (GFT) to capture the data's intrinsic structure. They also develop an anchor-based approach to enhance computational efficiency.
Results
The proposed methods demonstrate significant improvements in clustering performance across various benchmarks, showing robustness to sample permutations and outperforming existing TMC methods.
Implications
This research has implications for improving the robustness and effectiveness of multi-view clustering techniques, particularly in scenarios where sample ordering may vary, thus enhancing the reliability of clustering outcomes in real-world applications.
Capacity-Aware Deep Learning for Generalizable Traffic Volume Estimation Across Links and Cities
Graph Learning
Time Series
Theory
- Proposes a link-level learning framework for traffic volume estimation using widely available data.
- Introduces a capacity-aware formulation that incorporates traffic-theoretic constraints.
- Evaluates the model under intra-network and inter-network generalization settings.
- Demonstrates superior performance compared to state-of-the-art methods in spatial distribution shifts.
Read more
Capacity-Aware Deep Learning for Generalizable Traffic Volume Estimation Across Links and Cities
Summary
This paper addresses the challenge of estimating traffic volumes across urban networks, particularly in areas with sparse sensor coverage. The authors propose a novel link-level learning framework that utilizes widely available territorial data, such as probe speed profiles, road descriptors, and weather observations, to estimate hourly traffic volumes without relying on fixed sensor measurements. The framework is evaluated under two generalization settings: intra-network (unseen links within the same network) and inter-network (unseen cities). A key contribution is the introduction of a capacity-aware formulation that models traffic volume as a product of link-specific structural capacity and an hourly utilization ratio, embedding traffic-theoretic constraints into the learning process. The results demonstrate that this approach consistently outperforms state-of-the-art methods, particularly under conditions of spatial distribution shift, highlighting the effectiveness of integrating physical constraints into machine learning models for traffic estimation.
Methodology
The authors adopt an inductive, link-level formulation that estimates hourly traffic volumes from various data sources, including speed profiles, meteorological variables, and road descriptors. They implement a capacity-aware bounded prediction mechanism that decomposes traffic volume into link-specific capacity and a bounded utilization ratio, along with a regime-aware utilization modeling approach to capture different traffic dynamics. The model is trained using supervised learning on a sparse subset of links and is evaluated for its generalization capabilities to unseen links and cities.
Results
The proposed model shows significant improvements in traffic volume estimation accuracy compared to existing state-of-the-art methods, particularly in scenarios involving spatial distribution shifts. The incorporation of structural constraints enhances the model's robustness and generalization capabilities, allowing it to effectively estimate traffic volumes even in sparsely instrumented areas.
Implications
This research has important implications for urban planning, traffic management, and environmental monitoring, as it enables accurate traffic volume estimation without the need for extensive sensor networks. The approach can be applied in various urban settings, potentially improving traffic flow management and contributing to better air quality assessments.
XGRVFL-MV: Residual-Coupled Graph-Embedded Multi-View Random Vector Functional Link Network with FleXi Guardian Loss
Graph Learning
Multimodal
Optimization
- Introduction of a novel multi-view RVFL model that incorporates graph embedding and residual coupling.
- Utilization of FleXi Guardian loss to mitigate the impact of large prediction residuals.
- Demonstrated competitive performance on multiple benchmark datasets.
- Incorporation of geometric structure preservation through graph regularization.
Read more
XGRVFL-MV: Residual-Coupled Graph-Embedded Multi-View Random Vector Functional Link Network with FleXi Guardian Loss
Summary
This paper presents the XGRVFL-MV model, a novel approach to multi-view classification using a Residual-Coupled Graph-Embedded Multi-View Random Vector Functional Link Network (RVFL). The authors address the challenges of preserving view-specific geometric structures, managing large prediction residuals, and modeling relationships between multiple views. The proposed model constructs RVFL representations for each view and incorporates graph embedding through intrinsic and penalty graphs based on the Local Fisher Discriminant Analysis (LFDA) weighting scheme. A unique FleXi Guardian (XG) loss function is employed for residual learning, which is bounded and asymmetric, thus limiting the influence of large residuals. Additionally, a residual-coupling term is introduced to maintain consistency among view-specific prediction residuals while preserving their individual representations. The optimization problem is solved using an inversion-free first-order optimization method based on Nesterov accelerated gradient descent. The model is evaluated on several benchmark datasets, including UCI, KEEL, AwA, and Corel5k, demonstrating competitive classification performance compared to baseline methods. The results indicate that the XGRVFL-MV model effectively leverages multi-view information while addressing the limitations of existing methods.
Methodology
The XGRVFL-MV model constructs RVFL representations for each view and integrates graph embedding using intrinsic and penalty graphs. It employs a bounded and asymmetric FleXi Guardian loss for residual learning and introduces a residual-coupling term to ensure consistency among predictions across views. The optimization is performed using Nesterov accelerated gradient descent, which is efficient and avoids the need for matrix inversion.
Results
The experimental results show that the XGRVFL-MV model achieves competitive classification performance across various benchmark datasets, outperforming several baseline methods. Statistical analyses and hyperparameter sensitivity analyses further validate the robustness and effectiveness of the proposed model.
Implications
The findings suggest that the XGRVFL-MV model can be effectively applied in scenarios where multiple feature representations are available, such as image annotation, bioinformatics, and social network analysis. The model's ability to handle noisy data and preserve geometric structures makes it suitable for real-world applications where data quality may vary.
An Embarrassingly Simple Rule-based Visiting Circulation Approach to Trip Destination Prediction
Theory
- Introduction of the Rule-based Visiting Circulation (RVC) model for trip destination prediction.
- The model effectively utilizes origin information and individual trip behaviors without needing training data from the target area.
- RVC outperforms supervised learning methods and heuristics in prediction accuracy.
- The approach addresses significant challenges in destination prediction, including dataset variability and geographic complexity.
Read more
An Embarrassingly Simple Rule-based Visiting Circulation Approach to Trip Destination Prediction
Summary
This paper presents the Rule-based Visiting Circulation (RVC) model aimed at addressing the challenge of trip destination prediction as part of the IEEE Big Data Cup 2022. The task involves predicting trip destinations in a new metropolitan area (Kinki) based on travel information, personal attributes, and origin zones from training datasets of four different metropolitan areas in Japan (Tokyo, Chukyo, Kyushu, and Higashisurugawan). The authors highlight the difficulties posed by the absence of destination data for the target area, variations in datasets, and the complexity of geographic information. The RVC model leverages insights into revisiting behaviors and the relationships between origins and destinations to formulate simple rules for destination prediction without requiring transfer learning from the training areas. Experimental results demonstrate that the RVC model significantly outperforms traditional supervised learning methods and other heuristics, achieving second place in the competition leaderboard.
Methodology
The RVC model employs a rule-based approach that analyzes visiting circulation behaviors between origins and destinations. It formulates a set of comprehensive rules based on trip details and personal attributes to predict destinations in a new metropolitan area, avoiding the need for learning from previous datasets.
Results
The RVC model demonstrated superior performance in both offline evaluations and competition leaderboard submissions, significantly surpassing the baseline accuracy of 0.20879 set by the competition. The model's predictions were robust across different metropolitan areas, validating its effectiveness in the trip destination prediction task.
Implications
The findings have implications for city governance, emergency regulation, and location-based services. The RVC model can support policymaking, travel demand estimation, and transit planning, showcasing its potential utility in urban planning and management.
Quantum Speedups for Stochastic Optimization with Heavy-Tailed Noise
Optimization
Theory
- Introduction of quantum mean estimators for multivariate heavy-tailed distributions.
- Development of QNSGD and QPSGD methods that outperform classical algorithms in query complexity.
- Establishment of quantum lower bounds demonstrating optimality of the proposed estimators.
- Addressing the limitations of classical stochastic optimization methods under heavy-tailed noise.
Read more
Quantum Speedups for Stochastic Optimization with Heavy-Tailed Noise
Summary
This paper investigates stochastic optimization problems that are affected by heavy-tailed gradient noise. The authors introduce a novel quantum mean estimator for multivariate heavy-tailed random variables, which demonstrates lower query complexity compared to classical estimators in low-dimensional scenarios. They also develop an unbiased quantum mean estimator using a generalized multi-level Monte Carlo technique. The paper establishes quantum lower bounds, proving that the proposed quantum estimators are optimal under certain conditions. Building on these estimators, the authors propose two quantum optimization methods: Quantum Normalized Stochastic Gradient Descent (QNSGD) and Quantum Projected Stochastic Gradient Descent (QPSGD), which achieve improved query complexities compared to classical methods. The results indicate that quantum computing can provide significant speedups for stochastic optimization tasks in the presence of heavy-tailed noise, addressing a gap in existing quantum algorithms for this domain.
Methodology
The authors propose a quantum heavy-tailed mean estimator (QHTME) and an unbiased version (QUHTME) for estimating means of heavy-tailed random variables. They utilize quantum sampling oracles and apply a generalized multi-level Monte Carlo technique. The paper also includes theoretical proofs of quantum lower bounds for mean estimation, establishing the optimality of their methods in low-dimensional regimes.
Results
The proposed quantum estimators achieve lower query complexities than classical counterparts, with QNSGD requiring approximately O(βd Ο΅^(-5p-4)/(2p-2)) queries and QPSGD needing O(βd Ο΅^(-3p-2)/(2p-2) + Ο΅^(-2)) queries for convex functions. The results show significant improvements over classical lower bounds for both convex and non-convex optimization problems.
Implications
The findings suggest that quantum computing can effectively tackle stochastic optimization problems that are challenging for classical methods, particularly in scenarios with heavy-tailed noise. This could lead to advancements in various fields such as machine learning, finance, and operations research where such noise is prevalent.
Eliminating Propagation Delay: Attention-Based Spatial-Temporal Fusion Graph Convolution Network for Traffic Flow Prediction
Graph Learning
Time Series
Optimization
- Introduces A-STFGCN to address propagation delays in traffic flow prediction.
- Utilizes a spatial-temporal fusion block to enhance feature extraction.
- Implements a multi-head self-attention mechanism for temporal characteristics.
- Achieves superior performance on real-world datasets compared to existing methods.
Read more
Eliminating Propagation Delay: Attention-Based Spatial-Temporal Fusion Graph Convolution Network for Traffic Flow Prediction
Summary
This paper addresses the challenges in traffic flow prediction using graph convolutional networks (GCNs), specifically focusing on the limitations of existing models that overlook the propagation delays between adjacent nodes and often involve complex structures that increase computational time. The authors propose a novel model called the Attention-Based Spatial-Temporal Fusion Graph Convolution Network (A-STFGCN), which incorporates a spatial-temporal fusion block designed to effectively extract spatial-temporal feature correlations while eliminating propagation delay errors. The model employs a multi-head self-attention mechanism to capture both long-term and short-term temporal characteristics of traffic data. The proposed method is evaluated through extensive experiments on five real-world datasets, demonstrating superior performance compared to eight baseline methods in terms of accuracy, computational efficiency, and data utilization. The findings suggest that A-STFGCN can significantly enhance traffic flow prediction, contributing to improved urban mobility and transportation system optimization.
Methodology
The authors developed the A-STFGCN model, which integrates a spatial-temporal fusion block to capture spatial-temporal correlations while addressing propagation delays. The model employs a multi-head self-attention mechanism based on a mask matrix to effectively learn both long-term and short-term dependencies in traffic data. The performance of A-STFGCN was evaluated against eight baseline models using five real-world traffic datasets.
Results
The A-STFGCN model outperformed all baseline methods in terms of prediction accuracy and computational efficiency. The experiments demonstrated that the model effectively captures the dynamic spatial-temporal characteristics of traffic flow, leading to improved predictions.
Implications
The findings of this research have significant implications for the development of Intelligent Transportation Systems (ITS), as accurate traffic flow predictions can enhance urban mobility, optimize transportation systems, and facilitate better traffic management strategies.
Reinforcement Learning for Code Optimization
Reinforcement Learning
Optimization
- Reinforcement learning can be effectively adapted for code optimization by addressing measurement noise and reward sparsity.
- The DMC-Optim framework introduces a structured approach to testing, reward composition, and model adaptation.
- Significant improvements in optimization performance were achieved, with notable gains in pass rates for both correctness and speed.
- The methodology preserves correctness while enhancing execution speed, demonstrating the feasibility of combining these objectives in RL.
Read more
Reinforcement Learning for Code Optimization
Summary
This paper addresses the challenges of applying reinforcement learning (RL) to code optimization, particularly in achieving both correctness and execution speed. The authors identify that while RL has been effective for code correctness, extending it to optimization is complicated by issues such as measurement noise, reward sparsity, and instability in the optimization process. To tackle these challenges, the authors propose a structured approach that involves three key stages: enhancing the testing environment with large optimization tests and a calibrated sandbox, refining the reward mechanism to effectively combine correctness and speed, and adapting the learning model to cope with the noisier reward signals associated with execution time. The proposed framework, DMC-Optim, demonstrates significant improvements in optimization performance, achieving higher pass rates for code correctness and speed compared to traditional methods. The results indicate that the strongest configurations lead to substantial relative improvements in performance metrics, particularly at stricter evaluation percentiles, while maintaining correctness scores.
Methodology
The authors developed DMC-Optim, which consists of a three-stage process: (1) creating a testing environment with large optimization tests and a calibrated sandbox, (2) designing a reward system that integrates correctness and execution speed, and (3) modifying the learning model to handle the challenges of sparser and noisier reward signals. An offline RL simulator was used to evaluate different configurations before real training was conducted.
Results
The strongest optimization-aware configurations improved the pass@1 rate from 18.0% to 31.3% for Qwen 2.5 7B and from 30.7% to 50.4% for CWM 32B. At stricter percentiles, such as top-30%, CWM 32B achieved a 125% relative improvement. Robust optimization RL demonstrated up to 100%-200% improvement over standard RLVR in degraded conditions, and CWM 32B won 83% of median-sample speed comparisons against standard RLVR.
Implications
The findings suggest that RL can be effectively utilized for code optimization, potentially leading to more efficient code generation tools. This could have significant applications in software development, competitive programming, and automated code review processes, where both correctness and performance are critical.
Score-Based Stabilization for Time-Dependent Problems
Theory
Generative Models
Optimization
- Introduction of a two-stage stabilization framework that integrates a learned score-based correction into time-stepping schemes.
- The score correction effectively mitigates nonphysical behaviors and instabilities in numerical simulations of PDEs.
- Demonstration of conditional manifold stability, allowing for larger timesteps without traditional restrictions.
- Validation of the approach across multiple PDEs, showcasing its versatility and robustness.
Read more
Score-Based Stabilization for Time-Dependent Problems
Summary
This paper introduces a novel score-based stabilization framework for the numerical simulation of partial differential equations (PDEs). The proposed method utilizes a learned score model to define a stabilization operator that is applied to provisional numerical updates, enhancing standard time-stepping schemes. This operator ensures that numerical solutions adhere to the manifold of physically admissible states, thereby improving stability and preserving essential physical properties. The authors demonstrate that the score correction acts as a contraction toward this manifold, providing basin-conditional stability. Through numerical experiments on various PDEs, including Advection, Kortewegβde Vries (KdV), Nonlinear SchrΓΆdinger (NLS), and Burgersβ equations, the framework shows enhanced robustness, suppression of nonphysical instabilities, and maintenance of qualitative dynamics. The approach allows for larger timesteps without compromising the structural integrity of the solutions, making it a significant advancement in the field of numerical PDE simulation.
Methodology
The authors propose a two-stage numerical method where the first stage involves updating the solution using standard time-stepping techniques. In the second stage, a learned score-based correction is applied to stabilize the provisional update, ensuring it remains within the manifold of admissible states. The score function, representing the gradient of the log-density of a target distribution, is learned offline and used to inform the stabilization process.
Results
The numerical experiments conducted on various PDEs demonstrate that the score-based stabilization framework significantly improves the robustness of simulations, effectively suppressing nonphysical instabilities and preserving the qualitative dynamics of the solutions. The method allows for larger timesteps while maintaining the structural properties of the solutions, thus enhancing computational efficiency.
Implications
This score-based stabilization framework has the potential to revolutionize numerical simulations of complex physical systems, particularly in fields such as weather prediction and fluid dynamics, where maintaining stability and physical accuracy over long time horizons is crucial. The ability to reuse the learned score model across different initial conditions can lead to more efficient and reliable simulations.
Detecting CSAM Text-to-Image LoRAs From Weights
Generative Models
Computer Vision
Interpretability
- Introduces u1, a representation for detecting harmful LoRAs from their weights.
- Demonstrates that metadata for LoRAs is often incomplete and unreliable.
- Shows that the method can generalize across different base models and datasets.
- Proves robustness to noise and precision reduction in weight analysis.
Read more
Detecting CSAM Text-to-Image LoRAs From Weights
Summary
This paper addresses the challenge of detecting harmful low-rank adaptation (LoRA) models that can generate child sexual abuse material (CSAM) without relying on potentially misleading metadata or generating harmful outputs. The authors propose a novel approach that analyzes the weights of LoRA models to identify their training objectives. They introduce a representation called u1, derived from the top-left singular vectors of LoRA updates, which serves as a compact fingerprint of the model's learned changes. Using human-subject age as a benign proxy for CSAM, the authors demonstrate that u1 can effectively identify the training subject across different base models while abstaining from unrelated benign content. The method is robust against noise and rescaling, suggesting that harmful LoRAs can be screened directly from their weights, providing a safer alternative to existing moderation techniques.
Methodology
The authors extract the top-left singular vector from the weight updates of LoRA models to create the u1 representation. This method allows for inference-free screening of LoRAs by analyzing their weights directly, without the need for prompts or generated images. The study uses human-subject age as a benign proxy to evaluate the effectiveness of the detection method.
Results
The results indicate that the u1 representation can accurately identify the training subjects of LoRAs, achieving a high area under the receiver operating characteristic curve (AUROC) score of 0.998 on the age axis. The method successfully abstains from flagging unrelated benign adapters, demonstrating its effectiveness in distinguishing harmful content from benign content.
Implications
This research has significant implications for AI safety and child protection, as it provides a method to detect harmful models without the ethical and legal concerns associated with inspecting generated outputs. It could enhance moderation practices on public model-hosting platforms and contribute to the development of safer AI technologies.
From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
Generative Models
- Introduction of a diffusion-based generative model for air quality reconstruction.
- Joint modeling of multiple pollutants to capture inter-pollutant correlations.
- Evaluation on real-world data demonstrates high accuracy and generalization.
- Data augmentation methods facilitate transferability to real-world observations.
Read more
From Deterministic to Generative Deep Learning for Urban Air Quality Reconstruction from Sparse Observations
Summary
This paper addresses the challenge of reconstructing urban air quality from sparse observations of key pollutants (NO2, O3, PM2.5, PM10) using deep learning techniques. Traditional deterministic models face limitations due to uncertainties in emission inventories and meteorological data, which complicate real-time monitoring. The authors propose a novel diffusion-based generative framework that allows for multi-pollutant reconstruction directly from sensor observations, bypassing the need for chemistry-transport models during inference. The methodology involves training the model on simulated full-field data and evaluating its performance on real-world observations from Paris. The study introduces data augmentation techniques to enhance model generalization to real-world scenarios without retraining. Results indicate that the generative models achieve high structural similarity with simulated data and produce realistic spatial patterns in real-world observations, demonstrating robustness against noise and spatial variability. This work represents a significant advancement in the application of generative AI for environmental monitoring, highlighting its potential for reliable deployment in air quality reconstruction tasks.
Methodology
The authors employed a diffusion-based generative model trained on simulated full-field air quality data. The model was evaluated using real-world sensor data from Paris, with Voronoi tessellations used to address spatial representation issues. Data augmentation techniques were introduced to enhance model generalization without retraining.
Results
The generative models achieved high structural similarity with simulated validation data and produced realistic spatial patterns in real-world observations. The evaluation showed that the models generalize well to out-of-distribution data, indicating their robustness against noise and spatial variability.
Implications
The findings suggest that generative deep learning models can significantly improve urban air quality monitoring and reconstruction, providing valuable insights for public health decision-making and policy formulation. The ability to operate in real-time with sparse data could enhance the effectiveness of environmental monitoring systems.
Stable FP4 Training via Transposition-Invariant Block Quantization
Large Language Models
Efficient ML
Optimization
- Identified transposition-induced scale inconsistency as a major source of instability in FP4 training.
- Proposed a transposition-invariant FP4 quantization framework using 2D block quantization.
- Combined FP4 linear layers with MXFP8 attention for practical mixed-precision training.
- Achieved stable training for models up to 30 billion parameters with minimal performance degradation.
Read more
Stable FP4 Training via Transposition-Invariant Block Quantization
Summary
This paper addresses the challenges of training large language models (LLMs) using 4-bit floating point (FP4) precision, which has been hindered by instability during optimization. The authors identify a critical source of this instability in existing microscaling approaches, specifically the scale inconsistency caused by tensor transposition during forward and backward passes. Traditional 1D block quantization methods assign different scaling factors to the same values after transposition, leading to biased gradients and unstable training. To overcome this issue, the authors propose a novel low-precision training framework utilizing 2D block FP4 quantization, which maintains transposition-invariant scaling and ensures consistency between forward and backward computations. Additionally, they incorporate truncation-free scaling and stochastic rounding to minimize quantization error and preserve unbiased gradients. For the attention mechanisms, MXFP8 quantization is employed for query and key projections, resulting in a mixed-precision design. The proposed method is evaluated on dense LLMs with up to 7 billion parameters and a 30 billion parameter Mixture-of-Experts model, trained on 100 billion tokens. The results demonstrate that the approach achieves stable end-to-end FP4 training, closely matching the performance of BF16 with less than 1.3% degradation in perplexity and downstream accuracy, thus providing a viable pathway for more efficient LLM training.
Methodology
The authors developed a transposition-invariant FP4 training framework based on 2D block quantization, which ensures consistent scaling across forward and backward passes. They combined this with truncation-free scaling and stochastic rounding to control quantization errors. For attention mechanisms, MXFP8 quantization was utilized for query and key projections, creating a mixed-precision approach.
Results
The proposed method demonstrated stable end-to-end FP4 training across various model sizes, achieving performance comparable to BF16 with less than 1.3% degradation in perplexity and downstream accuracy. This indicates the effectiveness of enforcing forward-backward scaling consistency in enabling practical FP4 training.
Implications
The findings suggest that enforcing scaling consistency can significantly enhance the stability of low-precision training, paving the way for more efficient training of large language models, which is crucial given the rising computational and energy costs associated with LLMs.
Local Regularization Does Not Characterize Multiclass PAC Learnability
Theory
- Local regularization does not characterize multiclass PAC learnability.
- A countable hypothesis class with low complexity cannot be learned by local regularizers.
- The construction involves tournament structures that introduce complexity through cyclic relationships.
- The findings challenge existing assumptions about the effectiveness of local regularization in multiclass settings.
Read more
Local Regularization Does Not Characterize Multiclass PAC Learnability
Summary
In this paper, Eric Hou investigates the role of local regularization in characterizing multiclass PAC learnability. Local regularization assigns a score to each hypothesis based on the test point, and predictions are made using the hypothesis with the minimum score that is consistent with the sample. The author provides a counterexample to the claim made by Asilis et al. that local regularization can characterize multiclass PAC learnability. Specifically, Hou constructs a countable class of hypotheses with a DanielyβShalev-Shwartz dimension of at most two, which has a realizable PAC sample complexity of O(1/Ξ΅ log(1/Ξ΄)), but cannot be learned by any local regularizer. The hypotheses are represented as edges of complete graphs, and the instances are structured as tournaments. The findings reveal that cyclic triangles in the tournament structure lead to sufficient inversions, resulting in a constant population error regardless of sample size. This work highlights the limitations of local regularization in the context of multiclass learning, suggesting that alternative approaches may be necessary for effective learning in this setting.
Methodology
The author constructs a specific countable hypothesis class based on the edges of complete graphs and uses tournament structures to demonstrate the limitations of local regularization. The analysis involves deriving sample complexity bounds and exploring the implications of cyclic relationships within the tournament framework.
Results
The main result establishes that there exists a countable multiclass hypothesis class H with a realizable PAC sample complexity of O(1/Ξ΅ log(1/Ξ΄)) and a DanielyβShalev-Shwartz dimension of at most two, which cannot be learned by any local regularizer. This counterexample illustrates the failure of local regularization to capture the complexities of multiclass PAC learnability.
Implications
The findings suggest that local regularization may not be a sufficient learning principle for multiclass problems, indicating a need for alternative methods that can effectively handle the complexities inherent in such settings. This could influence future research directions in statistical learning theory and multiclass classification algorithms.
Random Forest-Based Prediction of Bone Volume Fraction and Fracture Position from S-Parameters
Theory
Efficient ML
- Introduces a Random Forest model for predicting bone volume fraction and fracture position.
- Utilizes a nine-antenna microwave scanning system to gather S-parameter data.
- Demonstrates the effectiveness of microwave sensing as a non-invasive alternative to traditional imaging techniques.
- Validates the method through both synthetic simulations and physical experiments with bone-mimicking phantoms.
Read more
Random Forest-Based Prediction of Bone Volume Fraction and Fracture Position from S-Parameters
Summary
This paper presents a novel approach for predicting bone volume fraction (BVF) and fracture position using a Random Forest model based on multichannel S-parameters obtained from a microwave sensing system. The authors designed a nine-antenna microwave scanning system to collect S-parameter data from bone-mimicking phantoms. The study highlights the limitations of traditional medical imaging techniques, such as X-rays and CT scans, which are often costly and involve harmful radiation. The proposed method leverages the advantages of microwave sensing, which is non-ionizing and cost-effective. The authors conducted extensive simulations and physical experiments, generating a dataset of 4,500 samples that include various BVF values and fracture scenarios. The Random Forest model was trained to predict both BVF and fracture position, demonstrating its capability to handle multi-output regression tasks effectively. The results indicate that the proposed method is valid and has the potential to improve bone health diagnostics.
Methodology
The methodology involves designing a microwave sensing system with a nine-antenna array to acquire multichannel S-parameters from bone-mimicking phantoms. The data is processed using a Random Forest model, which predicts BVF and fracture position based on the S-parameters. The model is trained on a dataset generated from simulations that include various BVF values and fracture scenarios.
Results
The Random Forest model successfully predicted bone volume fraction and fracture positions, demonstrating high accuracy in both synthetic and experimental validations. The study confirms the model's ability to handle complex nonlinear relationships and its robustness against overfitting.
Implications
The findings suggest that microwave sensing combined with machine learning can provide a non-invasive, cost-effective method for diagnosing bone health issues, potentially transforming clinical practices in osteoporosis and fracture management.
LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models
Time Series
Large Language Models
Optimization
- LAFP separates numerical generation and contextual reasoning, enhancing time-series forecasting accuracy.
- The framework uses Monte Carlo Tree Search (MCTS) to explore candidate trajectories effectively.
- LAFP shows consistent improvements over context-blind TSFMs and direct LLM forecasters across multiple datasets.
- The approach is training-free, allowing for easy integration with existing TSFMs and LLMs.
Read more
LLM as Forecasting Planner: Training-Free Text Conditioning for Time-Series Foundation Models
Summary
This paper introduces LAFP (LLM As Forecasting Planner), a novel framework for text-conditioned time-series forecasting that leverages the strengths of Time-Series Foundation Models (TSFMs) and Large Language Models (LLMs) without requiring additional training. The authors argue that while TSFMs excel at generating numerical forecasts based on historical data, they struggle to incorporate contextual information from text, which is crucial for accurate predictions. LAFP addresses this by treating forecasting as a planning problem, where the TSFM serves as a simulator to generate candidate numerical continuations, and the LLM acts as a policy and value function to guide the selection and evaluation of these candidates. The framework employs Monte Carlo Tree Search (MCTS) to explore the forecast horizon, allowing for the combination of promising trajectories. The authors demonstrate the effectiveness of LAFP through extensive experiments on two datasets, Context-is-Key and Time-MMD, using various TSFM and LLM combinations. The results indicate that LAFP consistently outperforms traditional context-blind TSFMs and direct LLM forecasting approaches, showcasing its potential for enhancing time-series predictions by integrating textual context.
Methodology
The LAFP framework formulates text-conditioned time-series forecasting as a planning problem, utilizing a frozen TSFM to generate numerical forecasts and a frozen LLM to rank and evaluate these forecasts. MCTS is employed to explore candidate trajectories, with the LLM providing policy guidance and value assessments without requiring retraining.
Results
Experiments conducted on the Context-is-Key and Time-MMD datasets demonstrate that LAFP consistently improves forecasting accuracy compared to context-blind TSFMs and direct LLM forecasting methods. The framework also exhibits robustness across different model configurations and search budgets.
Implications
The LAFP framework has significant implications for various domains requiring time-series forecasting, such as finance, energy, and public health. By effectively integrating textual context into numerical predictions, it enhances decision-making capabilities and allows for more informed responses to future events.
Rethinking Likelihood distributions: Student's t Likelihood Boosts Bayesian Neural Network Performance
Theory
- Student's t likelihood distribution outperforms Gaussian likelihood in BNNs.
- The choice of likelihood distribution significantly affects predictive accuracy.
- Student's t distribution can lead to shorter training times while being easy to implement.
- The study fills a gap in the literature regarding the exploration of likelihood distributions in BNNs.
Read more
Rethinking Likelihood distributions: Student's t Likelihood Boosts Bayesian Neural Network Performance
Summary
This paper investigates the impact of likelihood distribution choices on the performance of Bayesian Neural Networks (BNNs), specifically focusing on the commonly used Gaussian distribution versus alternative distributions. The authors highlight that while Gaussian distributions are prevalent due to their computational ease, they may lead to model misspecification and suboptimal performance. The study systematically evaluates various likelihood distributions, including skewed and heavy-tailed options, across regression tasks using multilayer perceptrons (MLPs) on both artificial and real-world datasets. The findings reveal that the Student's t distribution consistently outperforms the Gaussian likelihood, enhancing predictive accuracy and, in some cases, reducing training times. The research emphasizes the importance of selecting appropriate likelihood distributions in BNNs to improve uncertainty quantification and model reliability, particularly in high-stakes applications where uncertainty estimates are crucial.
Methodology
The authors conducted controlled experiments comparing various likelihood distributions, including Gaussian and Student's t, across different regression tasks. They utilized multilayer perceptrons (MLPs) to evaluate the performance of these distributions on both synthetic and real-world datasets, measuring predictive accuracy and training times.
Results
The results demonstrated that the Student's t distribution consistently provided better predictive performance than the Gaussian distribution across all tested datasets and MLP architectures. Additionally, in certain scenarios, the Student's t distribution resulted in shorter training times, indicating its efficiency and effectiveness as a likelihood choice in BNNs.
Implications
The findings suggest that adopting the Student's t likelihood distribution in Bayesian Neural Networks can enhance model performance and reliability, particularly in applications requiring robust uncertainty quantification. This has significant implications for fields such as autonomous driving, finance, and energy systems, where accurate uncertainty estimates are critical for decision-making.
GAUGE: Grading Agent-Built Financial Models Without a Golden Answer
Large Language Models
Optimization
Theory
- GAUGE provides a more realistic benchmark for evaluating financial models by considering multiple analyst outputs instead of a single expert reference.
- The study reveals significant discrepancies in financial model outputs among professional analysts, questioning the validity of point-tolerance grading.
- GAUGE incorporates a multi-layered scoring system that evaluates both mechanical correctness and judgment-based facets.
- Results show that current AI agents perform better in model construction than in valuation judgment compared to human analysts.
Read more
GAUGE: Grading Agent-Built Financial Models Without a Golden Answer
Summary
The paper introduces GAUGE, a novel benchmark for evaluating financial models built by AI agents, addressing the limitations of existing benchmarks that rely on a single expert reference for grading outputs. The authors argue that financial models often yield multiple reasonable outputs due to the subjective nature of certain inputs, such as forecasts and discount rates. They present empirical evidence showing that using a single reference can unfairly penalize agents for diverging from expert opinions, as demonstrated by a study involving 137 workbooks from 65 companies, where the median score was low and discrepancies were common. GAUGE is constructed from 1,001 analyst workbooks and includes a comprehensive evaluation set of 196 tasks, employing a scoring system that incorporates mechanical checks, judgment facets, and validity gates. The benchmark allows for a more nuanced evaluation of agent performance by comparing outputs against a range of observed analyst practices rather than a single point estimate. The results indicate that while agents excel in model construction, they lag in valuation judgment compared to human analysts, highlighting the need for further development in AI capabilities for financial decision-making.
Methodology
The authors developed GAUGE by analyzing 1,001 analyst-built financial workbooks and creating a 196-task evaluation set. The scoring system includes mechanical checks, judgment facets, and validity gates to assess the quality of agent-built models against observed analyst practices.
Results
In a study involving 24 agents and 1,011 scored generations, the best-performing agent scored 53.4, which is above the average score of finance students (43.2) but below that of senior analysts (88.3) and junior analysts (66.0). The agents excelled in mechanical checks (93% pass rate) but performed poorly in judgment facets (78% pass rate).
Implications
The findings suggest that AI agents need further development in valuation judgment to match human analysts' capabilities. GAUGE can serve as a foundation for future research in AI financial modeling and evaluation, promoting more robust benchmarks in the field.
K-Survival Means
Optimization
Theory
Time Series
- K-SurvMeans explicitly incorporates survival outcomes into the clustering process.
- The method uses Particle Swarm Optimization to solve the non-differentiable optimization problem.
- It operates in a learned low-dimensional latent space to improve cluster separation and efficiency.
- Experiments show superior performance in cluster separation compared to existing methods.
Read more
K-Survival Means
Summary
This paper introduces K-SurvMeans, a novel extension of the K-Means clustering algorithm specifically designed for survival data. The primary innovation of K-SurvMeans is its explicit incorporation of survival outcomes into the clustering process, which optimizes cluster centers to maximize pairwise survival differences between clusters. This approach addresses the limitations of traditional clustering methods that do not consider survival information, leading to clusters that may not effectively stratify risk. The optimization problem posed by K-SurvMeans is non-differentiable, prompting the use of the Particle Swarm Optimization algorithm to find optimal cluster centers. To enhance flexibility and combat the curse of dimensionality, the framework is extended to operate in a learned low-dimensional latent space, improving cluster separation and optimization efficiency. Experimental results on multiple benchmark survival datasets demonstrate that K-SurvMeans consistently produces clusters with better separation in survival distributions compared to existing deep learning-based survival clustering methods.
Methodology
K-SurvMeans extends the traditional K-Means algorithm by integrating survival outcomes into the clustering objective function. It employs Particle Swarm Optimization to handle the non-differentiable nature of the optimization problem and utilizes dimensionality reduction techniques to operate in a low-dimensional latent space.
Results
The experiments conducted on publicly available benchmark survival datasets indicate that K-SurvMeans achieves significantly improved cluster separation in terms of survival distributions when compared to existing deep learning-based survival clustering methods.
Implications
K-SurvMeans has potential applications in healthcare for patient stratification, enabling clinicians to identify distinct patient groups based on survival profiles. This can support personalized treatment planning and resource allocation. Additionally, the method can be applied in various fields where survival analysis is relevant, such as predictive maintenance and financial risk assessment.
FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting
NLP
Large Language Models
Multimodal
- Introduces a multimodal RAG framework for financial forecasting that incorporates uncertainty calibration.
- Utilizes a point-in-time retrieval mechanism to ensure only relevant, publicly available data is used.
- Develops a hybrid uncertainty score to guide selective predictions and abstentions.
- Demonstrates that calibrated abstention can reduce errors and drawdowns in financial predictions.
Read more
FinAbstain: Uncertainty-Calibrated Multimodal RAG for Selective Financial Forecasting
Summary
The paper introduces FinAbstain, a framework designed to enhance financial forecasting by addressing the overconfidence of large language models (LLMs) when evidence is sparse or contradictory. FinAbstain employs a retrieval-augmented generation (RAG) approach that ensures only publicly available information at the forecast timestamp is utilized. The framework integrates multiple agents that assess evidence from various modalities, including financial filings, news, and technical indicators, while also exposing disagreements among these sources. A novel hybrid uncertainty score is developed to calibrate predictions and determine when to abstain from making forecasts. The system's controller predicts market outcomes (bullish, bearish, or neutral) only when uncertainty is below a validated threshold; otherwise, it abstains or requests further evidence. The evaluation of FinAbstain includes metrics for accuracy, calibration, and risk coverage, demonstrating that calibrated abstention can lead to lower selective errors and drawdowns. The contributions of this work include a time-safe architecture, a composite uncertainty formulation, and a reproducible evaluation blueprint for selective financial forecasting.
Methodology
The methodology involves a point-in-time multimodal retrieval system that aggregates evidence from various financial sources. It employs multiple agents to assess the evidence, calculates a hybrid uncertainty score, and uses a controller to make selective predictions based on uncertainty thresholds. The evaluation is conducted using simulated results to ensure transparency and reproducibility.
Results
The results indicate that the FinAbstain framework effectively reduces selective errors and drawdowns in financial forecasting by implementing calibrated abstention. The evaluation metrics show improved accuracy and risk coverage compared to traditional forecasting methods.
Implications
The implications of this research suggest that financial forecasting systems can be made more reliable by integrating uncertainty calibration and selective prediction strategies. This approach can enhance decision-making in finance, particularly in environments characterized by noisy and contradictory information.
Multi-Agent Privacy Game in Federated Learning: A Unified Mean-Field View
Federated Learning
Theory
Optimization
- Introduces a mean-field privacy game framework for privacy-preserving federated learning.
- Achieves tractable equilibrium for arbitrarily many clients while accommodating heterogeneous privacy preferences.
- Recovers existing privacy-preserving methods as special cases of the proposed framework.
- Demonstrates effective privacy-utility trade-offs through empirical experiments.
Read more
Multi-Agent Privacy Game in Federated Learning: A Unified Mean-Field View
Summary
This paper addresses the privacy concerns in Federated Learning (FL), where model updates can potentially leak information about local datasets. The authors propose a novel framework that combines existing privacy-preserving methods by formulating the problem as a mean-field privacy game. In this framework, each client strategically selects its privacy budget while interacting with the population through a single mean-field statistic. This approach allows for a tractable equilibrium even with a large number of clients and accommodates heterogeneous privacy preferences. The framework recovers existing methods as special cases and provides an exponentially decaying privacy guarantee. The authors validate their approach through experiments on various tasks, demonstrating that it achieves a favorable privacy-utility trade-off compared to traditional methods.
Methodology
The authors formalize privacy-preserving FL as a mean-field stochastic differential game, defining key components such as state, type, control, and dynamics. They prove the existence of a mean-field Nash equilibrium and certify differential privacy at equilibrium using log-Sobolev contraction properties. The framework is implemented through a ParticleβSinkhorn JKO step, allowing for comparisons across different configurations.
Results
The proposed mean-field privacy game framework successfully balances privacy and utility, achieving results comparable to the entropic privacy baseline while providing personalized privacy guarantees. Experiments on quadratic regression, logistic regression, and MNIST datasets show that the framework outperforms traditional methods in terms of privacy-utility trade-offs.
Implications
This work has significant implications for enhancing privacy in federated learning systems, particularly in applications where data sensitivity is paramount. The proposed framework can be applied to various domains requiring collaborative learning while ensuring data privacy, such as healthcare and finance.
Contextual Deconvolution for Variance-Stable Demand Sensing: Kernel-Modulated Operators in Promotional Retail
Time Series
- Introduction of Contextual Deconvolution (CD) for improved demand sensing in promotional retail.
- CD utilizes a kernel-modulated banded operator to separate transient shocks from structural demand.
- The method is scalable and does not require per-SKU training, making it suitable for large retail catalogs.
- Empirical results show significant reductions in safety stock and order variance compared to traditional methods.
Read more
Contextual Deconvolution for Variance-Stable Demand Sensing: Kernel-Modulated Operators in Promotional Retail
Summary
This paper addresses the challenges of demand sensing in promotional retail, where traditional machine learning forecasts often lead to excess operational volatility. The author introduces Contextual Deconvolution (CD), a two-stage estimator that separates transient demand shocks from structural baselines using a kernel-modulated banded operator. This operator encodes promotional carryover without requiring per-SKU training, making it scalable for large catalogs. The methodology involves a two-stage pipeline that combines robust detrending with distributed lag regression, allowing for effective kernel recovery. The empirical evaluation on 30,490 SKUs from the M5 competition and 2,845 items from Favorita demonstrates that CD significantly reduces safety stock and order variance while maintaining operational stability. The results indicate that CD achieves a lower Variance Ratio and less safety stock compared to traditional methods, although it may under-provision during promotional spikes. The study highlights the importance of structural decomposition in improving inventory management and operational efficiency in retail environments.
Methodology
The methodology involves a two-stage estimation pipeline that includes Robust Huber Detrending for baseline estimation and Distributed Lag Regression for kernel recovery. The kernel-modulated operator is designed to encode promotional dynamics and is regularized through hierarchical partial pooling to enhance scalability across SKUs.
Results
The evaluation of CD on M5 and Favorita datasets shows a median Variance Ratio of 0.0007 compared to 0.0135 for Tuned XGBoost, and a substantial reduction in safety stock (0.16x on M5 and 0.38x on Favorita). While accuracy gains are modest at the median, CD achieves the lowest cross-sectional dispersion of per-SKU error and ranks first in mis-forecasts across multiple independent draws.
Implications
The findings suggest that CD can significantly enhance inventory management practices in promotional retail by reducing operational volatility and improving demand forecasting accuracy. This has potential applications in supply chain optimization and inventory control strategies.
Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs
Graph Learning
Time Series
- Introduces a contrastive representation learning framework for longitudinal disease trajectories.
- Models disease trajectories as temporal graphs with nodes and edges representing observations and relationships.
- Utilizes structure-aware random walks for effective contrastive learning.
- Achieves improved clustering performance on real-world biomedical datasets.
Read more
Contrastive Representation Learning of Longitudinal Disease Trajectories on Temporal Graphs
Summary
This paper addresses the challenge of understanding disease trajectories from longitudinal clinical data, which often exhibit complex temporal dynamics and involve heterogeneous patient cohorts. The author proposes a novel contrastive representation learning framework called RankWalk, which models multivariate disease trajectories as temporal graphs. In this framework, nodes represent patient observations over time, while edges capture both temporal continuity and structural similarity between trajectories. The methodology employs structure-aware random walks to guide contrastive learning, enabling the generation of embeddings that maintain temporal context and trajectory topology. The learned representations facilitate robust clustering of patients with similar disease progression patterns, revealing latent structures within longitudinal data. The framework is evaluated through two simulation studies and four real-world biomedical datasets, demonstrating consistent improvements in clustering performance compared to existing longitudinal clustering methods.
Methodology
The proposed RankWalk framework constructs a heterogeneous graph from longitudinal observations, combining temporal edges for consecutive observations of the same subject with similarity edges for subjects with comparable characteristics. A weighted contrastive objective is employed to learn node representations that emphasize informative positive pairs, facilitating unsupervised longitudinal clustering.
Results
The experimental results indicate that the RankWalk framework consistently outperforms traditional longitudinal clustering approaches across various datasets, demonstrating enhanced clustering performance and the ability to uncover meaningful patterns in disease progression.
Implications
The findings suggest that the RankWalk framework can significantly improve patient stratification and disease subtyping in clinical settings, offering a more nuanced understanding of disease trajectories and potentially informing personalized treatment strategies.
Mind the Missing Split: Resolving Feature Heterogeneity in Swarm Learning with Random Forests
Federated Learning
- Addresses feature heterogeneity in Swarm Learning with Random Forests.
- Proposes inference-time strategies to resolve missing splits without restricting feature training.
- Evaluates methods on nine datasets, showing improved performance over traditional approaches.
- Highlights the importance of decentralized learning in maintaining data privacy.
Read more
Mind the Missing Split: Resolving Feature Heterogeneity in Swarm Learning with Random Forests
Summary
This paper addresses the challenge of feature heterogeneity in Swarm Learning (SL), a decentralized collaborative learning framework where multiple organizations train a shared model without central coordination or direct data sharing. In typical horizontal SL, datasets across sites are assumed to share the same feature set, but in practice, they often have partially overlapping features due to differences in measurements and protocols. This heterogeneity can lead to issues in machine learning algorithms, particularly Random Forests, where encountering a split on a feature not available at a site can complicate inference. The authors propose several deterministic and probabilistic strategies to handle missing splits during inference, allowing for collaborative learning without restricting training to the intersection of features. The proposed methods are evaluated on nine datasets, demonstrating superior performance compared to both the intersection baseline and locally trained models across various scenarios.
Methodology
The authors introduce several deterministic and probabilistic strategies for handling missing splits in Random Forests during inference. These methods allow for the use of site-specific variables without requiring all sites to share the same feature set. The strategies are tested on multiple datasets to assess their effectiveness.
Results
The proposed methods outperformed both the intersection baseline and locally trained models across a broad range of scenarios, indicating their effectiveness in resolving feature heterogeneity in Swarm Learning.
Implications
The findings suggest that it is possible to conduct collaborative learning in decentralized environments while accommodating feature heterogeneity, which is particularly relevant in fields like healthcare where data privacy is crucial. This can enhance model generalizability and accuracy by leveraging diverse datasets.
Charging Phase Health Indicators for Battery State-of-Health Estimation: A Systematic Comparison of CC, CV, and Combined Approaches under Cross-Battery Validation
Time Series
Interpretability
Efficient ML
- Combined CC+CV indicators yield the highest SOH estimation accuracy.
- A significant performance gap exists between standard cross-validation and LOBO validation.
- SHAP analysis identifies the CV-to-CC time ratio as a critical indicator for SOH estimation.
- The study provides practical guidelines for indicator selection under data constraints.
Read more
Charging Phase Health Indicators for Battery State-of-Health Estimation: A Systematic Comparison of CC, CV, and Combined Approaches under Cross-Battery Validation
Summary
This paper presents a systematic comparison of health indicators derived from constant-current (CC) and constant-voltage (CV) charging phases for estimating the state-of-health (SOH) of lithium-ion batteries. The authors address the inadequacy of existing studies that primarily focus on CC indicators by evaluating the effectiveness of CC-only, CV-only, and combined CC+CV indicators using rigorous Leave-One-Battery-Out (LOBO) validation on the NASA battery aging dataset. The findings reveal that the combined approach significantly outperforms the individual methods, achieving a coefficient of determination (RΒ²) of 0.874, indicating that CC and CV phases provide complementary information regarding battery degradation. Furthermore, the study highlights a substantial performance gap (119% increase in RMSE) between standard cross-validation and LOBO validation, suggesting that conventional evaluation methods may overestimate the accuracy of SOH estimation. The authors also employ SHAP analysis to identify the CV-to-CC time ratio as the most influential indicator, offering practical guidance for selecting indicators in battery management systems. Overall, this work contributes to the field by providing insights into effective SOH estimation methods and emphasizing the importance of realistic validation techniques.
Methodology
The authors conducted a systematic comparison of health indicators derived from CC and CV charging phases using the NASA battery aging dataset. They employed Leave-One-Battery-Out (LOBO) validation to rigorously evaluate the performance of CC-only, CV-only, and combined indicator sets. Additionally, SHAP analysis was utilized to interpret the influence of various indicators on SOH estimation.
Results
The combined CC+CV approach achieved the best performance with an RΒ² value of 0.874. The study found a 119% increase in RMSE when comparing LOBO validation to standard 5-fold cross-validation, indicating that traditional evaluation methods may overstate the accuracy of SOH estimation.
Implications
The findings suggest that incorporating both CC and CV indicators can enhance the reliability of SOH estimation in battery management systems. The study's insights into validation techniques can help improve the robustness of SOH estimation methods, particularly in real-world applications where data and computational resources may be limited.
Calibrated Partial Resets: Preventing Policy Collapse in Continual Reinforcement Learning
Reinforcement Learning
Optimization
Robotics
- CPR is introduced as a solution to prevent policy collapse in continual RL by selectively adjusting low-utility neurons.
- The method maintains a balance between network plasticity and stability, avoiding the pitfalls of binary resets.
- CPR outperforms existing methods in multiple benchmarks, demonstrating its effectiveness in long-term training scenarios.
- Ablation studies reveal a tunable trade-off between plasticity and peak performance, emphasizing the utility-scaled approach.
Read more
Calibrated Partial Resets: Preventing Policy Collapse in Continual Reinforcement Learning
Summary
This paper addresses the challenge of policy collapse in continual reinforcement learning (RL) due to the accumulation of dormant neurons and loss of expressivity in neural networks during non-stationary training. The authors propose a novel approach called Calibrated Partial Resets (CPR), which periodically adjusts low-utility neurons toward their initialization without fully reinitializing them. This method aims to maintain network plasticity while avoiding the instability associated with binary resets. CPR selectively applies a continuous adjustment based on each neuron's utility, allowing for smoother updates compared to traditional methods. The authors demonstrate that CPR effectively prevents policy collapse over extensive training periods and outperforms existing decay and reset methods across various benchmarks, highlighting its potential for enhancing continual learning in RL environments.
Methodology
The authors developed CPR, which periodically pulls low-utility neurons toward their initialization based on a utility score that reflects each neuron's contribution to learning. This utility score is computed using a layer-normalized exponential moving average of neuron activations. The method contrasts with binary reset techniques by providing a continuous adjustment that is both selective and smooth, thereby enhancing training stability in continual RL settings.
Results
The results show that CPR successfully avoids policy collapse over 400 million training steps in the SlipperyAnt environment, achieving the highest episodic return compared to other methods. It also demonstrates improved performance in SlipperyHumanoid, Continual MetaWorld, and Continual MinAtar benchmarks. The ablation studies confirm that the utility-scaled resets allow for a controlled trade-off between maintaining plasticity and achieving peak performance.
Implications
The findings suggest that CPR could significantly enhance the training of RL agents in non-stationary environments, making it a promising approach for applications in robotics and other domains requiring continual learning. By improving the stability and adaptability of neural networks, CPR may facilitate the development of more robust AI systems capable of learning over extended periods.
Data-Dependent Regret and Polyak Corrections for Constrained Online Convex Optimization
Optimization
Theory
- Introduces a refined regret analysis for constrained OCO using data-dependent quantities.
- Proposes AdaOGD-PFS, an adaptive-step-size algorithm that achieves improved regret bounds.
- Demonstrates significant performance improvements in experiments, with reductions in regret bounds by 38-43%.
- Identifies the Polyak correction term as a crucial factor in achieving tighter regret bounds.
Read more
Data-Dependent Regret and Polyak Corrections for Constrained Online Convex Optimization
Summary
This paper addresses the problem of constrained online convex optimization (OCO), where a learner aims to minimize regret against adversarially chosen convex costs while adhering to convex constraints. The author proposes an enhancement to the online gradient descent (OGD) algorithm by incorporating a Polyak feasibility step, which allows for efficient constraint satisfaction with minimal computational overhead. The analysis reveals that existing regret bounds can be tightened by using data-dependent quantities instead of worst-case estimates. Specifically, the paper introduces a refined regret bound that incorporates the observed gradient accumulation and a Polyak correction term, which accounts for the cumulative displacement caused by the feasibility projection. The proposed AdaOGD-PFS algorithm, an adaptive-step-size variant, achieves a regret bound of O(βGT), which is significantly tighter than the traditional O(GfβT) bound. Experimental results demonstrate a 38-43% improvement in bound performance across various constrained instances, highlighting the effectiveness of both data-dependent gradients and Polyak corrections in enhancing the algorithm's performance.
Methodology
The paper utilizes online gradient descent augmented with a Polyak feasibility step, which involves a closed-form half-space projection. The analysis employs a refined approach to regret bounds by replacing worst-case gradient norms with observed data-dependent sums and incorporating a Polyak correction term that captures cumulative displacement during feasibility projections.
Results
The refined regret bound shows that the total improvement is non-negative and decomposes into two independent sources. The proposed AdaOGD-PFS algorithm achieves a regret bound of O(βGT), which is significantly better than the traditional O(GfβT) bound, with experimental results confirming a 38-43% reduction in regret across various problem instances.
Implications
The findings have significant implications for safety-critical domains such as power systems, autonomous control, and clinical decision-making, where maintaining feasibility while minimizing regret is essential. The proposed methods can enhance decision-making processes in these areas by providing more reliable and efficient optimization strategies.
When Does Deep Representation Learning Help Single-Cell Clustering? A Sensitivity-Aware Diagnostic Benchmark for Biomedical AI Pipelines
Generative Models
Optimization
Efficient ML
- Deep representation learning can enhance clustering performance but is not universally superior to classical methods.
- The study identifies three distinct dataset regimes where different methods excel.
- Learning rate and latent dimensionality are critical hyperparameters influencing clustering outcomes.
- A dataset-aware and compute-conscious framework is proposed for optimizing biomedical AI pipelines.
Read more
When Does Deep Representation Learning Help Single-Cell Clustering? A Sensitivity-Aware Diagnostic Benchmark for Biomedical AI Pipelines
Summary
This paper investigates the effectiveness of deep representation learning in single-cell RNA sequencing (scRNA-seq) clustering, a crucial step in biomedical analysis. The authors propose a diagnostic benchmark that evaluates nine clustering pipelines across ten real datasets, each varying in size and complexity. The study aims to determine whether the computational cost of deep learning methods, such as autoencoders and variational autoencoders, is justified compared to classical methods like principal component analysis (PCA). The methodology integrates hyperparameter optimization, robustness testing, and sensitivity analysis to assess the performance of different approaches. Results indicate that while deep learning methods can outperform classical techniques in specific scenarios, particularly with mid-sized datasets, classical PCA remains competitive when the data structure allows for effective linear projections. The findings emphasize the importance of dataset characteristics in selecting appropriate clustering methods, advocating for a dataset-aware approach to enhance efficiency in biomedical AI applications.
Methodology
The authors conducted a diagnostic benchmark involving nine clustering pipelines on ten datasets, utilizing techniques such as Optuna for hyperparameter search, repeated-run robustness checks, and Sobol sensitivity analysis to evaluate the impact of various hyperparameters on clustering performance.
Results
The contrastive autoencoder achieved the highest mean Adjusted Rand Index (0.7872), but statistical tests did not confirm its dominance over classical baselines. The analysis revealed that probabilistic VAE variants performed best on smaller datasets, while deep autoencoders excelled on mid-scale datasets. Classical PCA remained effective when the data's dominant variation could be captured linearly.
Implications
The findings suggest that practitioners should consider dataset-specific characteristics when choosing clustering methods for scRNA-seq data. This approach can lead to more efficient use of computational resources and improved outcomes in precision medicine applications.
Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning
Federated Learning
Interpretability
- Federated Learning was utilized to ensure patient data privacy while predicting CKD.
- A VotingClassifier combining Random Forest, AdaBoost, and XGBoost was employed for model prediction.
- GridSearchCV was used for optimizing model performance on the client side.
- Explainable AI techniques, particularly LIME, were integrated to enhance model interpretability.
Read more
Explainable AI for Chronic Kidney Disease Prediction Using Simulated Federated Learning
Summary
This study addresses the critical issue of Chronic Kidney Disease (CKD) prediction using Federated Learning (FL) combined with explainable AI (XAI) techniques. CKD is a significant public health concern characterized by the gradual loss of kidney function, making early detection essential for improving patient outcomes. The authors employed a VotingClassifier that integrates Random Forest, AdaBoost, and XGBoost algorithms to predict CKD from clinical datasets. The models were optimized using GridSearchCV to enhance performance on the client side. A notable aspect of the study is the incorporation of XAI methods, specifically LIME, to interpret the prediction mechanisms, thereby increasing model transparency and trustworthiness. The results demonstrated an impressive average accuracy of 99% for the global model, indicating the effectiveness of interpretable FL models in supporting early CKD diagnosis and advancing data-driven healthcare solutions. This research highlights the potential of federated learning in maintaining patient data privacy while delivering high-performance predictive models.
Methodology
The study implemented Federated Learning with a VotingClassifier that included Random Forest, AdaBoost, and XGBoost algorithms. GridSearchCV was applied for hyperparameter tuning on the client side, and LIME was used for model interpretation to enhance explainability.
Results
The global model achieved an average accuracy of 99%, indicating a high level of performance in predicting Chronic Kidney Disease. The integration of explainable AI techniques allowed for better understanding and trust in the model's predictions.
Implications
The findings suggest that federated learning can effectively support the development of predictive models in healthcare while preserving patient privacy. The high accuracy and interpretability of the model can facilitate early diagnosis of CKD, potentially leading to improved patient outcomes and more informed clinical decisions.
EchoBridge: Long-Tail-Aware ECG-Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings
Multimodal
- Introduction of EchoBridge for ECG-echocardiography text alignment.
- Utilization of CSPP and APBC to address cross-modal representation interference and long-tailed distributions.
- Significant performance improvements over baseline methods in classification metrics.
- Effective in identifying low-prevalence cardiac findings.
Read more
EchoBridge: Long-Tail-Aware ECG-Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings
Summary
The paper presents EchoBridge, a novel approach for aligning ECG data with echocardiography-derived cardiac findings, particularly addressing the challenges posed by long-tailed distributions of cardiac conditions. The authors propose two key methodologies: Complementary SharedβPrivate Projection (CSPP) and Adaptive Prototype Boundary Calibration (APBC). CSPP effectively reduces redundancy in modality-specific factors by creating shared and private projections for each modality, while APBC organizes the shared hypersphere using class-specific prototypes and adaptive angular margins. The evaluation of EchoBridge is conducted on multiple datasets, including EchoNext-Mini and independent cohorts, using various protocols such as prompt-based inference and cross-center transfer. The results demonstrate significant improvements in classification metrics, including AUROC, AUPRC, and F1 scores, outperforming existing baselines. The findings indicate that EchoBridge is particularly effective for low-prevalence cardiac conditions, showcasing its potential for enhancing ECG interpretation in clinical settings.
Methodology
The methodology involves two main components: CSPP, which creates shared and private projections to reduce redundancy and align modalities, and APBC, which organizes class-specific prototypes in a shared hypersphere with adaptive margins. The model is evaluated through various protocols, including classifier-free inference and cross-center transfer learning.
Results
EchoBridge achieved improvements of 7.88 points in AUROC, 5.61 points in AUPRC, and 4.54 points in F1 score compared to the strongest baselines. It also demonstrated the highest performance across all evaluation protocols and datasets, particularly excelling in identifying low-prevalence cardiac findings.
Implications
The findings suggest that EchoBridge can enhance the accuracy of ECG interpretations in clinical practice, particularly for rare cardiac conditions, thereby improving patient outcomes. The approach may also be applicable to other domains requiring multimodal data alignment.
Raven: High-Recall Sequence Modeling with Sparse Memory Routing
NLP
Large Language Models
Efficient ML
- Raven introduces Routing Slot Memories (RSMs) to enhance long-context recall in sequence models.
- The model selectively updates and decays memory slots, reducing interference and improving token recoverability.
- Raven outperforms existing linear-time models on recall-intensive benchmarks, maintaining effectiveness even with extended context lengths.
- The architecture simplifies prior models and does not depend on convolutional or sliding-window attention mechanisms.
Read more
Raven: High-Recall Sequence Modeling with Sparse Memory Routing
Summary
The paper introduces Raven, a novel linear-time sequence model designed to enhance long-context recall by utilizing a mechanism called Routing Slot Memories (RSMs). Traditional state-based linear models, such as state-space models (SSMs) and linear Transformers, face challenges in memory management due to dense updates that lead to interference and difficulty in recovering specific past tokens. Conversely, sliding-window attention (SWA) maintains explicit token representations but suffers from limited recall once tokens are evicted from the fixed window. Raven addresses these limitations by maintaining a fixed set of memory slots and employing learned, input-dependent routing to selectively update and decay only a subset of these slots. This approach mitigates the drawbacks of both SSMs and SWA, allowing for improved long-range content preservation. The authors demonstrate that Raven achieves competitive or superior performance on recall-intensive benchmarks compared to existing linear-time models, even when extrapolating to context lengths significantly larger than those seen during training. The model's architecture simplifies previous linear models and does not rely on convolutional or sliding-window attention layers, making it a versatile option for various sequence modeling tasks.
Methodology
Raven employs a fixed set of memory slots and uses a sparse, input-dependent routing mechanism to update and decay selected slots at each time step. This allows the model to manage memory more effectively, balancing the trade-offs between dense updates and sparse retention seen in traditional models.
Results
Raven demonstrates strong performance on various recall-intensive benchmarks, achieving perfect recall on specific tasks where SWA fails. It also shows competitive results against recent SSMs and linear models, maintaining accuracy even when extrapolating context lengths up to 16 times greater than those encountered during training.
Implications
The development of Raven has significant implications for tasks requiring long-context recall, such as language modeling and information retrieval. Its architecture could be applied in hybrid models and other sequence-based applications, potentially leading to advancements in natural language processing and related fields.
Minimizing Targeted Activations: Input-Only Suppression of Evaluation-Awareness Latents in Large Language Models
NLP
Large Language Models
Optimization
- Introduces the first input-only method for minimizing targeted internal latents in LLMs.
- Develops a behavioral + erasure protocol to differentiate genuine behavioral changes from mere activation drops.
- Demonstrates successful suppression of evaluation-awareness latents across multiple constructions.
- Finds that suppression does not equate to behavioral control, challenging assumptions about model evaluation.
Read more
Minimizing Targeted Activations: Input-Only Suppression of Evaluation-Awareness Latents in Large Language Models
Summary
This paper introduces a novel approach to control internal activations in large language models (LLMs) by optimizing prompts to suppress specific internal latents, particularly an 'evaluation-awareness' latent, without requiring access to the model during inference. The authors adapt techniques from Fluent Dreaming and EPO to minimize targeted activations through prompt optimization, focusing on five different constructions related to evaluation-awareness. The study reveals that while the suppression of these latents is achievable, it does not necessarily translate to behavioral control, as evidenced by various controls and tests that demonstrate the model's evaluation judgments remain unchanged or even increase despite successful latent suppression. This highlights a critical distinction between activation readability and behavioral controllability, emphasizing the need for robust methodologies in evaluating model behavior under different contexts.
Methodology
The authors adapt Fluent Dreaming and EPO techniques to create an input-side control mechanism that minimizes targeted internal latents through prompt optimization. This involves GCG-style discrete token-gradient optimization with a self-cross-entropy fluency regularizer, focusing on five constructions related to evaluation-awareness. The methodology includes a behavioral + erasure protocol to assess the impact of latent suppression on model behavior.
Results
The results indicate that prompt-side minimization effectively drives targeted latents toward zero across all five constructions, including a causally-validated sparse-autoencoder feature. However, the expected behavioral changes do not follow, as the model's evaluation judgments remain largely unchanged or even increase in certain contexts, underscoring the distinction between activation suppression and behavioral control.
Implications
The findings suggest that while it is possible to suppress certain internal signals in LLMs, this does not guarantee control over the model's behavior during evaluations. This has significant implications for safety evaluations and the reliability of model behavior in deployment, indicating a need for more nuanced approaches to model interpretability and control.
Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework
Computer Vision
Theory
Interpretability
- Introduces a multi-scale structural feature representation for visual recognition.
- Demonstrates continual learning without destructive adaptation or reliance on past data.
- Achieves higher accuracy on class-incremental MNIST compared to traditional methods.
- Retains previously learned classes while integrating new observations.
Read more
Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework
Summary
This paper addresses the challenges of continual learning in machine learning, particularly the issues of knowledge retention and comprehensibility in visual recognition tasks. The author introduces a developmental, gradient-free learning framework that builds a discrete, topological model of inputs, allowing for continual learning without the need for replay buffers or predefined task boundaries. The key contribution is a novel multi-scale structural feature representation that captures shape structure across various scales, enhancing the expressiveness of the feature representation. This new approach is integrated with a network-refinement learning process, improving learning dynamics and prediction accuracy. The methodology is evaluated using the class-incremental MNIST dataset, demonstrating significant improvements in accuracy while retaining previously learned classes. The results indicate that the proposed method not only matches or exceeds traditional replay and regularization-based approaches but also preserves the integrity of earlier learned information, making the learned representations interpretable and human-readable.
Methodology
The methodology involves a developmental learning framework that constructs a topological model of visual inputs through local variation and selection. The multi-scale representation encodes shape structures at different levels of granularity, allowing for effective refinement and retention of learned features. The learning process is gradient-free and operates without the need for replay buffers or task boundaries, integrating information sample by sample.
Results
The proposed method achieved a held-out accuracy of 0.87 on the class-incremental MNIST dataset, surpassing previous benchmarks and traditional methods that rely on replay or regularization. The system demonstrated the ability to retain knowledge of earlier learned classes while successfully incorporating new data.
Implications
The findings suggest that this developmental learning framework could be applied to various visual recognition tasks, enhancing the ability of machine learning systems to learn continuously and interpretably. This approach may have implications for real-world applications where continual learning is essential, such as robotics, autonomous systems, and interactive AI.
Generator-Aligned Representation Interfaces for Diagnostic Soft Equivariance
Computer Vision
Multimodal
Theory
- Introduction of the Generator-Aligned Representation Interface (GARI) for soft equivariance.
- GARI allows for the integration of transformation generators into generic sequence backbones.
- Direct Equivariance Error (DEE) is proposed as a measure for assessing representation-level soft equivariance.
- Experiments show improved task performance and representation consistency across various data modalities.
Read more
Generator-Aligned Representation Interfaces for Diagnostic Soft Equivariance
Summary
This paper introduces the Generator-Aligned Representation Interface (GARI), a novel framework designed to bridge the gap between exact-equivariant architectures and generic sequence backbones in machine learning. Traditional equivariant architectures encode specific group actions in their operators, which complicates their adaptability across different data modalities. GARI addresses this by exposing transformation generators to a generic backbone through aligned canonical and generator-induced views. The framework formalizes soft equivariance, distinguishing representation consistency from task robustness and exact equivariance. GARI is instantiated as GARI-Net, which processes generator-indexed streams and allows for cross-stream information exchange while repairing context mismatches. The authors propose Direct Equivariance Error (DEE) as a diagnostic measure for assessing the representation relation under known transformations. Experiments conducted on genomic sequences, images, and 3D point clouds demonstrate that GARI-Net supports task-relevant transformation consistency and generalization without requiring group-specific redesigns. This makes GARI a portable complement to hard-equivariant architectures, making generator structures accessible and measurable while maintaining a distinction from exact equivariance certification.
Methodology
The authors developed GARI as a representation-level design principle that utilizes selected transformation generators to create aligned views of input data. GARI-Net was instantiated to process these views through a shared sequence backbone, allowing for information exchange and context mismatch repair. The performance was evaluated using Direct Equivariance Error (DEE) and task accuracy metrics across different datasets.
Results
GARI-Net demonstrated improved rotation accuracy on ImageNet-1K by 1.34 percentage points over a baseline, with a significant reduction in DEE from 0.983 to 0.831. On MNIST, broader generator coverage led to enhanced performance, indicating that GARI effectively supports transformation consistency and generalization.
Implications
The GARI framework has the potential to enhance the adaptability and performance of machine learning models across various data modalities, making it easier to implement equivariant architectures without extensive redesign. This could lead to more robust models in fields such as computer vision and bioinformatics.
Perturbative-NeuSA: A Structured Spectral Framework for Time-Dependent PDEs
Theory
Efficient ML
Interpretability
- Perturbative-NeuSA separates the solution into a low-fidelity background and a high-resolution perturbation, optimizing learning efficiency.
- The framework outperforms the baseline NeuSA without requiring neural network training, particularly excelling in the Burgers equation.
- The effectiveness of the neural closure is conditional, varying with background fidelity and residual organization.
- The method provides a structured approach to diagnosing the contributions of different components in the solution process.
Read more
Perturbative-NeuSA: A Structured Spectral Framework for Time-Dependent PDEs
Summary
The paper introduces Perturbative-NeuSA, a novel framework designed to enhance neural spectral solvers for time-dependent partial differential equations (PDEs). Traditional neural spectral PDE solvers often attempt to learn the entire unresolved vector field, which can be inefficient when a simpler approximate model suffices. Perturbative-NeuSA addresses this by decomposing the target solution into a low-fidelity background and a high-resolution perturbation, focusing the learning on the unresolved dynamics. The method employs a fixed spectral operator, a background-dependent correction, and an optional neural closure to improve accuracy. The authors demonstrate the effectiveness of Perturbative-NeuSA across three PDEs: 2D Burgers, KleinβGordon, and heterogeneous 2D wave equations. The results show significant performance improvements over the baseline NeuSA, particularly in the Burgers equation, where training and extrapolation errors are reduced by factors of 24 and 44, respectively. The study also reveals that the benefits of the neural closure depend on the background resolution and the organization of the residual, indicating that the closure's effectiveness is conditional rather than universally beneficial.
Methodology
The authors propose a residual formulation that decomposes the target solution into a low-fidelity background and a high-resolution perturbation. The method utilizes a fixed spectral operator, a background-dependent correction, and an optional neural closure to model unresolved dynamics. The structured perturbative framework allows for a clear separation of contributions from the background, correction, and neural closure.
Results
Perturbative-NeuSA demonstrated substantial improvements in accuracy over the NeuSA baseline across three PDEs. Specifically, for the Burgers equation, training and extrapolation errors were reduced by factors of 24 and 44, respectively. The study also found that the neural closure's benefits varied with the background resolution and the organization of the residual, indicating a nuanced relationship between these factors.
Implications
The findings suggest that Perturbative-NeuSA could be applied to various time-dependent PDE problems, potentially leading to more efficient and interpretable neural solvers. The conditional nature of the neural closure's effectiveness may guide future research in optimizing neural network architectures for PDE applications.
Semantic Space Search Trajectory Networks
Optimization
Theory
- Introduction of a clustering-based discretization strategy for constructing STNs in high-dimensional semantic spaces.
- Enables qualitative and quantitative comparisons of learning dynamics across different algorithm families.
- Reveals systematic differences in learning dynamics between standard training and label randomization in neural networks.
- Semantic Space STNs provide insights into the interaction between learning algorithms and data.
Read more
Semantic Space Search Trajectory Networks
Summary
This paper introduces Semantic Space Search Trajectory Networks (STNs), a novel methodology for visualizing and analyzing the behavior of optimization algorithms in high-dimensional semantic spaces. Traditional STNs rely on discretization of the search space, which limits their applicability to low-dimensional or combinatorial problems. The authors propose a clustering-based approach that aggregates semantic vectors into network nodes using agglomerative clustering with complete linkage under a normalized Hamming distance. This allows for the comparison of learning dynamics across different machine learning algorithms by summarizing their behavior through semantic vectors. The methodology is applied to both classification and regression tasks, revealing known qualitative differences between various algorithms. Additionally, the authors investigate neural network generalization by contrasting standard training with a label randomization regime, finding that training on real labels results in denser and more centralized graphs compared to training on shuffled labels. The findings suggest that Semantic Space STNs effectively capture the functional dynamics of learning, providing a valuable tool for analyzing and comparing machine learning models and training regimes.
Methodology
The authors developed a methodology for constructing Semantic Space STNs by discretizing semantic vectors and aggregating them into network nodes using agglomerative clustering with complete linkage based on normalized Hamming distance. This approach allows for the visualization of optimization behavior without relying on explicit search space discretization.
Results
The application of Semantic Space STNs to various machine learning tasks demonstrated known qualitative differences in learning dynamics among different algorithms. The analysis of neural network training revealed that training with real labels produced denser and more efficient graphs compared to training with randomized labels, indicating distinct structural differences in learning behavior.
Implications
The findings suggest that Semantic Space STNs can be utilized as a powerful tool for analyzing and comparing the dynamics of different machine learning models and training regimes, potentially leading to improved understanding and optimization of learning processes.
A Coulomb Particle Model for Learning Kernel Attention in Transformers
NLP
Theory
Efficient ML
- Introduces a particle-based method for learning feature distributions in kernel attention.
- Utilizes a Hamiltonian framework to optimize kernel-target alignment.
- Demonstrates improved performance in accuracy and robustness on NLP benchmarks.
- Maintains linear inference complexity while enhancing attention mechanisms.
Read more
A Coulomb Particle Model for Learning Kernel Attention in Transformers
Summary
This paper introduces a novel approach to learning kernel attention in Transformers using a Coulomb particle model. The authors propose a method that optimizes the distribution of randomized features to enhance kernel-target alignment while employing a Riesz/Coulomb repulsive potential to maintain diversity among the features. The method is implemented in a two-phase training process for linearized Transformer attention, where the first phase focuses on learning positive random-feature maps, followed by freezing these features to train the remaining network parameters. The experiments conducted on synthetic classification tasks and sentence-level benchmarks demonstrate that the learned kernelized attention significantly improves accuracy, calibration, and robustness across various feature maps, all while maintaining linear inference complexity. This work bridges the gap between kernel methods and deep learning, providing a scalable and effective way to incorporate kernelized attention mechanisms into Transformer architectures.
Methodology
The authors develop a particle Hamiltonian that facilitates target-aligned random-feature learning. They employ a Langevin optimization procedure, which leads to a mean-field continuum limit. The method is instantiated in the context of linearized Transformer attention, where the feature particles are learned in an initial alignment phase before freezing them for subsequent training of the network parameters using cross-entropy loss.
Results
The experimental results indicate that the learned kernelized attention mechanism outperforms traditional methods in terms of accuracy, calibration, and robustness across multiple benchmarks, including synthetic classification tasks and natural language processing datasets. The approach successfully preserves linear-time inference complexity, making it computationally efficient.
Implications
This research has significant implications for enhancing Transformer models in various applications, particularly in NLP tasks where attention mechanisms are crucial. The ability to learn kernelized attention dynamically could lead to more adaptable and robust models that perform better across diverse datasets and tasks.
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
NLP
Large Language Models
Graph Learning
- HYSET treats tool retrieval as a set-level problem, allowing for joint evaluation of tool sets.
- The method captures size-dependent compatibility among tools through cardinality-specific interactions.
- HYSET functions as a pre-selection module, requiring no changes to downstream LLM agents.
- Experimental results indicate significant improvements over existing retrieval methods in both accuracy and task success.
Read more
Tools Are Not Islands: Set-Level Tool Retrieval for LLM Agents via Query-Conditioned Hyperedge Prediction
Summary
The paper introduces HYSET, a novel approach to tool retrieval for large language model (LLM) agents, addressing the limitations of existing retrieval methods that evaluate tools in isolation or sequentially. HYSET formulates tool retrieval as a query-conditioned hyperedge prediction problem on a tool co-invocation hypergraph, allowing for the joint evaluation of tool sets rather than individual tools. The authors emphasize the importance of capturing size-dependent tool compatibility through cardinality-specific interactions. HYSET operates as a pre-selection module that integrates seamlessly with existing LLM agents without requiring modifications. Experimental results on the ToolBench dataset demonstrate that HYSET significantly outperforms state-of-the-art baselines in both tool retrieval accuracy and overall task success rates. Additionally, HYSET shows strong performance in zero-shot and few-shot transfer scenarios, effectively generalizing to new tools and domains with minimal supervision.
Methodology
The authors propose a hyperedge-based framework for tool retrieval, where tools are represented in a co-invocation hypergraph. The retrieval process involves predicting hyperedges conditioned on the query, allowing for the assessment of tool sets as a whole. This approach contrasts with traditional methods that score tools independently or assemble them sequentially, thus failing to capture the interdependencies among tools.
Results
HYSET consistently outperformed state-of-the-art retrieval methods on the ToolBench dataset, achieving higher tool retrieval performance and end-to-end task success rates. The method also demonstrated robust performance in zero-shot and few-shot settings, indicating its ability to generalize to unseen tools and categories.
Implications
The findings suggest that adopting a set-level approach to tool retrieval can enhance the efficiency and effectiveness of LLM agents in real-world applications. This could lead to improved task completion rates and a more streamlined interaction with extensive tool libraries, making LLMs more capable in diverse operational contexts.
Spectral-Aware Analytic Class-Incremental Learning for Long-Tailed Distributions
Theory
Efficient ML
- Identification of spectral collapse as a critical issue in RLS-based ACL methods under long-tailed distributions.
- Introduction of Geometry-Spectral Rectification (GSR) as a solution that selectively inflates eigenvalues for tail classes.
- Theoretical proof of GSR's effectiveness in improving numerical stability and stable rank of the Gram matrix.
- Demonstration of GSR's superior performance compared to existing ACL methods on benchmark datasets.
Read more
Spectral-Aware Analytic Class-Incremental Learning for Long-Tailed Distributions
Summary
This paper addresses the challenges faced by Analytic Continual Learning (ACL) methods, particularly those based on Recursive Least Squares (RLS), in scenarios characterized by long-tailed distributions. The authors identify that the performance degradation in these scenarios is due to a phenomenon termed 'spectral collapse', where tail classes suffer from severe numerical instability. To mitigate this issue, they propose a novel framework called Geometry-Spectral Rectification (GSR), which treats long-tailed learning as a spectral regularization problem. Unlike standard isotropic regularization methods, GSR employs an anisotropic spectral filter that selectively inflates the collapsed eigenvalues of tail classes, thus improving the numerical stability of the Gram matrix. The theoretical foundation of GSR is established, demonstrating its effectiveness in enhancing the stable rank of the Gram matrix. Extensive experiments validate that GSR achieves state-of-the-art performance in analytic class-incremental learning, effectively balancing computational efficiency and robust generalization in long-tailed settings.
Methodology
The authors propose Geometry-Spectral Rectification (GSR), which involves constructing a data-dependent spectral perturbation matrix that selectively inflates the collapsed eigenvalues of the Gram matrix. This approach is grounded in statistical geometry and employs a class-adaptive manifold mixing strategy to enhance the representation of tail classes without distorting the head classes.
Results
The experiments conducted on benchmark datasets show that GSR significantly outperforms existing ACL baselines, achieving new state-of-the-art results in analytic class-incremental learning. The method demonstrates a robust trade-off between computational efficiency and generalization performance in long-tailed settings.
Implications
The findings suggest that GSR can be effectively applied in real-world applications where data streams exhibit long-tailed distributions, such as in edge devices and real-time systems. This approach could enhance the performance of machine learning models in scenarios where class imbalance is prevalent.
Empirical Evaluation of Out-Of-Distribution Performance of Tabular Foundation Models
Theory
Efficient ML
Interpretability
- First empirical evaluation of TFMs under distribution shifts across three real-world datasets.
- Systematic performance degradation of TFMs under OOD conditions, regardless of pre-training strategy.
- Identification of a scalability gap in high-performing models requiring significant computational resources.
- Extension of OOD analysis from traditional models to TFMs, highlighting the relationship between in-distribution and OOD performance.
Read more
Empirical Evaluation of Out-Of-Distribution Performance of Tabular Foundation Models
Summary
This paper presents an empirical evaluation of the Out-Of-Distribution (OOD) performance of nine Tabular Foundation Models (TFMs) across various pre-training strategies and architectures. The study addresses the gap in research regarding the robustness of TFMs under distribution shifts, which are common in real-world applications. The authors evaluate TFMs on three datasets from the TableShift study, which include different types of distribution shifts: label shift, socioeconomic shift, and geographic shift. The results indicate that all evaluated TFMs experience systematic performance degradation under distribution shifts, with observed shift gaps ranging from 0.003 to 0.060. The study also finds a significant scalability gap, as high-performing models require substantial computational resources, limiting their practical deployment. This research extends existing benchmarks for OOD performance in tabular data and provides insights into the implications for high-stakes domains characterized by structural distribution shifts.
Methodology
The authors conducted an empirical evaluation of nine TFMs with varying pre-training strategies on three datasets (HELOC, Childhood Lead, Voting) that represent different types of distribution shifts. The performance of these models was assessed under OOD scenarios to understand their robustness and scalability.
Results
The evaluation revealed that all TFMs exhibited performance degradation under distribution shifts, with shift gaps ranging from 0.003 to 0.060. The study confirmed that the relationship between in-distribution and OOD performance observed in classical models is also applicable to TFMs. Additionally, the analysis highlighted significant computational constraints for deploying high-performing TFMs.
Implications
The findings suggest that while TFMs show promise for tabular predictive tasks, their deployment in real-world applications must consider the potential for distribution shifts and the associated computational requirements. This research underscores the importance of developing robust models capable of maintaining performance across varying data distributions, particularly in critical domains such as healthcare and finance.
Physics-Informed Neural Networks for Predicting Nitrous Oxide Flux
Theory
- Introduces Physics-Informed Neural Networks (PINNs) for predicting N2O emissions.
- Demonstrates significant performance improvement over traditional models like Cycles.
- Finds that physics constraints enhance robustness but may degrade accuracy on familiar data.
- Highlights the challenges of cross-site generalization in N2O prediction.
Read more
Physics-Informed Neural Networks for Predicting Nitrous Oxide Flux
Summary
This paper addresses the challenge of predicting nitrous oxide (N2O) emissions, a significant greenhouse gas primarily produced through agricultural processes. The authors propose the use of Physics-Informed Neural Networks (PINNs) to improve prediction accuracy by incorporating biogeochemical constraints derived from established process-based models like DayCent. The study utilizes a multi-site agricultural dataset from four distinct locations in the U.S. to train an MLP-based PINN. The results demonstrate that the PINN significantly outperforms traditional uncalibrated simulation models, achieving a mean R2 of 0.411 across various configurations. While the inclusion of physics constraints enhances out-of-distribution robustness, it also leads to a trade-off in in-distribution accuracy, particularly affecting performance on unseen geographical sites. This research highlights the potential of PINNs in environmental modeling, particularly for complex systems like N2O emissions, where traditional methods struggle to capture the underlying biogeochemical processes.
Methodology
The authors constructed a PINN by integrating mechanistic equations from the DayCent model into the loss function of a multi-layer perceptron (MLP). The model was trained on a dataset from four agricultural sites, optimizing for both data loss and physics loss to ensure predictions adhered to biogeochemical principles.
Results
The PINN consistently outperformed the uncalibrated Cycles simulation, achieving a mean R2 of 0.411 across ten random seeds. The physics constraints improved model performance in leave-one-site-out validation but resulted in performance degradation in holdout validation, particularly at higher physics loss weighting.
Implications
The findings suggest that PINNs can provide a more reliable framework for predicting N2O emissions, which is crucial for climate mitigation strategies. This approach could lead to better-informed agricultural practices and policies aimed at reducing greenhouse gas emissions.
Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls
NLP
Large Language Models
Theory
- Introduces role-stratified per-field conformal risk control for language-model tool calls.
- Addresses the inadequacy of aggregate risk control methods that obscure failures in high-risk fields.
- Demonstrates improved risk budget compliance and robustness under various conditions.
- Provides formal guarantees for role-specific risk management.
Read more
Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls
Summary
This paper addresses the risks associated with language-model agents that execute structured tool calls, emphasizing the need for role-specific risk control. Traditional statistical methods often overlook high-risk fields by averaging risks across entire actions, which can mask failures in critical areas. The authors propose a novel approach called role-stratified per-field conformal risk control (CRC), which assigns separate thresholds and risk budgets to different semantic roles within tool calls. This method ensures that each role, such as target, credential, or command, is certified independently, thus preventing the dilution of risk associated with rare but critical roles. The paper demonstrates that this stratification leads to improved compliance with risk budgets across various models and scenarios, including model transfer, unseen tool suites, and adaptive attacks. The findings suggest that structured tool calls should be certified at the semantic-role level rather than as a whole action, providing a more robust framework for managing risks in language model applications.
Methodology
The authors developed a calibration layer that applies class-conditional conformal risk control at the semantic role level. Each argument in a tool call is assigned a specific role with its own risk budget, allowing for tailored risk management. The method incorporates pooled certification for rarer roles and ensures compliance through empirical validation across multiple language models and benchmarks.
Results
The proposed role-stratified CRC method showed significant improvements in risk budget compliance compared to traditional aggregate methods. It maintained robustness against model and attack transfer, detector noise, gradual drift, and adaptive attacks. The empirical utility gap aligned with theoretical predictions, confirming the effectiveness of role-specific calibration.
Implications
This research has significant implications for the development of safer language-model applications, particularly in sensitive areas such as finance and personal data management. By ensuring that high-risk fields are managed independently, the approach can enhance the reliability and trustworthiness of automated systems that rely on language models.
A Statistical Difference between Single-Layer Learning and Hierarchical Learning in Wide Neural Networks
Theory
- Training input-to-hidden weights reduces generalization error compared to fixing them.
- Fixed input-to-hidden weights lead to singularities in the parameter space.
- Both learning schemes are universal function approximators but exhibit different approximation properties.
- Singularities are significant in wide neural networks and influence representation learning.
Read more
A Statistical Difference between Single-Layer Learning and Hierarchical Learning in Wide Neural Networks
Summary
This paper investigates the mathematical properties of hierarchical neural networks, particularly focusing on the differences between single-layer learning and hierarchical learning in the context of wide neural networks. The author examines a three-layer neural network with a finite number of hidden units and compares two training schemes: one where the input-to-hidden weights are fixed and another where both input-to-hidden and hidden-to-output weights are optimized. The study reveals that optimizing the input-to-hidden weights results in a smaller generalization error compared to keeping them fixed. Furthermore, the fixed setting leads to singularities in the parameter space, which do not occur when weights are optimized. These findings suggest that singularities play a crucial role in the learning dynamics of wide neural networks, challenging the notion that their effects diminish in large-scale models. The paper emphasizes the importance of hierarchical structures in achieving effective representation learning, even in the infinite-width limit.
Methodology
The author analyzes a three-layer neural network and compares two learning schemes: one with fixed input-to-hidden weights and another with both input-to-hidden and hidden-to-output weights optimized. The study employs statistical inference methods to derive generalization errors and free energies associated with each learning scheme.
Results
The paper demonstrates that the learning curves for the two schemes differ fundamentally, with the scheme that optimizes both sets of weights achieving a lower generalization error. It also establishes that the fixed scheme is regular while the optimized scheme is singular, highlighting the impact of singularities on learning performance.
Implications
The findings suggest that understanding and addressing singularities in neural network training could lead to improved performance in various AI applications. The results may influence future research on neural network architectures and training methodologies, emphasizing the need for hierarchical structures in deep learning.
Physically Verifiable Evidence and LLM-Based Reporting for Bearing Fault Diagnosis
Time Series
Interpretability
Large Language Models
- Introduction of the Diagnostic Evidence Network (DENet) for bearing fault diagnosis.
- DENet provides a structured evidence record that includes classification, characteristic frequency prediction, and impulse localization.
- The framework allows for independent validation of predictions against physical reality.
- A QLoRA-adapted language model translates evidence into natural language, ensuring traceability and reducing hallucination rates.
Read more
Physically Verifiable Evidence and LLM-Based Reporting for Bearing Fault Diagnosis
Summary
This paper addresses the critical need for verifiable outputs in AI-based fault diagnosis systems, particularly in safety-critical applications like bearing fault diagnosis. Current systems typically output a class label with a softmax confidence score, which lacks independent verification against physical reality. The authors propose the Diagnostic Evidence Network (DENet), an encoder-agnostic multi-task framework that enhances the output to include a structured evidence record. This record comprises a fault classification, a predicted characteristic frequency that can be compared to theoretical values based on bearing geometry and shaft speed, and a temporal localization of transient impulses that can be inspected against raw waveform data. The proposed framework incurs no significant accuracy cost while providing a validation signal that can detect misclassifications effectively. Additionally, a QLoRA-adapted language model is employed to translate the evidence into natural language reports, ensuring that all statements are traceable to specific evidence fields, thereby reducing the risk of hallucinated content. Overall, the work emphasizes the importance of verifiable outputs in AI diagnostics and presents a novel approach to achieving this goal.
Methodology
The authors developed the Diagnostic Evidence Network (DENet), which is an encoder-agnostic framework that outputs a structured evidence record instead of a simple class label and confidence score. The framework includes a predicted characteristic frequency and temporal localization of impulses. A bidirectional multi-scale Mamba encoder is used as the default backbone, and a QLoRA-adapted language model is employed to generate natural language reports based on the evidence provided by DENet.
Results
DENet demonstrated no statistically significant accuracy loss while providing a mean absolute error of approximately 6 Hz for frequency predictions on 1,024-point segments. The validation signal derived from the deviation between predicted and theoretical frequencies effectively detected misclassifications with AUROC scores of 0.970 and 0.871, particularly in high-confidence scenarios where traditional confidence scores may fail.
Implications
The proposed framework enhances the reliability of AI-based fault diagnosis systems in safety-critical applications by providing verifiable outputs. This approach can lead to increased trust in AI diagnostics, facilitating broader industrial adoption and improving maintenance decision-making processes.
A2TTA: Anchored-and-Agile Test-Time Adaptation for Evolving Traffic Sensor Networks
Graph Learning
Time Series
Optimization
- A2TTA effectively addresses the challenges of topology expansion and temporal distribution shifts in traffic forecasting.
- The framework separates adaptation into global corrections and context-specific specializations for better performance.
- Extensive experiments validate the robustness and efficiency of A2TTA across multiple real-world traffic networks.
Read more
A2TTA: Anchored-and-Agile Test-Time Adaptation for Evolving Traffic Sensor Networks
Summary
The paper presents A2TTA, an innovative framework for test-time adaptation (TTA) specifically designed for evolving traffic sensor networks. Traditional traffic forecasting models often assume static sensor graphs, which leads to performance degradation when faced with dynamic changes in real-world traffic networks, such as new sensor deployments and shifting human mobility patterns. A2TTA addresses two main challenges: topology expansion, which alters the sensor graph structure, and temporal distribution shifts, which affect the input-output relationships over time. The framework transforms the problem of topology-induced forecasting errors into an expandable output calibration issue and separates temporal adaptation into two components: persistent global correction and agile context-specific specialization. Through extensive experiments on ten real-world traffic networks, A2TTA demonstrates significant improvements in forecasting performance across various backbones, datasets, and prediction horizons, showcasing its robustness and efficiency in adapting to continuously evolving traffic environments.
Methodology
The A2TTA framework employs a dual approach to adaptation: it first calibrates outputs to account for topology changes and then applies a two-tier adaptation strategy to handle both long-term and short-term temporal shifts. This involves using persistent global corrections for stable shifts and agile context-specific specializations for transient deviations.
Results
The experiments conducted on ten real-world traffic networks indicate that A2TTA consistently enhances forecasting accuracy compared to traditional methods, demonstrating its capability to adapt effectively to evolving conditions in traffic sensor networks.
Implications
The findings suggest that A2TTA can significantly improve traffic management and route planning in smart cities, enabling more efficient responses to dynamic traffic conditions and enhancing the overall functionality of intelligent transportation systems.
SpectONet: A Physics-Guided Spectral Deep Operator Network for Euler-Bernoulli Beam Dynamics
Theory
Efficient ML
- Introduction of SpectONet, a physics-guided spectral deep operator network for EBB dynamics.
- Utilizes nonuniform spectral sensor placement for improved boundary response representation.
- Incorporates physics-informed constraints into the training objective for better generalization.
- Demonstrates significant performance improvements over traditional models in both synthetic and real-world datasets.
Read more
SpectONet: A Physics-Guided Spectral Deep Operator Network for Euler-Bernoulli Beam Dynamics
Summary
This paper introduces SpectONet, a novel physics-guided spectral deep operator network designed to address Euler-Bernoulli beam (EBB) vibration problems. The framework enhances the operator-learning capabilities of DeepONet by integrating physics-informed constraints and employing Chebyshev-Gauss-Lobatto (CGL) sensor placement. Unlike traditional DeepONet models that utilize uniformly distributed sensors, SpectONet strategically places sensors nonuniformly, concentrating more near the boundaries of the domain. This approach improves the representation of boundary-sensitive structural responses while minimizing the number of required branch-network inputs. The training objective incorporates the governing beam equation along with initial and boundary conditions, ensuring that the predictions are physically consistent and generalizable. The effectiveness of SpectONet is validated through numerical experiments on three synthetic EBB vibration problems and a real-world bridge vibration dataset, demonstrating its superior performance compared to established baselines such as Vanilla DeepONet, PI-DeepONet, PINN, and CNN-UNet. The results indicate that SpectONet achieves at least a 64% improvement in prediction accuracy for synthetic problems and a 37% improvement for real-world applications, showcasing its potential as an accurate, efficient, and physically consistent framework for structural vibration analysis.
Methodology
The methodology involves integrating the operator-learning capabilities of DeepONet with physics-informed constraints and employing a nonuniform sensor placement strategy using Chebyshev-Gauss-Lobatto points. The training objective incorporates the governing beam equation along with initial and boundary conditions to ensure physically consistent predictions.
Results
SpectONet outperformed baseline models, achieving at least a 64% reduction in prediction errors for synthetic EBB problems and a 37% reduction for real-world bridge vibration data. This demonstrates its effectiveness in providing accurate and efficient predictions.
Implications
The implications of this research suggest that SpectONet can be applied to various structural dynamics problems, enhancing the accuracy and efficiency of vibration analysis in engineering applications. Its ability to generalize across different conditions without retraining could significantly streamline the modeling process in real-time applications.
A Unified Algorithmic Framework for Hybrid Reinforcement Learning in Tabular MDPs with Shifted Transition Dynamics
Reinforcement Learning
Theory
Efficient ML
- Introduction of a unified framework for hybrid RL in tabular MDPs with shifted transition dynamics.
- Development of two algorithms: MIN-UCB-VI for regret minimization and MAX-LCB-VI for best policy identification.
- Provision of rigorous theoretical guarantees on regret and sub-optimality gaps.
- Demonstration of the framework's effectiveness through extensive experimental validation.
Read more
A Unified Algorithmic Framework for Hybrid Reinforcement Learning in Tabular MDPs with Shifted Transition Dynamics
Summary
This paper addresses the challenges of hybrid reinforcement learning (RL) in tabular Markov Decision Processes (MDPs) where offline data may originate from environments with shifted transition dynamics. The authors propose a unified algorithmic framework that includes two algorithms: MIN-UCB-VI for regret minimization and MAX-LCB-VI for best policy identification. These algorithms utilize fine-grained bias information to effectively exploit offline data, even when transition dynamics differ between the source and target environments. The paper provides theoretical guarantees, including instance-dependent and independent upper bounds on regret and sub-optimality gaps, alongside matching lower bounds to affirm the optimality of the proposed methods. Extensive experiments validate the theoretical findings, demonstrating the framework's effectiveness in improving sample efficiency in hybrid RL settings.
Methodology
The authors developed a unified algorithmic framework that includes MIN-UCB-VI and MAX-LCB-VI algorithms. These algorithms leverage bias information between source and target environments to enhance online learning efficiency. The framework also incorporates a model selection scheme for scenarios lacking prior knowledge of environment bias.
Results
The theoretical analysis provided both instance-dependent and independent upper bounds on regret and sub-optimality gaps, with matching lower bounds confirming the optimality of the proposed algorithms. Experimental results showed that the algorithms effectively improved learning efficiency in hybrid RL settings with shifted transition dynamics.
Implications
The proposed framework can significantly enhance the sample efficiency of RL agents in practical applications where offline data may not perfectly align with online environments. This has potential applications in various fields such as robotics, game playing, and real-world decision-making scenarios.
Explainable Reinforcement Learning via Physics-Aware Policy Distillation
Reinforcement Learning
Robotics
Interpretability
- Introduces a policy distillation framework for enhancing interpretability in DRL systems.
- Achieves performance parity with a high-performance DRL agent while using a simpler decision tree model.
- Demonstrates the trade-off between continuous and discrete control methods, affecting system dynamics.
- Maintains BIBO stability, ensuring safety in autonomous applications.
Read more
Explainable Reinforcement Learning via Physics-Aware Policy Distillation
Summary
This paper addresses the challenge of interpretability in Deep Reinforcement Learning (DRL), particularly in safety-critical applications such as robotics and automotive engineering. The authors propose a policy distillation framework that transforms a high-performance, opaque DRL agent (Twin Delayed DDPG) into an interpretable decision tree model. This is achieved through the use of a custom physics-aware feature and 'Noisy Oracle Rollouts' for dataset generation. The study demonstrates that the distilled decision tree model maintains performance comparable to the original DRL agent while providing both global and local interpretability. The authors also analyze the trade-offs involved in transitioning from continuous to discrete control, highlighting the emergence of Bang-Bang actuation and a stable bimodal limit cycle. The results indicate that the proposed method ensures Bounded-Input Bounded-Output (BIBO) stability, which is crucial for the deployment of safe autonomous systems.
Methodology
The authors implemented a policy distillation framework using the Inverted Pendulum benchmark. A Twin Delayed DDPG (TD3) agent served as the teacher model, while a shallow Decision Tree was used as the student model. The distillation process utilized a custom physics-aware feature and 'Noisy Oracle Rollouts' to generate a dataset for training the decision tree, ensuring that the distilled model retained high performance and interpretability.
Results
The distilled decision tree model achieved performance equivalent to the TD3 agent while providing clear interpretability. The comparative analysis revealed a trade-off in control dynamics, where the transition from continuous to discrete control led to high-frequency Bang-Bang actuation and a stable bimodal limit cycle. The method also ensured BIBO stability, which is essential for the safe operation of autonomous systems.
Implications
The findings suggest that the proposed framework can significantly enhance the interpretability and trustworthiness of DRL systems in safety-critical applications. This is particularly relevant for regulatory compliance and improving human-agent interactions in robotics and autonomous vehicles.
Emergent Latent-State Computation under Stochastic Volatility
Time Series
Interpretability
Theory
- Demonstrates a two-stage computation in neural sequence models for forecasting latent volatility states.
- Identifies specific architectural stages in Transformers where latent-state decodability emerges, varying with volatility periods.
- Shows that in long-cycle regimes, the computation simplifies to a learned linear projection followed by normalization.
- Highlights that forecast degradation under noisy training is linked to readout misalignment rather than latent state encoding failure.
Read more
Emergent Latent-State Computation under Stochastic Volatility
Summary
This paper investigates the mechanistic interpretability of neural sequence models, particularly focusing on how they represent latent stochastic dynamics in a multivariate stochastic volatility setting. The authors propose a controlled framework where models predict future squared returns based on noisy observations of latent volatility states. Through experiments with Transformers and MLPs, the study reveals a two-stage computation process: hidden representations encode significant information about the next latent volatility state, while the output head translates this representation into return forecasts. The findings indicate that in Transformers, the decodability of latent states emerges at specific architectural stages, influenced by the volatility period. In long-cycle regimes, the computation simplifies into a learned linear projection followed by normalization. Additionally, the paper highlights that performance degradation during noisy training is partly due to readout misalignment rather than representation failure. Overall, the study positions stochastic volatility models as valuable benchmarks for understanding mechanistic interpretability under conditions of partial observability and latent dynamics.
Methodology
The authors employed a controlled multivariate stochastic volatility process to simulate returns, using small Transformers and MLPs to forecast future squared returns. They analyzed the internal representations through linear probes, stage-wise representation analysis, and targeted interventions to understand how models recover latent volatility states.
Results
The study found that both Transformers and MLPs exhibit a robust two-stage computation where hidden states encode information about latent volatility. Transformers showed identifiable stages of latent-state decodability influenced by volatility periods. In long-cycle regimes, the computation became a simplified filter, and replacing the output head with a probe-aligned readout improved performance under noisy training conditions.
Implications
The findings suggest that stochastic volatility models can enhance our understanding of how neural models interpret and utilize latent stochastic structures, which could have implications for financial forecasting and the development of interpretable machine learning models in other domains.
LC-SEPLM: long-range contact-supervised adaptation for sequence-only protein representation learning
NLP
Large Language Models
Theory
- Introduces LC-SEPLM, which integrates long-range contact supervision into protein representation learning.
- Utilizes low-rank adaptation (LoRA) and cross-attention mechanisms to enhance model performance.
- Demonstrates significant improvements in protein-level tasks, especially in remote-homology recognition.
- Achieves superior performance on the ESM-S EC benchmark compared to previous models.
Read more
LC-SEPLM: long-range contact-supervised adaptation for sequence-only protein representation learning
Summary
The paper introduces LC-SEPLM (Long-range Contact-supervised ESM Protein Language Model), which enhances protein representation learning by incorporating long-range residue-pair contact supervision into the ESM2 model while maintaining sequence-only inference. Traditional protein language models primarily focus on contextual dependencies along amino-acid sequences, which limits their ability to learn about three-dimensional residue contacts formed during protein folding. LC-SEPLM addresses this limitation by utilizing low-rank adaptation (LoRA) and training on a dataset of 500,000 AlphaFold Swiss-Prot proteins. The model employs pair-specific queries that leverage cross-attention over the entire sequence to capture global context related to long-range spatial contacts. The evaluation of LC-SEPLM across eight protein-level tasks demonstrated significant improvements over ESM2, particularly in remote-homology recognition, where the macro-F1 score increased from 0.6122 to 0.6769. Additionally, LC-SEPLM outperformed ESM-S on the official ESM-S EC benchmark, achieving a maximum absolute gain of 0.1771. These findings suggest that residue-pair contact supervision is an effective method for integrating structural information into protein sequence representations without compromising the sequence-only inference capability.
Methodology
LC-SEPLM adapts the ESM2 model using low-rank adaptation (LoRA) and incorporates long-range residue-pair contact supervision derived from AlphaFold structures. The model employs pair-specific queries that utilize cross-attention over the entire sequence to capture global context related to spatial contacts. The training process focuses on informative long-range relations while ensuring that downstream inference remains sequence-only.
Results
LC-SEPLM improved all evaluated protein-level tasks compared to ESM2, with the most notable enhancement in remote-homology recognition (macro-F1 score increased by 6.47 percentage points). On the ESM-S EC benchmark, LC-SEPLM achieved a maximum absolute gain of 0.1771 over ESM-S, indicating its effectiveness in integrating structural information into sequence representations.
Implications
The findings suggest that incorporating long-range residue-pair contact supervision can significantly enhance the performance of protein language models, making them more effective for tasks such as remote-homology recognition and other protein-level predictions. This approach may also facilitate advancements in protein engineering and functional annotation.