AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
Retrofitting Linear Attention into Diffusion Language Models
NLP
Large Language Models
Efficient ML
- Introduction of block-hybrid attention to improve efficiency in dLLMs.
- LLADA-HYBRID model shows up to 1.7× higher decoding throughput.
- Minimal post-training required to retrofit linear attention into existing models.
- Maintains competitive performance on benchmark tasks despite architectural changes.
Read more
Retrofitting Linear Attention into Diffusion Language Models
Summary
This paper explores the integration of linear attention into diffusion language models (dLLMs) to enhance inference speed. dLLMs, which utilize blockwise semi-autoregressive decoding, face inefficiencies due to prefix-attention costs during denoising steps. The authors propose a novel approach called block-hybrid attention, which combines exact softmax attention within the active denoising block with linear attention over previous blocks. This method allows for efficient inference while maintaining performance. The authors retrofit this hybrid attention into a pre-trained dLLM, LLADA 2.1, with minimal post-training effort, demonstrating that it can replace six out of twenty attention layers effectively. The resulting model, LLADA-HYBRID, achieves significant improvements in decoding throughput and supports more concurrent requests without exhausting memory, showcasing the potential of linearizing pretrained dLLMs for faster inference.
Methodology
The authors propose a two-stage post-training process to retrofit block-hybrid attention into the LLADA 2.1 model. In the first stage, they freeze the backbone and train the hybrid layers to match the output of the original softmax-attention layers using an MSE loss. In the second stage, they apply LoRA (Low-Rank Adaptation) to fine-tune the hybrid model on the masked-token diffusion objective, ensuring that downstream performance is preserved.
Results
The retrofitted LLADA-HYBRID model achieves a decoding throughput improvement of up to 1.7 times compared to a strong baseline. Performance on benchmark tasks remains high, with scores of 72.0% vs. 75.6% on HumanEval, 63.0% vs. 57.7% on MBPP+, and 86.7% vs. 88.3% on CMATH, indicating that the hybrid attention approach effectively balances speed and accuracy.
Implications
The findings suggest that linear attention can be effectively integrated into existing diffusion language models, leading to faster inference times and improved scalability for applications requiring high throughput. This could have significant implications for real-time language processing tasks and large-scale deployments of language models.
Efficient Test-Time Scaling for LLM-based Time Series Forecasting
Time Series
Large Language Models
Efficient ML
- SCALER is a two-stage framework that combines a lightweight forecaster with a pretrained LLM for efficient time series forecasting.
- The framework uses a coarse-to-fine approach, predicting a coarse future shape before refining it with fewer tokens.
- SCALER achieves significant reductions in inference time, being seven times faster than standard test-time scaling methods.
- The method stabilizes predictions by anchoring refinements to an explicit coarse shape, reducing reliance on long prompts.
Read more
Efficient Test-Time Scaling for LLM-based Time Series Forecasting
Summary
This paper introduces SCALER, a novel two-stage framework designed to enhance long-term time series forecasting using large language models (LLMs). Traditional LLM-based forecasting methods often struggle with computational efficiency and global-shape consistency, particularly over extended prediction horizons. SCALER addresses these challenges by first employing a lightweight Transformer model to predict a coarse representation of future dynamics, capturing essential global structures such as trends and seasonality. This coarse prediction serves as a guide for a pretrained LLM, which performs iterative refinement in a fixed-step manner, processing fewer tokens at each step. This approach not only reduces computational overhead but also mitigates the risk of global-shape mismatch that can occur with longer prediction horizons. Experimental results demonstrate that SCALER outperforms existing forecasting baselines across various scenarios, including long-term, short-term, and zero-shot forecasting, while significantly lowering inference costs associated with LLMs.
Methodology
SCALER employs a two-stage process: in the first stage, a lightweight Transformer model predicts a coarse future shape of the time series, capturing essential global dynamics. In the second stage, a pretrained LLM refines this coarse prediction through a fixed-step iterative process, generating finer-scale token blocks while maintaining coherence with the predicted shape. This design minimizes computational costs and avoids the pitfalls of long prompt dependencies and global-shape drift.
Results
The experimental evaluation indicates that SCALER consistently outperforms strong forecasting baselines in long-term, short-term, and zero-shot forecasting scenarios. Additionally, SCALER demonstrates a significant reduction in inference costs, achieving an average inference time that is seven times faster than traditional test-time scaling methods.
Implications
The SCALER framework has practical implications for various domains requiring time series forecasting, such as finance, healthcare, and energy demand modeling. Its efficiency and stability make it suitable for real-time applications where computational resources are limited and accurate predictions are crucial.
CommitKV: Lifecycle-Aware KV Cache Compression via Commit Transitions for Multi-Turn Agents
Large Language Models
NLP
Efficient ML
- CommitKV distinguishes between dormant and completed KV states to optimize cache usage.
- The method employs commit transitions to evaluate the relevance of KV pages before and after tool calls.
- CommitKV achieves up to 22.24 percentage points improvement in accuracy over existing compression methods.
- The approach provides significant memory savings (up to 5.00×) and speed improvements (up to 5.62×) in inference.
Read more
CommitKV: Lifecycle-Aware KV Cache Compression via Commit Transitions for Multi-Turn Agents
Summary
The paper introduces CommitKV, a novel method for key-value (KV) cache compression tailored for multi-turn reasoning-and-acting (ReAct) agents. As these agents accumulate extensive trajectories of reasoning, tool calls, and observations, their KV caches expand, leading to increased memory usage and attention costs during inference. Traditional KV cache compression techniques often rely on attention scores to evict low-importance states, but this approach fails to account for the evolving significance of information across turns. CommitKV addresses this limitation by identifying the lifecycle of KV states through 'commit transitions.' The method partitions completed agent events into token pages and evaluates their deletion effects before and after tool-call commits. By distinguishing between dormant pages and those that have completed their roles, CommitKV employs a greedy joint validation strategy to selectively retire pages while protecting those that may still be relevant. Experimental results demonstrate that CommitKV significantly reduces memory usage, accelerates inference speed, and enhances accuracy compared to existing methods, showcasing its effectiveness in managing the complexities of multi-turn interactions.
Methodology
CommitKV partitions completed tool calls and observations into event pages, measuring the deletion effects of these pages before and after tool-call commits. It uses a greedy joint validation approach to determine which pages can be safely retired while protecting those that may still be useful in future interactions.
Results
The experiments conducted on six large language models across eight benchmarks show that CommitKV outperforms the best existing compression methods by up to 22.24 percentage points in average accuracy, achieves up to 5.00× peak KV-memory savings, and provides a 5.62× speedup in end-to-end inference.
Implications
The findings suggest that CommitKV can enhance the performance of multi-turn agents in various applications, such as conversational AI, where efficient memory management and faster response times are critical. This method could be applied to improve the efficiency of large language models in real-time applications.
Causal State-Space Model for Causal Inference: Estimating Longitudinal Individual Treatment Effects
Theory
Time Series
Optimization
- Introduction of the Causal State-Space model with Direct decoder (CSSD) for improved causal inference.
- Development of CSSPD, which integrates predictive regularization techniques to resolve mutual information conflicts.
- Formalization of the balancing-prediction mutual information conflict and its implications for counterfactual prediction accuracy.
- Empirical results show CSSPD achieves state-of-the-art performance on real-world and synthetic datasets.
Read more
Causal State-Space Model for Causal Inference: Estimating Longitudinal Individual Treatment Effects
Summary
This paper addresses the challenge of estimating counterfactual outcomes from longitudinal observational data, particularly in clinical decision support. The authors identify a fundamental issue with existing methods that utilize domain confusion, which leads to a mutual information conflict by suppressing treatment-correlated signals necessary for accurate predictions. To tackle this, they propose two models: the Causal State-Space model with Direct decoder (CSSD) and the Causal State-Space model with Predictive regularisation and Direct decoder (CSSPD). CSSD employs a parallel multi-step decoder to produce simultaneous predictions, thereby eliminating accumulated rollout errors. CSSPD enhances CSSD by incorporating Contrastive Predictive Coding (CPC) and Local Information Maximisation (LIM) to recover lost covariate information and reinforce temporal predictability. Empirical evaluations on the MIMIC-III dataset and a synthetic Cancer Simulation dataset demonstrate that CSSPD outperforms existing methods, achieving lower counterfactual RMSE and confirming the theoretical predictions regarding mutual information conflicts.
Methodology
The authors developed two models: CSSD, which uses a selective State Space Model with a parallel multi-step decoder to produce all predictions in a single forward pass, and CSSPD, which augments CSSD with CPC and LIM to address mutual information conflicts. They derived a Jensen–Shannon divergence bound to formalize the prediction error associated with domain confusion.
Results
CSSPD achieved lower counterfactual RMSE than the Causal Transformer method at every horizon τ ≥2 on the MIMIC-III dataset, with improvements ranging from 0.02 to 0.07. On the Cancer Simulation dataset, CSSPD outperformed the Causal Transformer at confounding strengths γ ≤3, with performance margins between 25.9% and 37.0%. CSSD also demonstrated a 12.7% reduction in average RMSE compared to the Causal Transformer.
Implications
The findings suggest that the proposed models can significantly enhance the accuracy of individualized treatment effect estimations in clinical settings, potentially leading to better patient outcomes and more informed clinical decision-making. The formalization of the mutual information conflict provides a new perspective on the limitations of existing causal inference methods.
ED-CSP: Crystal Structure Prediction from Electron Diffraction
Generative Models
- ED-CSP is a novel model for predicting crystal structures from sparse electron diffraction data.
- The model utilizes a unique combination of relational encoding and multi-view aggregation techniques.
- A large dataset of 4.85 million simulated structures was created to train and evaluate the model.
- ED-CSP outperforms existing methods, achieving a structural match rate of 57.5% and up to 66.3% with dataset expansion.
Read more
ED-CSP: Crystal Structure Prediction from Electron Diffraction
Summary
The paper introduces ED-CSP, a novel machine learning model designed for crystal structure prediction from electron diffraction (ED) data. Unlike previous methods that focused on predicting crystallographic labels or retrieving structures from finite libraries, ED-CSP aims to predict lattice and fractional atomic coordinates based on known composition and atom count, utilizing multiple sets of sparse, unindexed ED observations. The model employs a relational set encoder, permutation-invariant multi-view aggregation, and a periodic flow generator to effectively generate crystal structures. To train ED-CSP, the authors created a comprehensive dataset, Electron Diffraction Crystal Structures (ED-CS), comprising 4.85 million simulated structures filtered from various materials repositories. The performance of ED-CSP was evaluated against the CHILI-100K dataset, achieving a structural match rate of 57.5% at five candidates per query, surpassing the state-of-the-art PXRDGen model. Furthermore, the study demonstrates that increasing the dataset size significantly enhances model performance, with a warm-start from a one-million-structure precursor improving the match rate to 66.3%. The findings indicate that ED-CSP can recover structures even when exact-formula lookups are unavailable, showcasing its potential for generative inference from sparse ED data and future experimental applications.
Methodology
ED-CSP employs a relational set encoder to process multiple views of electron diffraction data, aggregates the information in a permutation-invariant manner, and utilizes a periodic flow generator to produce the final crystal structure predictions. The model is trained on a large dataset of simulated electron diffraction patterns corresponding to known crystal compositions.
Results
ED-CSP achieved a structural match rate of 57.5% at five candidates per query on the CHILI-100K dataset, outperforming the PXRDGen model. With a warm-start from a one-million-structure precursor, the match rate improved to 66.3%. The model also maintained a 53.5% match rate on queries with absent reduced formulas, indicating its ability to recover structures without exact matches.
Implications
The development of ED-CSP has significant implications for materials science, particularly in predicting crystal structures from electron diffraction data, which is crucial for understanding material properties and guiding experimental efforts. The model's ability to generate structures from sparse data may facilitate advancements in the discovery of new materials and enhance the efficiency of crystallography research.
Density-aware Hierarchical Clustering Based on Element-Categorized Connection Subgraphs
Graph Learning
Theory
Efficient ML
- Introduction of DHC-ECS, a density-aware hierarchical clustering method.
- Novel inter-cluster similarity metric that incorporates density and connectivity.
- Extensive evaluations show superior performance compared to baseline methods.
- Potential for reduced reliance on manual parameter tuning in clustering tasks.
Read more
Density-aware Hierarchical Clustering Based on Element-Categorized Connection Subgraphs
Summary
This paper introduces a novel density-aware hierarchical clustering method called DHC-ECS, which integrates hierarchical, density-based, and graph clustering techniques. Traditional hierarchical clustering methods often rely on pairwise distance metrics that fail to account for density variations and structural connectivity in data. DHC-ECS addresses these limitations by proposing a new inter-cluster similarity metric that incorporates element categorization within KNN connection subgraphs, kernel density estimation, and local connectivity. The method is evaluated against several baseline clustering algorithms, including AChameleon, RNN-DBSCAN, McDPC, and G-RMS, using heterogeneous benchmark datasets. The results demonstrate that DHC-ECS significantly outperforms these methods in clustering accuracy and robustness to parameter variations. The findings suggest that leveraging local density and graph-structured connectivity can enhance clustering performance, particularly for low-dimensional datasets, while also reducing the need for manual parameter tuning.
Methodology
The proposed DHC-ECS method combines hierarchical clustering with density-based and graph clustering approaches. It utilizes a new inter-cluster similarity metric that factors in distances, element categorization, kernel density estimation, and local connectivity within sub-clusters, enhancing the robustness of clustering outcomes.
Results
DHC-ECS outperformed baseline clustering algorithms in terms of accuracy and robustness across various datasets. The method effectively identified clusters with varying densities and structures, demonstrating its capability to handle low-dimensional data without extensive parameter tuning.
Implications
The DHC-ECS algorithm has significant implications for applications in fields such as GIS, image segmentation, and bioinformatics, where effective clustering of complex data structures is crucial. Its ability to integrate density and connectivity can lead to better pattern recognition and data analysis outcomes.
Evaluating Generative Time-Series Models on Data with Point Masses
Generative Models
Time Series
- Standard evaluation protocols may not accurately reflect model performance on datasets with point masses.
- An autoregressive occurrence hurdle model outperforms conditional flow models on five out of six datasets.
- The evaluation window can differ significantly from the dataset, leading to misleading conclusions.
- The choice of occurrence statistics can change model rankings, indicating variability in performance assessment.
Read more
Evaluating Generative Time-Series Models on Data with Point Masses
Summary
This paper investigates the performance of generative time-series models on datasets characterized by a significant presence of point masses, particularly zeros. The author highlights the inadequacies of standard evaluation protocols, such as the rolling-origin method, which can lead to misleading conclusions about model performance. The study reveals that the evaluation windows often do not reflect the underlying dataset's structure, causing discrepancies in performance metrics. A control experiment is introduced to isolate the contribution of temporal coupling to model statistics. The author benchmarks seven models across six datasets, finding that an autoregressive hurdle model significantly outperforms a conditional flow model in most cases. The results indicate that the choice of occurrence statistics can drastically affect model rankings, suggesting that continuous-state models may be systematically disadvantaged when dealing with data that has a point mass. The findings emphasize the need for careful evaluation practices in generative modeling, particularly for intermittent demand scenarios.
Methodology
The study employs a series of experiments to benchmark seven generative time-series models across six datasets. It utilizes the rolling-origin evaluation protocol and introduces a control to measure the impact of temporal coupling on performance metrics. The models are assessed based on their ability to predict time-series data characterized by point masses.
Results
The autoregressive hurdle model outperformed the conditional flow model by up to a factor of 153 across five datasets. The flow model's occurrence statistics varied by up to 62% across different training seeds, while the autoregressive model remained deterministic. The study also found that the evaluation windows did not match the dataset distributions, leading to reversed conclusions in prior analyses.
Implications
The findings suggest that researchers should reconsider the evaluation methods used for generative time-series models, particularly in contexts where point masses are prevalent. Improved evaluation practices could lead to more accurate assessments of model performance and better model selection for applications involving intermittent demand.
DreOPD: Degraded-Reference Extrapolative On-Policy Distillation for Flow-matching Models
Generative Models
Reinforcement Learning
Optimization
- DreOPD bridges the gap between reinforcement learning and on-policy distillation for flow-matching models.
- The method converts implicit reward extrapolation into closed-form velocity regression, enhancing stability.
- A mildly degraded reference is used to strengthen teacher-reference contrast, improving extrapolation.
- DreOPD shows superior performance compared to existing methods and specialized teachers in various metrics.
Read more
DreOPD: Degraded-Reference Extrapolative On-Policy Distillation for Flow-matching Models
Summary
The paper introduces DreOPD, a novel method for enhancing flow-matching models in image generation through a combination of on-policy distillation (OPD) and reward extrapolation. Traditional flow-matching models often require post-training for diverse tasks, leading to conflicts among optimization objectives. While reinforcement learning (RL) allows for direct optimization of task-specific rewards, it suffers from high variance and cross-task interference. DreOPD addresses these challenges by converting implicit reward extrapolation into a closed-form velocity regression, which stabilizes the training process. The method employs a mildly degraded reference to improve teacher-reference contrast, facilitating clearer extrapolation directions. Experimental results demonstrate that DreOPD outperforms both OPD and multi-task RL baselines, achieving superior average performance and surpassing specialized teachers on most metrics.
Methodology
DreOPD employs a closed-form target that extends flow-based OPD from teacher imitation to reward extrapolation. It utilizes shared-covariance Gaussian transitions to define a conditional transition objective at each student-visited state. By introducing controlled reference degradation, the method enhances the contrast between teacher and reference, guiding the student model towards improved performance.
Results
DreOPD achieved the highest average performance across evaluated reward metrics in both single- and multi-teacher settings, outperforming traditional OPD and multi-task RL baselines. The method also surpassed specialized teachers on most metrics, demonstrating its effectiveness in consolidating knowledge from multiple sources.
Implications
The findings suggest that DreOPD can significantly enhance the performance of flow-matching models in diverse image generation tasks, making it a valuable tool for applications requiring high-quality image synthesis and adaptability to various objectives.
Synthetic LiDAR Data Generation and Deterministic Downsampling for Point Cloud Classification on the Edge
Computer Vision
Robotics
Efficient ML
- Introduction of a synthetic LiDAR data generation method to address the reality gap in training datasets.
- Development of a Critical Points Layer (CPL) for efficient downsampling of point clouds.
- Demonstration of real-time classification capabilities on low-power hardware with high accuracy.
- Emphasis on the importance of sensor-aware training for effective point cloud classification.
Read more
Synthetic LiDAR Data Generation and Deterministic Downsampling for Point Cloud Classification on the Edge
Summary
This paper addresses the challenges of deploying deep learning frameworks for 3D point cloud classification on low-power embedded processors, specifically the Raspberry Pi 5. The authors propose a hardware-constrained workflow that includes the generation of synthetic LiDAR data using a physics-based simulation to create a repository that accounts for sensor noise, thus bridging the reality gap between clean CAD datasets and real-world data. They highlight the necessity of sensor-aware training, as traditional networks trained on CAD files exhibit significant drops in classification accuracy when applied to synthetic LiDAR data. To optimize the processing pipeline, the authors introduce a Critical Points Layer (CPL) that deterministically downsamples point clouds from 1024 points to a more manageable 40-60 points while retaining critical features for classification. This approach not only reduces latency but also enhances throughput, achieving approximately 50 FPS on the ARM Cortex-A76 processor with a classification accuracy of 88.36%. The findings demonstrate the feasibility of real-time 3D perception on edge devices, paving the way for advancements in applications such as autonomous driving and robotic navigation.
Methodology
The authors generated synthetic LiDAR data using a physics-based simulation and the BLAINDER tool, which incorporates sensor noise to create realistic datasets. They integrated the Critical Points Layer (CPL) into the PointNet architecture to perform feature-driven downsampling, reducing the number of points in a point cloud while preserving essential information for classification tasks.
Results
The proposed pipeline achieved an inference throughput of approximately 50 FPS on the Raspberry Pi 5, with a high instance classification accuracy of 88.36%. The integration of the CPL allowed for effective compression of point clouds from 1024 points to 40-60 points, demonstrating the method's efficiency and effectiveness in real-time applications.
Implications
The findings suggest that the proposed methods can significantly enhance the deployment of 3D perception systems in resource-constrained environments, such as autonomous vehicles and robotics, where real-time processing of spatial data is critical. This work opens avenues for further research into optimizing deep learning frameworks for edge computing.
From Objectives to What Models Learn: A Landau Theory of Invariant Learning
Theory
- Introduces a Landau theory perspective to unify predictive-mode learning and invariant regularization.
- Derives low-order objective signatures that predict phase boundaries and regularization behaviors.
- Validates theoretical predictions through controlled experiments on bilinear and ReLU networks.
- Offers a structural analysis of existing invariant-learning objectives, revealing method-specific behaviors.
Read more
From Objectives to What Models Learn: A Landau Theory of Invariant Learning
Summary
This paper addresses the objective-behavior gap in invariant learning, which aims to develop representations that maintain predictive power across varying environments. The authors propose a novel framework based on Landau theory, likening representation learning to multimode magnetization. They derive an effective free energy from invariant-learning objectives, where low-order coefficients act as objective signatures that dictate distinct regularization behaviors. The study reveals how effective quadratic and quartic corrections influence phase boundaries and mode behaviors, leading to predictions about model learning as regularization strength varies. Through experiments on bilinear models and ReLU networks, the authors validate their theoretical predictions regarding phase boundaries and regularization phenotypes, demonstrating that the derived signatures can effectively predict qualitative behaviors across different architectures. This framework not only organizes invariant objectives by their regularization effects but also provides a structural analysis of existing objectives, enhancing our understanding of their performance in practice.
Methodology
The authors employ a Landau-type effective free energy approach to analyze invariant-learning objectives, reducing their local equilibria to derive low-order signatures. This involves examining the effects of quadratic, quartic, and higher-order terms on phase boundaries and mode behaviors. Controlled experiments are conducted on bilinear models and one- and two-hidden-layer ReLU networks to empirically validate the theoretical predictions.
Results
The study successfully identifies closed-form phase boundaries and steady-state loadings in a canonical bilinear model. Experiments confirm the predicted behaviors, including distinct regularization phenotypes such as finite-threshold elimination and non-monotonic tails in ReLU networks. The findings indicate that the derived objective signatures are predictive of qualitative regularization-path behavior, even with depth-dependent shifts.
Implications
This framework provides a deeper understanding of invariant learning objectives, potentially guiding the design of more effective learning algorithms that maintain predictive power across diverse environments. It may also inform future research on the structural properties of learning objectives and their impact on model performance.
Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction
Computer Vision
Optimization
Time Series
- Introduces an unsupervised fire-zone segmentation algorithm that improves wildfire prediction accuracy.
- Demonstrates that spatial discretization significantly affects predictive performance, more than model choice.
- Achieves consistent performance improvements over grid-based methods with mean IoU gains of 3-6%.
- Method is computationally efficient and can be parallelized, making it suitable for real-time applications.
Read more
Data-Driven Fire-Zone Segmentation for Improved Short-Term Wildfire Prediction
Summary
This paper addresses the limitations of traditional grid-based approaches in wildfire prediction by proposing a novel unsupervised fire-zone segmentation algorithm. The authors argue that the way data is discretized significantly impacts predictive performance, often more so than the choice of model. Their method combines watershed detection with K-means clustering to create semantically meaningful fire zones based on historical ignition patterns. Experiments conducted across six French departments demonstrate that this segmentation approach consistently outperforms conventional grid-based methods, achieving mean Intersection over Union (IoU) improvements of 3-6% depending on the spatial scale. The proposed method is computationally efficient, taking less than 10 seconds per configuration and is fully parallelizable. The findings suggest that optimizing spatial discretization can lead to substantial and reproducible enhancements in short-term wildfire forecasting, opening new avenues for research in wildfire prediction methodologies.
Methodology
The proposed methodology involves three main stages: generating a continuous 3D signal of fire risk, detecting fire-prone areas using watershed segmentation, and merging these areas with K-means clustering to create prediction units that reflect historical ignition distributions. This unsupervised approach avoids the need for pre-labeled datasets, making it adaptable and efficient.
Results
The results indicate that models trained on the proposed fire-zone segmentation outperform those using standard grid-based approaches across multiple spatial scales. The improvements in mean IoU metrics validate the effectiveness of the segmentation algorithm in enhancing predictive accuracy for short-term wildfire forecasting.
Implications
The findings suggest that optimizing spatial discretization in wildfire prediction can lead to better resource allocation and early warning systems, potentially reducing the economic and humanitarian impacts of wildfires. This research opens new pathways for integrating AI with IoT sensor networks for improved wildfire management.
Momba: Network Modernization Improves Multi-Objective Reinforcement Learning
Reinforcement Learning
Robotics
Optimization
- Integration of advanced neural network architectures improves MORL performance.
- Proposed architecture enhances sample efficiency and solution quality without complex updates.
- Introduction of a categorical critic adapted for multi-objective domains.
- Empirical validation across multiple continuous control tasks shows significant performance gains.
Read more
Momba: Network Modernization Improves Multi-Objective Reinforcement Learning
Summary
This paper addresses the underexplored area of neural network architectures in multi-objective reinforcement learning (MORL), contrasting with the predominant focus on algorithmic innovations. The authors propose enhancements to the architecture of MORL algorithms by integrating advanced neural network design principles such as observation and feature normalization, weight normalization, and modeling of distributional returns. These modifications are applied to an entropy-regularized MORL algorithm. The empirical results demonstrate significant improvements in performance across standard continuous control benchmarks, showing that the proposed architecture can yield better solution sets without necessitating complex preference selection algorithms or specialized update rules. The study also introduces a novel adaptation of the categorical critic to the multi-objective domain, allowing for the direct prediction of scalarized return distributions, which is validated through extensive empirical studies. Overall, the findings suggest that modernizing neural network architectures can substantially enhance the effectiveness of MORL.
Methodology
The authors employed a combination of advanced neural network techniques, including observation and feature normalization, weight normalization, and distributional return modeling, within an entropy-regularized MORL framework. They adapted the categorical critic to predict scalarized return distributions directly, facilitating better performance in multi-objective scenarios.
Results
The results indicate that the proposed architecture, referred to as Momba, outperforms existing baselines in seven continuous control tasks, achieving higher normalized hypervolume (HV) and expected utility (EUM) metrics. The improvements were substantial, demonstrating that the architectural enhancements led to better coverage and quality of solution sets.
Implications
The findings suggest that enhancing neural network architectures can lead to significant advancements in MORL applications, potentially benefiting real-world scenarios such as robotics, medical treatment optimization, and multi-objective decision-making tasks. This work encourages further exploration of architectural innovations in reinforcement learning.
Latent Fact-Checking: Detecting Misinformation through Activation Engineering
NLP
Large Language Models
Interpretability
- Introduces a novel misinformation detection framework based on activation engineering.
- Utilizes the geometric properties of transformer models to identify misinformation directions.
- Achieves competitive performance on fact-checking benchmarks without requiring model fine-tuning.
- Demonstrates that truthfulness is a structured concept in the latent space of language models.
Read more
Latent Fact-Checking: Detecting Misinformation through Activation Engineering
Summary
The paper addresses the growing challenge of misinformation online by proposing a novel framework for misinformation detection that utilizes activation engineering within transformer models. Unlike traditional methods that rely on linguistic features or external knowledge, this approach treats truthfulness as a geometric property within the model's representation space. The authors introduce a method that identifies a 'misinformation direction' by contrasting activations from paired truthful and false statements, leveraging the difference-in-means principle of Contrastive Activation Addition (CAA). During inference, the last-token activation of a claim is projected onto this misinformation direction, and the resulting representation is classified using a Multilayer Perceptron (MLP). This method does not require fine-tuning of the backbone model, external evidence retrieval, or extensive supervision beyond the contrastive pairs used for direction estimation. The authors evaluate their approach across 11 models from various families on three fact-checking benchmarks: AVeriTeC, LIAR, and FACTors. The results indicate that the misinformation direction is recoverable across different model architectures and scales, with the last-token projection achieving competitive performance against zero-shot and few-shot prompting baselines, particularly benefiting smaller models. However, performance on AVeriTeC was limited, attributed to its evidence-grounded labeling scheme. Overall, the findings suggest that truthfulness can be effectively represented in the latent space of pretrained language models, highlighting the potential of interpretability-driven misinformation detection as a viable alternative to retrieval-based methods.
Methodology
The authors employ activation engineering to analyze the internal representations of transformer models. They compute the mean activation difference between paired truthful and false statements to derive a misinformation direction. This direction is used to project the last-token activation of unseen claims, which is then classified using a Multilayer Perceptron (MLP). The method is evaluated across various model architectures and scales on real-world fact-checking datasets.
Results
The proposed method successfully recovers the misinformation direction across different model scales and architectures. The last-token projection outperforms zero-shot and few-shot prompting baselines on the LIAR and FACTors datasets, with significant improvements noted for smaller models. However, the performance on the AVeriTeC dataset was less robust, likely due to its specific evidence-grounded labeling scheme.
Implications
This research indicates that leveraging the latent geometry of language models can enhance misinformation detection capabilities. The findings support the development of more interpretable and efficient automated fact-checking systems, which could serve as practical tools in combating misinformation in various online contexts.
CONFER: Conflict-Aware Evidence Negotiation for Regime-Calibrated Weak Supervision in Multimodal Emotion Recognition
Multimodal
Graph Learning
- CONFER is a unified framework that integrates cross-modal conflict negotiation with weak-label calibration.
- The framework utilizes a dynamic graph mechanism for effective message passing between modality experts.
- Empirical results show significant performance improvements in high-conflict scenarios and robustness against weak-label corruption.
- The approach provides a novel perspective on the utility of cross-modal conflict in multimodal emotion recognition.
Read more
CONFER: Conflict-Aware Evidence Negotiation for Regime-Calibrated Weak Supervision in Multimodal Emotion Recognition
Summary
The paper introduces CONFER, a novel framework designed to enhance weakly supervised multimodal emotion recognition by addressing the unreliability of self-reported labels and the conflicts that arise between different modalities. Traditional approaches often treat self-reported emotional states as reliable, ignoring the inherent biases and conflicts that can occur. CONFER employs a graph-based structure where each modality is represented as a node, incorporating predictive beliefs, uncertainty measures, and reliability estimates based on historical performance. The framework facilitates iterative message passing among modalities, guided by compatibility and reliability-aware edge weights, to negotiate conflicts and improve predictions. The authors categorize the negotiation outcomes into three regimes—Consensus, Dissent, and Ambiguity—to better calibrate weak-label supervision. Evaluations on multiple datasets (AMIGOS, MAHNOB-HCI, and DEAP) demonstrate that CONFER achieves competitive accuracy and robustness, particularly in high-conflict scenarios, suggesting that cross-modal conflict can be beneficial for improving modality coordination and supervision reliability.
Methodology
CONFER employs a graph-based architecture where each modality expert is represented as a node. It calculates predictive beliefs and uncertainty scores, and uses historical performance to estimate reliability. The framework facilitates message passing between nodes using reliability-directed asymmetric edge weights, allowing for conflict negotiation and peer-supported predictions. The outcomes are categorized into three regimes to calibrate weak-label supervision effectively.
Results
CONFER achieved an accuracy of 0.873 on the AMIGOS-V dataset and 0.854 on the MAHNOB-V dataset under strict leave-one-subject-out evaluation. The framework demonstrated larger gains in performance on high-conflict samples and showed improved robustness to weak-label corruption, indicating its effectiveness in leveraging cross-modal conflict for better predictions.
Implications
The findings suggest that CONFER can enhance the reliability of emotion recognition systems in real-world applications, where self-reported data may be unreliable. This approach can be applied in various domains, including mental health monitoring, human-computer interaction, and affective computing, where accurate emotion detection is crucial.
Learning Suffers More Than the Policy Class Under Partial Observability: A Closed-Form Analysis
Reinforcement Learning
Theory
- The learning gap, not the policy-class gap, is the primary issue in reinforcement learning under partial observability.
- The critic's misinterpretation of unexplained variations leads to biased value estimates, causing the actor to deviate from optimal policies.
- Closed-form expressions for the policy and its cost provide precise insights into the learning dynamics.
- Adjusting the lookahead parameter can effectively address the learning gap, while adding memory does not yield improvements.
Read more
Learning Suffers More Than the Policy Class Under Partial Observability: A Closed-Form Analysis
Summary
This paper investigates the challenges faced by reinforcement learning (RL) agents operating under partial observability, specifically distinguishing between the policy-class gap and the learning gap. The author demonstrates that even when a good policy exists and the agent's value function can represent it, learning can still fail significantly due to biases in the critic's learning process. The study focuses on a partially observed linear-quadratic problem, where a standard actor-critic learner is analyzed in closed form. The findings reveal that the learning gap is more detrimental than the policy-class gap, with the agent settling on a policy that is 35% worse than the optimal one it could represent. The paper identifies the root cause of this issue as the critic misinterpreting unexplained variations in observations, leading to incorrect value estimates. The author derives closed-form expressions for the resulting policy and its cost, and highlights a design choice regarding the lookahead parameter that can mitigate the learning gap. Empirical results from deep reinforcement learning experiments support the theoretical predictions, indicating that memory of past observations does not improve learning outcomes, while adjusting the lookahead parameter does.
Methodology
The study employs a closed-form analysis of a partially observed linear-quadratic problem, utilizing a standard actor-critic framework. The author computes the critic's fixed point and the actor's expected updates in closed form, allowing for precise identification of the learning dynamics and the factors contributing to the learning gap.
Results
The analysis reveals that the best memoryless policy is only 10.4% worse than the optimal policy, yet the learning process results in a policy that is 35% worse than the best memoryless policy. The paper provides exact formulas for the equilibrium gain at which learning stabilizes and the associated costs, demonstrating that the learning gap dominates the policy-class gap.
Implications
The findings suggest that improving RL agents' performance under partial observability may require focusing on the learning process and the biases inherent in value estimation rather than solely enhancing policy representation. This has implications for the design of RL algorithms and architectures, particularly in environments where full state information is unavailable.
Stochastic Autoregressive Learning
NLP
Large Language Models
Theory
- Introduces a PAC-learning model for binary stochastic autoregressive learning.
- Generalizes deterministic autoregressive learning frameworks to include stochastic elements.
- Establishes sample complexity bounds for different forms of supervision: base, CoT, and e2e.
- Demonstrates fundamental differences between stochastic and deterministic learning paradigms.
Read more
Stochastic Autoregressive Learning
Summary
This paper introduces a PAC-learning model for binary stochastic autoregressive learning, motivated by the behavior of large language models (LLMs) that generate outputs by sampling from next-token distributions. The authors generalize the deterministic autoregressive learning framework by defining a model where a fixed generator assigns a Bernoulli next-token distribution to each prompt string. The learning process involves sampling tokens iteratively for a specified number of steps (M), with three forms of supervision: base one-step samples, chain-of-thought (CoT) samples revealing full trajectories, and end-to-end (e2e) samples showing only the final token. The study reveals that stochastic autoregressive learning differs fundamentally from deterministic learning, with no universal comparison between the three tasks at scale ε. The authors derive bounds on the sample complexity required for learning in each supervision model and demonstrate that CoT learning is bounded by base learning at a modified scale, while e2e learning is bounded by a combination of CoT learning and the number of steps M. The paper also explores logistic functions in the model, providing upper bounds for e2e learning and establishing the necessity of these dependencies.
Methodology
The authors define a stochastic autoregressive learning model where a generator assigns a probability distribution over the next token for each prompt. They analyze three supervision models (base, CoT, and e2e) and derive sample complexity bounds under squared loss error. The methodology includes theoretical proofs and comparisons of learning tasks, focusing on logistic functions and the implications of stochasticity in learning.
Results
The paper shows that the sample complexities for CoT and e2e learning can be significantly larger than for base learning, with specific upper bounds established for each case. CoT learning is bounded by base learning at a modified scale, while e2e learning is bounded by a combination of CoT learning and the number of steps M. The authors also prove the necessity of these bounds and provide insights into the learning dynamics of logistic functions.
Implications
The findings have significant implications for the design and understanding of large language models, particularly in how they learn and generate language. The results suggest that different forms of supervision can lead to varying complexities in learning, which could influence future research on model training and architecture.
Graph Machine: Exploring Edge Mechanisms as an Inductive Bias
Graph Learning
- Graph Machine introduces edge-augmented attention and edge-centric referral mechanisms.
- The architecture allows for dynamic construction of relational graphs, enhancing reasoning capabilities.
- In Sudoku experiments, Graph Machine outperforms traditional Transformer models.
- The model effectively learns to construct higher-level relations relevant to the task.
Read more
Graph Machine: Exploring Edge Mechanisms as an Inductive Bias
Summary
The paper introduces Graph Machine, a novel architecture that incorporates explicit edge-based mechanisms to enhance reasoning capabilities in machine learning models. Unlike traditional Transformers that rely heavily on global content-based matching, Graph Machine employs two key mechanisms: edge-augmented attention, which allows edges to influence attention between nodes, and edge-centric referral, enabling nodes to share addresses to update their edges. This architecture facilitates the dynamic construction and revision of relational graphs across layers, promoting iterative traversal of latent relations. The effectiveness of Graph Machine is evaluated through controlled experiments on Sudoku, where it outperforms Transformer baselines. Ablation studies and mechanistic analyses reveal that the gains are primarily due to the edge mechanisms, which allow the model to discover a compact edge-based structure relevant to Sudoku geometry. The findings suggest that explicit edge mechanisms can serve as a beneficial inductive bias for structured reasoning tasks.
Methodology
The Graph Machine architecture operates on a graph with nodes and edges, utilizing edge-augmented attention to update node features and edge-centric referral to construct new edges. The model maintains edge features and addresses, allowing for a flexible representation of relational structures. Controlled experiments were conducted using Sudoku to evaluate performance against Transformer baselines, with mechanistic analyses to understand the contributions of the edge mechanisms.
Results
Graph Machine demonstrated superior performance in Sudoku tasks compared to same-scale and prior-advantaged Transformer models. The model was able to construct higher-level Sudoku relations through a compact edge-based approach, indicating that it was not merely fitting solutions but actively building relevant latent structures.
Implications
The findings suggest that incorporating explicit edge mechanisms can enhance the reasoning capabilities of machine learning models, particularly in structured tasks. This could lead to advancements in various applications requiring complex relational reasoning, such as graph-based learning and problem-solving in combinatorial domains.
Fairness in Link Prediction Beyond Demographic Parity: A Reproducibility Study
Graph Learning
- Demographic parity can obscure within-group exposure biases in link prediction.
- Normalized Discounted KL-divergence (NDKL) is a more effective, rank-aware fairness metric.
- The MORAL algorithm successfully reduces exposure biases while preserving utility.
- The study extends previous work by testing robustness across different homophily settings and sensitive attribute types.
Read more
Fairness in Link Prediction Beyond Demographic Parity: A Reproducibility Study
Summary
This paper addresses the limitations of demographic parity (∆DP) as a fairness metric in link prediction, particularly in the context of social networks. The authors reproduce findings from Mattos et al. (2025), which argue that ∆DP fails to account for exposure bias by ignoring the ranking of links. They introduce a rank-aware metric, Normalized Discounted KL-divergence (NDKL), which effectively identifies disparities in link rankings among different subgroup pairs. The study also evaluates the MORAL post-processing algorithm, which enhances exposure-based fairness while maintaining competitive utility. The authors extend the original research by testing the robustness of these methods across various settings, including synthetic homophily scenarios and categorical sensitive attributes. Their findings confirm that exposure-based metrics reveal biases that ∆DP overlooks and demonstrate that MORAL can mitigate these biases with minimal impact on utility across diverse datasets.
Methodology
The authors reproduce the experimental setup of Mattos et al. (2025) and extend it by conducting a robustness analysis using synthetic graphs, varying homophily levels, and evaluating categorical sensitive attributes. They employ NDKL to assess fairness and Precision@K to measure utility, ensuring a comprehensive evaluation of the MORAL algorithm's performance.
Results
The results indicate that ∆DP fails to detect significant subgroup-pair exposure disparities, while NDKL successfully identifies these biases. The MORAL algorithm consistently mitigates these biases across various datasets and settings, achieving strong fairness-utility trade-offs.
Implications
The findings suggest that relying solely on demographic parity can lead to unfair outcomes in link prediction tasks, particularly in sensitive applications like social network recommendations. The introduction of NDKL and MORAL provides practitioners with robust tools to enhance fairness in machine learning models, potentially improving equity in resource allocation and decision-making processes.
Capacity Confounds and Coverage Guarantees in Adaptive Sub-model Federated Learning
Federated Learning
- Client heterogeneity estimates are confounded by device capacity, not just data differences.
- Adaptive allocation can lead to model corruption if clients are under-capacitated.
- A coverage guarantee can prevent failure modes in adaptive sub-model training.
- Random capacity allocation may perform similarly to heterogeneity-aware strategies.
Read more
Capacity Confounds and Coverage Guarantees in Adaptive Sub-model Federated Learning
Summary
This paper explores the challenges of adaptive sub-model federated learning (FL), particularly focusing on how capacity allocation can be improved by considering client data heterogeneity. The authors introduce HAS-FL, an adaptive capacity-allocation framework, and investigate whether client data heterogeneity can be accurately estimated from observed updates. The study reveals three significant findings: first, the estimates of client heterogeneity are primarily influenced by device capacity rather than actual data differences, indicating a confound that affects existing methods. Second, the authors identify a failure mode in adaptive allocation where clients with limited capacity can lead to model corruption due to untrained parameters. A coverage guarantee is proposed to mitigate this issue. Lastly, the research shows that random allocation of capacity can perform comparably to heterogeneity-aware policies, suggesting that the benefits of adaptive allocation stem more from effective capacity budgeting and parameter coverage than from accurate heterogeneity estimation. Overall, the findings challenge the current understanding of heterogeneity-aware sub-model allocation and highlight the need for future designs to differentiate between capacity effects and true data signals.
Methodology
The authors employed a systematic analysis of the HAS-FL framework, validating their findings against ground-truth label-distribution divergence across multiple datasets. They utilized corrected estimators to assess the relationship between client heterogeneity and device capacity, and conducted experiments to compare the performance of different capacity allocation strategies.
Results
The study found that client heterogeneity estimates were negatively correlated with device capacity, indicating that capacity confounds the estimation process. The introduction of a coverage guarantee successfully mitigated the identified failure mode in adaptive allocation. Additionally, the matched-budget control experiments demonstrated that random allocation could yield similar performance to heterogeneity-aware policies, particularly in image benchmarks.
Implications
These findings suggest a reevaluation of how adaptive capacity allocation is approached in federated learning, emphasizing the importance of parameter coverage over mere heterogeneity estimation. This could lead to more robust federated learning systems that effectively include resource-constrained clients without compromising model quality.
Risk-Aware Decision Policies for Agents Under Noisy Perception
Robotics
Theory
Reinforcement Learning
- Noisy perception significantly affects decision-making in predator-prey scenarios.
- Blindly trusting perceptual labels can lead to catastrophic failures.
- Uncertainty-aware strategies improve survival rates and reduce errors.
- Agents exhibit behavioral regime shifts from exploration to conservatism with increased uncertainty.
Read more
Risk-Aware Decision Policies for Agents Under Noisy Perception
Summary
This paper explores the impact of noisy perception on decision-making in artificial life agents within a predator-prey model. It highlights the challenges organisms face when making decisions under uncertainty, where misclassification can have severe consequences. The study compares various decision-making policies: a blind strategy that relies solely on current perceptual labels, a skeptical strategy that acts only when confidence in the label is high, and a verify strategy that gathers additional information before making decisions. Through simulations, the research demonstrates that blindly trusting perceptual labels leads to significant failures as noise increases, while uncertainty-aware strategies enhance survival and reduce fatal errors. The findings reveal qualitative shifts in agent behavior, transitioning from exploration to conservatism as uncertainty rises. This work connects ecological decision-making with artificial life, emphasizing the importance of incorporating uncertainty into decision policies to improve robustness in environments characterized by unreliable perception.
Methodology
The research employs a simulation-based approach, modeling agents in a two-dimensional foraging environment where they receive noisy signals about food and predator identities. Three decision strategies are compared: blind, skeptical, and verify, allowing for an analysis of how these strategies perform under varying levels of perceptual noise.
Results
The results indicate that as perceptual noise increases, agents relying on blind strategies experience significant performance degradation, while those using skeptical or verify strategies demonstrate improved survival and reduced fatal errors. The study identifies clear behavioral regime shifts, with agents adapting their strategies based on the level of uncertainty in their perceptions.
Implications
The findings suggest that incorporating uncertainty into decision-making processes can enhance the robustness of agents in artificial life and may inform approaches in machine learning where label noise is prevalent. This research could influence the design of more resilient AI systems that better handle uncertain information.
CrystalGRPO: Target-Aligned and Coverage-Preserving Reinforcement Learning for Flow-Based Crystal Structure Prediction
Reinforcement Learning
Generative Models
- Introduces CrystalGRPO, a CSP-aligned reinforcement learning framework for flow-based crystal generators.
- Combines predicted energy with a StructureMatcher-based recovery score to improve target polymorph identification.
- Offers two operational modes to balance between single-draw accuracy and candidate coverage.
- Demonstrates significant improvements in recovery rates compared to existing methods.
Read more
CrystalGRPO: Target-Aligned and Coverage-Preserving Reinforcement Learning for Flow-Based Crystal Structure Prediction
Summary
The paper presents CrystalGRPO, a novel post-training framework for crystal structure prediction (CSP) that enhances flow-based generative models by aligning them with target recovery objectives and preserving candidate coverage. Traditional flow-based models generate candidate crystal structures but do not optimize for the specific polymorphs required in CSP. Existing reinforcement learning (RL) approaches primarily focus on energy rewards, which can mislead the model towards low-energy non-target structures. CrystalGRPO addresses this by integrating a hybrid reward system that combines predicted energy with a StructureMatcher-based recovery score, allowing for better identification of the target polymorph. The framework operates in two modes: CrystalGRPO-Q, which emphasizes single-draw recovery, and CrystalGRPO-C, which maintains coverage across multiple candidates. The authors demonstrate that both modes significantly reduce root mean square error (RMSE) in candidate predictions compared to traditional coordinate-only reinforcement methods, achieving improved Top-1 and Top-20 recovery rates across various datasets.
Methodology
CrystalGRPO extends existing stochastic policy constructions for flow models to a joint coordinate-lattice state, allowing for more accurate candidate generation. It employs a hybrid reward system that integrates physics-informed energy predictions with a recovery score based on structural matching. The framework includes mechanisms for coverage preservation to ensure a diverse set of candidates while optimizing for target recovery.
Results
The evaluation of CrystalGRPO on datasets MP-20 and MPTS-52 shows that both CrystalGRPO-Q and CrystalGRPO-C significantly reduce one- and twenty-sample RMSE compared to coordinate-only reinforcement methods. CrystalGRPO-Q consistently improves Top-1 recovery rates, while CrystalGRPO-C achieves higher Top-20 recovery rates across all tested settings.
Implications
The proposed framework has the potential to enhance the efficiency and accuracy of crystal structure prediction in materials science, enabling better identification of polymorphs and facilitating the discovery of new materials with desired properties.
Label-Free Parkinson's Disease Screening from Face and Voice through Mechanistic Interpretability
Multimodal
Audio & Speech
Computer Vision
- Introduces a label-free screening method for Parkinson's Disease using face and voice data.
- Utilizes frozen pretrained encoders, avoiding the need for PD labels during model training.
- Establishes the alignment principle to explain the effectiveness of synthetic-degradation detection.
- Achieves an AUROC of 0.802 and NPV of 0.95 on the YouTubePD benchmark.
Read more
Label-Free Parkinson's Disease Screening from Face and Voice through Mechanistic Interpretability
Summary
This paper addresses the challenge of screening for Parkinson's Disease (PD) using machine learning methods that do not rely on labeled data, which is often scarce due to privacy concerns and the need for expert annotation. The authors propose a novel approach that utilizes frozen pretrained encoders for both face and voice modalities, specifically a Vision Transformer for facial expressions and HuBERT for voice. The methodology involves constructing a synthetic-dysarthria contrastive activation addition (CAA) direction for voice, derived from healthy speech degradation, and employing a k-nearest-neighbor anomaly score for the face modality. The central contribution is the introduction of the alignment principle, which posits that the effectiveness of the synthetic-degradation CAA detector is contingent upon the cosine similarity between synthetic and real disease directions being positive. The proposed model was evaluated on the YouTubePD benchmark, achieving an area under the receiver operating characteristic curve (AUROC) of 0.802 with a negative predictive value (NPV) of 0.95, indicating its potential for rule-out triage in clinical settings. The results suggest that while the voice modality shows promising transferability, the face modality's performance may be overly optimistic and requires further validation.
Methodology
The authors developed a label-free screening framework that constructs scoring functions for both voice and face modalities using only control data. The voice modality employs a synthetic-dysarthria CAA direction created from healthy speech degradation, while the face modality uses a k-nearest-neighbor anomaly score against a control embedding cluster. The alignment principle is introduced as a post-hoc analysis to assess the effectiveness of the detection methods.
Results
The voice modality achieved a cosine similarity of +0.37, resulting in an AUROC of 0.765, while the face modality had a cosine similarity of -0.48, with the anomaly detection yielding an AUROC of 0.751. The combined late fusion of both modalities reached an AUROC of 0.802, indicating strong performance in distinguishing PD from healthy controls.
Implications
This research has significant implications for the early detection of Parkinson's Disease, offering a non-invasive, label-free screening method that could be implemented in clinical settings. The findings suggest that machine learning can effectively leverage existing healthy data to identify deviations indicative of PD, potentially improving patient outcomes through earlier intervention.
PRISM: Principled Reference Identification for Schrodinger Bridge Model
Generative Models
Theory
Computer Vision
- PRISM provides a theoretical foundation for reference design in Schrödinger bridge models.
- An invisibility principle shows that the reference choice is irrelevant in the ideal limit but matters under finite resources.
- The optimal noise spectrum is derived and shown to be proportional to the destroyed information spectrum.
- Experiments confirm theoretical predictions and highlight the limitations of Gaussian assumptions in real image processing.
Read more
PRISM: Principled Reference Identification for Schrodinger Bridge Model
Summary
The paper introduces PRISM, a theoretical framework for designing reference processes in Schrödinger bridge models, which are used to restore clean signals from degraded observations. Traditionally, references have been chosen heuristically, often as white noise with hand-tuned schedules. PRISM characterizes time-varying Gaussian references that maintain tractability and establishes an invisibility principle, indicating that the choice of reference is only significant under finite computational resources. The authors derive a closed-form finite-step objective and demonstrate that the optimal noise spectrum is proportional to the spectrum of information destroyed by the sensor. Through experiments, the authors validate their theoretical predictions regarding noise levels and reference design, revealing that while white noise often outperforms matched references in certain settings, the non-Gaussian statistics of real images complicate the idealized assumptions of the model. Overall, PRISM transforms reference design from a hyperparameter tuning process into a principled calculation, identifying the limitations of Gaussian assumptions in real-world applications.
Methodology
The authors develop a solvable linear-Gaussian degradation model to analyze the performance of different reference processes. They derive theoretical results regarding the tractability of time-varying Gaussian references, establish an invisibility principle, and formulate a closed-form expression for the finite-step terminal law. Experiments are conducted in Gaussian settings to validate the theoretical findings.
Results
The study finds that the optimal reference noise level is predictable and varies with the solver budget. The derived closed-form expressions for the finite-step objectives align with experimental results, confirming the theoretical predictions. Additionally, the experiments reveal that while white noise often performs better than matched references, the non-Gaussian characteristics of real images lead to deviations from the model's idealized assumptions.
Implications
The findings suggest that reference design in generative models can be systematically approached rather than relying on heuristic methods. This has potential applications in improving the performance of signal restoration and generative tasks in various domains, particularly in computer vision.
Fixed and Adaptive Topological DeepONets: Functional Measurements on Hausdorff Locally Convex Spaces
Theory
Efficient ML
Interpretability
- Introduces Fixed and Adaptive Topological DeepONets for functional measurements in locally convex spaces.
- Replaces point samples with continuous linear functionals for more efficient operator learning.
- Demonstrates the ability to learn operators in non-normable input spaces.
- Achieves significant error reduction in various applications, including fluid dynamics.
Read more
Fixed and Adaptive Topological DeepONets: Functional Measurements on Hausdorff Locally Convex Spaces
Summary
This paper presents Fixed and Adaptive Topological DeepONets, an advancement in the framework of Deep Operator Networks (DeepONets) that allows for functional measurements in Hausdorff locally convex spaces. Traditional DeepONets rely on fixed discretization and point samples, which can be inefficient for certain operator learning tasks. The authors propose a method that utilizes continuous linear functionals from the continuous dual of the input space, enabling a more flexible and efficient representation of input functions. The framework includes both fixed and adaptive measurement systems, integrating a Two-Step procedure for stability and regularization. The authors derive a discrete error decomposition to analyze the performance of their approach and evaluate it on various operators, including the antiderivative operator and Navier–Stokes vorticity operators. The results demonstrate that the Adaptive Topological DeepONet achieves superior accuracy compared to traditional methods, particularly in scenarios where the input topology is generated by seminorms rather than a single norm.
Methodology
The authors develop two instantiations of Topological DeepONets: a fixed version using global functionals and an adaptive version that learns measurement functionals from data. They employ a Two-Step procedure to construct a stable output basis and train the branch network to predict coefficients, while also deriving a discrete error decomposition to analyze performance.
Results
The Adaptive Topological DeepONet outperforms traditional DeepONets, achieving a mean relative L2 error of 1.685% ± 0.017% with only 128 functional coordinates in the fixed-time Navier–Stokes problem. In the heterogeneous Darcy problem, the method maintains errors around 5.5%–5.6% on unseen grids, showcasing its robustness across different scenarios.
Implications
The proposed framework allows for more efficient and flexible operator learning in computational science and engineering, particularly in applications involving nonlinear operators and varying input resolutions. It opens avenues for further research in functional approximation and operator learning in complex spaces.
Optimal Learning Under Tsybakov Noise
Theory
- The paper resolves a 20-year open question regarding the gap between upper and lower bounds for learning under Tsybakov noise.
- An adaptive partitioning method is introduced to handle varying noise levels in the instance space.
- The proposed algorithm achieves optimal error guarantees, matching the best-known lower bound.
- The work builds on and connects with recent advances in non-realizable learning.
Read more
Optimal Learning Under Tsybakov Noise
Summary
This paper addresses a significant gap in the understanding of learning algorithms under Tsybakov noise, a model that allows for varying levels of label noise in binary classification tasks. The authors improve upon previous upper bounds for error guarantees in the Probably Approximately Correct (PAC) learning framework, specifically under the Tsybakov noise model. They present a learning algorithm that adaptively partitions the instance space into regions based on noise levels, ensuring that the hypothesis produced meets specific error constraints for each region. This work resolves a long-standing open question by matching the upper bound with the best-known lower bound, thus establishing optimal error guarantees for learning in the presence of Tsybakov noise. The findings contribute to a deeper understanding of learning rates and error bounds in non-realizable settings, enhancing the theoretical foundations of machine learning.
Methodology
The authors develop a learning algorithm that adaptively partitions the instance space into regions corresponding to different noise levels. This approach allows the algorithm to return a hypothesis that satisfies specific error constraints for each region, thereby optimizing the learning process under Tsybakov noise.
Results
The main result is the establishment of optimal error guarantees for learning under Tsybakov noise, with the improved upper bound matching the previously known lower bound. This resolves a significant theoretical gap in the literature and provides a robust framework for understanding learning rates in the presence of label noise.
Implications
The findings have important implications for the design of learning algorithms in real-world scenarios where label noise is prevalent. By providing optimal error guarantees, this work can inform the development of more robust machine learning systems that can better handle noisy data, potentially improving performance in various applications such as classification tasks in fields like healthcare, finance, and social media.
SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction
Graph Learning
Efficient ML
Optimization
- SNI-GNN significantly reduces communication overhead in full-graph GNN training by predicting embeddings in-network.
- The system achieves 1.3-3.6x speedups over existing methods while maintaining high accuracy.
- Theoretical foundations ensure controlled predictor bias and convergence despite using inexact gradients.
- SNI-GNN scales efficiently to 16 GPUs and handles graphs with tens of millions of edges.
Read more
SNI-GNN: SmartNIC-Assisted Full-Graph GNN Training with In-Network Embedding Prediction
Summary
The paper presents SNI-GNN, a SmartNIC-assisted system designed to enhance the training of full-graph Graph Neural Networks (GNNs) in distributed multi-server environments. Full-graph training is known for its high accuracy but suffers from significant communication overhead due to the need for frequent inter-node embedding exchanges. SNI-GNN addresses this issue by implementing a lightweight linear-trend predictor on SmartNICs to forecast remote embeddings in-network, thereby reducing the volume of communication required. The system employs an importance-based boundary-node sampling policy and an asynchronous data pipeline that allows for intermediate-result reuse, optimizing the training process. The authors provide theoretical guarantees regarding error and convergence, demonstrating that the predictor bias is controlled and that the system achieves standard non-convex convergence even with inexact gradients. The implementation on NVIDIA BlueField-3 shows that SNI-GNN can reduce communication by 21-45% and achieve end-to-end speedups of 1.3-3.6x over existing methods while maintaining accuracy loss below 0.01. The results indicate that SmartNIC-based in-network prediction is a viable solution for communication-efficient full-graph GNN training at scale.
Methodology
SNI-GNN employs a SmartNIC-assisted architecture that utilizes a lightweight linear-trend predictor for in-network embedding prediction, coupled with boundary-node importance sampling and an asynchronous DPU-GPU pipeline to optimize communication and computation overlap.
Results
The implementation of SNI-GNN on NVIDIA BlueField-3 demonstrated a 21-45% reduction in inter-node communication and achieved end-to-end speedups of 1.3-3.6x compared to BNS-GCN and a 1.29x speedup over the SANCUS baseline, with an accuracy loss of less than 0.01.
Implications
The findings suggest that SmartNIC-based in-network prediction can effectively complement existing communication optimization techniques in large-scale GNN training, making it a practical solution for handling large graphs in distributed environments.
KReF: Training-Free Retrieval for Long-Term Time-Series Forecasting and Predictive Uncertainty
Time Series
- KReF is a training-free retrieval framework for long-term time-series forecasting.
- It constructs a query-local empirical predictive distribution from historical data.
- KReF achieves the lowest CRPS across multiple benchmarks and outperforms some trained models.
- The framework adapts interval boundaries using validation-selected expansion and shrinkage rates.
Read more
KReF: Training-Free Retrieval for Long-Term Time-Series Forecasting and Predictive Uncertainty
Summary
The paper introduces KReF, a novel training-free retrieval framework designed for long-term time-series forecasting (LTSF) and predictive uncertainty estimation. Traditional probabilistic forecasting methods often rely on trained models, which can be limited by the need for parameter optimization and delayed feedback in sequential predictions. KReF addresses these issues by utilizing historical lookback-future pairs to create a query-local empirical predictive distribution. The framework preprocesses lookbacks using handcrafted statistics or frozen random Fourier features, retrieves similar historical pairs, and uses their similarity weights to define predictive masses, quantiles, and point forecasts. KReF also constructs a probability-integral-transform map based on the observed query lookback and applies validation-selected rates to adjust interval boundaries. The authors demonstrate KReF's effectiveness across six LTSF benchmarks, achieving the lowest Continuous Ranked Probability Score (CRPS) in all dataset-embedding settings and surpassing trained baselines in point forecasts on two datasets. The findings suggest that retrieval-based methods can serve as a powerful inductive bias for LTSF, offering a promising alternative to traditional training-dependent approaches.
Methodology
KReF preprocesses historical lookbacks and embeds them using handcrafted statistics or frozen random Fourier features. It retrieves similar historical lookback-future pairs and uses their similarity weights to define a local empirical predictive distribution. The framework constructs a probability-integral-transform map based on the observed query lookback and adjusts interval boundaries through validation-selected rates.
Results
KReF achieved the lowest CRPS in all 12 dataset-embedding settings and the lowest IS90 in 9 settings across six LTSF benchmarks. Its point forecasts matched or surpassed trained baselines on two out of six datasets. An archive-oracle analysis indicated significant potential for improvement through finer-grained retrieval.
Implications
The findings suggest that training-free retrieval methods like KReF can effectively address challenges in long-term time-series forecasting and predictive uncertainty, potentially leading to more robust and efficient forecasting systems in various applications.
Bootstrap-Conditioned Action Selection with Tabular Foundation Models
Reinforcement Learning
Efficient ML
Theory
- BC-ICL combines bootstrap resampling with frozen in-context learning models for effective action selection in contextual bandits.
- The arm-context conditioning architecture enhances exploration by sharing statistical strength across actions.
- Empirical results indicate that BC-ICL outperforms established contextual bandit baselines under strict online protocols.
- The approach effectively addresses cold-start and data-efficiency issues in personalized recommendation systems.
Read more
Bootstrap-Conditioned Action Selection with Tabular Foundation Models
Summary
This paper addresses the challenges of deploying contextual bandits for personalized recommendation systems, particularly in scenarios with sparse and biased interaction data. The authors introduce BC-ICL (Bootstrap-conditioned action selection using In-Context Learning), a novel approach that leverages pretrained tabular foundation models to create randomized policies for online decision-making. BC-ICL operates by drawing bootstrap resamples of interaction history, conditioning a frozen ICL model on these resamples, scoring all possible actions, and selecting the one with the highest score. This method promotes shared statistical strength across actions through an arm-context conditioning architecture, which mitigates common issues faced by isolated-arm bandits. The empirical evaluation demonstrates that BC-ICL significantly outperforms traditional linear, kernelized, and neural contextual bandit methods, especially in early rounds of interaction, showcasing its effectiveness in real-world applications.
Methodology
The methodology involves a stochastic contextual bandit framework where at each round, a bootstrap resample of the interaction history is drawn. A frozen pretrained ICL predictor is conditioned on this resample to score actions. The arm-context representation is enhanced through a multiplicative feature map that captures interactions between context and actions, facilitating shared exploration. The selected action is the one with the highest predicted score from the conditioned model.
Results
The empirical evaluation shows that BC-ICL achieves strong regret performance across standard contextual bandit suites, outperforming linear, kernelized, and neural bandit methods. The results highlight the effectiveness of the bootstrap conditioning and arm-context architecture in enhancing exploration and decision-making.
Implications
The findings suggest that BC-ICL can be effectively utilized in real-world recommendation systems, particularly in scenarios with limited data. Its ability to leverage pretrained models for exploration can lead to more efficient and personalized user experiences in various online decision-making applications.
When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition
Large Language Models
Theory
Optimization
- Defined a functional interaction surface that separates non-additivity from effects along each axis.
- Demonstrated transfer beyond original task pairs and parameterization through prospective tests.
- Identified boundaries relevant to deployment, emphasizing the dependence of public transfer on input format.
- Established that continuous non-additivity cannot replace evaluation based on intended prompts.
Read more
When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition
Summary
This paper investigates the concept of task arithmetic in machine learning, which treats fine-tuning displacements as composable directions in weight space. The authors aim to clarify when parameter addition leads to predictable changes in model function by separating parameter geometry from functional geometry. They measure pairwise functional non-additivity over a two-dimensional task-vector surface using a first-token predictive-distribution interaction ratio, conditioned on input distribution and evaluated with norm-matched controls. The study employs the Qwen2.5-1.5B model to analyze interactions between task vectors, revealing that certain combinations, such as code and safety prompts, exhibit more non-additivity than others, like code and math prompts. The authors conduct prospective tests on unseen task pairs and targeted stress audits, confirming that functional interference is influenced by both the task pair and the input distribution. They establish that while weight-space composition can support functional statements across various adaptation methods and scales, it is not a universal predictor of performance. The findings highlight the importance of input format on public transfer and the necessity of evaluating models on intended prompts and behavioral metrics.
Methodology
The authors utilized a two-dimensional composition surface to diagnose functional non-additivity by comparing actual output distributions with additive predictions. They conducted norm-matching experiments, prospective transfer tests, and targeted stress audits to assess the interactions of task vectors across different input distributions and model architectures.
Results
The study found that the combination of code and safety prompts showed significantly higher non-additivity compared to code and math prompts. All eight comparisons of unseen task pairs had the predicted sign, and the functional contrast persisted across various parameterizations and models. However, the results indicated that public transfer is format-dependent, with certain wrappers collapsing the expected contrasts.
Implications
The findings suggest that while task vectors can be composed in weight space, their effectiveness is contingent on the specific tasks and input formats used. This has implications for model deployment and adaptation strategies, highlighting the need for careful evaluation based on intended use cases.
FreSH: Frequency-Segmented Hierarchical Multi-Expert Framework for Multivariate Time Series Classification
Time Series
Optimization
Efficient ML
- FreSH introduces a frequency-aware modeling paradigm for MTSC, allowing for adaptive multi-scale analysis.
- The framework combines localized specialization with global context modeling to enhance representational capacity.
- An adaptive gating mechanism dynamically emphasizes the most informative components of the input data.
- A new polynomial loss function improves learning stability across varying sample difficulties and class distributions.
Read more
FreSH: Frequency-Segmented Hierarchical Multi-Expert Framework for Multivariate Time Series Classification
Summary
The paper introduces FreSH, a novel Frequency-Segmented Hierarchical Multi-Expert Framework aimed at improving Multivariate Time Series Classification (MTSC). The authors identify key challenges in existing MTSC approaches, such as difficulties in capturing complex temporal patterns, class imbalance, and high computational costs. FreSH addresses these issues by enabling adaptive, multi-scale analysis of temporal signals through a combination of localized specialization and holistic context modeling. The framework utilizes a multi-expert system where local experts focus on specific frequency segments while a global expert captures overall context. An adaptive gating mechanism dynamically fuses the outputs of these experts, enhancing the model's flexibility and representational capacity. Additionally, the authors propose a polynomial loss function to improve learning stability across varying sample difficulties and class distributions. Extensive evaluations on 30 UEA benchmark datasets and real-world vibration data demonstrate that FreSH outperforms state-of-the-art methods in classification accuracy while significantly reducing model size and computational overhead.
Methodology
FreSH employs a hierarchical multi-expert architecture that includes dedicated local experts for specific frequency segments and a global expert for overall context. The framework transforms time series data into the frequency domain, where the Multi-Segment Expert Module learns segment-wise patterns, and the Global Expert Module captures full-spectrum dependencies. An adaptive gating mechanism is used to fuse these outputs for final predictions. The authors also introduce a polynomial loss function to optimize the model's performance across different sample difficulties and class distributions.
Results
FreSH was evaluated on 30 UEA benchmark datasets and real-world vibration data, demonstrating superior classification accuracy compared to state-of-the-art methods. The framework also achieved significant reductions in model size and computational efficiency, making it suitable for real-time applications.
Implications
The proposed FreSH framework has potential applications in various fields that rely on multivariate time series data, including healthcare, industrial equipment fault diagnosis, and human action recognition. Its ability to effectively handle class imbalance and varying sample difficulties makes it a robust choice for real-world scenarios.
Multi-Agent AI Safety as an Institutional Design Problem
Theory
- The paper presents a novel approach to AI safety by framing it as an institutional design problem.
- Different components of algorithmic institutions, such as rules and authority states, significantly affect agent behavior and safety outcomes.
- Both constitutional prompts and provenance-aware guards can achieve zero realized violations, but they operate differently in handling unsafe proposals.
- The study reveals that the same final safety counts can result from different mechanisms, emphasizing the complexity of safety in multi-agent systems.
Read more
Multi-Agent AI Safety as an Institutional Design Problem
Summary
This paper investigates the safety of multi-agent AI systems through the lens of institutional design, focusing on how different components of an algorithmic institution can influence safety outcomes. The research is part of the POLIS program and employs a frozen 5,280-episode study suite to analyze the effects of delegation rules and authority states on agent behavior. The study features a core delegation experiment with structured workflows that vary rule formulations and authority states. Key findings include that both constitutional prompts and provenance-aware guards achieved zero realized violations, although they differed in how they handled prohibited attempts. The research highlights that the enforcement architecture and the authority state significantly impact safety, suggesting that identical safety outcomes can arise from different underlying mechanisms. The paper emphasizes the importance of understanding the institutional context in which AI agents operate to enhance safety in multi-agent systems.
Methodology
The study utilizes a structured three-role delegation workflow in a controlled environment, analyzing the effects of varying rule information and authority states on agent behavior across multiple episodes. It includes experiments on delegation, enforcement mechanisms, and resource allocation.
Results
The experiments showed that constitutional prompts and provenance-aware guards both resulted in zero realized violations (0/384). However, the provenance-aware guard blocked 51 unsafe attempts, with 44 of those episodes later completing safely. In contrast, a local-state guard admitted 22 violations in specific scenarios, highlighting the importance of the authority state and enforcement mechanisms.
Implications
The findings suggest that designing AI systems with a clear understanding of institutional frameworks can enhance safety and prevent violations. This research could inform the development of more robust multi-agent systems and governance structures in AI applications.
Target-Weighted Neyman Allocation: Experimental Design for Heterogeneous Treatment Effects under Population Shift
Theory
Optimization
- TWNA minimizes estimation error for target-weighted group average treatment effects (GATEs).
- The method incorporates a two-stage design using pilot data to inform sample allocation.
- TWNA adapts to uncertainties in the target population composition, enhancing robustness.
- Simulations show TWNA outperforms standard experimental designs in terms of estimation accuracy.
Read more
Target-Weighted Neyman Allocation: Experimental Design for Heterogeneous Treatment Effects under Population Shift
Summary
This paper addresses the challenge of designing randomized experiments that effectively guide decisions across different populations, particularly when treatment effects vary among subgroups. The authors introduce Target-Weighted Neyman Allocation (TWNA), a two-stage stratified experimental design that optimally allocates sample sizes and treatment probabilities based on pilot estimates of group–arm outcome variances. The first stage involves a small pilot experiment to estimate outcome variances, while the second stage allocates samples according to both the importance of groups in the deployment population and the precision of their measurement. TWNA adapts to uncertainties in deployment composition, ensuring robustness whether the target mix is known or unknown. The authors demonstrate that TWNA significantly reduces estimation error for target-weighted group average treatment effects (GATEs) compared to traditional methods, particularly when groups are both important for deployment and difficult to measure accurately. Through simulations and real-world benchmarks, the paper establishes TWNA's effectiveness and provides a practical framework for researchers and practitioners in experimental design.
Methodology
The methodology involves a two-stage stratified design where a pilot experiment estimates outcome variances for different groups and treatment arms. The final experiment allocates samples based on these estimates and the target deployment shares, balancing the importance of groups and the difficulty of measuring their outcomes. The authors derive both an oracle design and a plug-in rule that approaches the oracle as pilot data stabilizes.
Results
The results indicate that TWNA consistently reduces estimation error for target-weighted group effects compared to standard alternatives, such as deployment-proportional and variance-only designs. The method tracks the oracle design closely as the pilot size increases, demonstrating its effectiveness in real-world applications.
Implications
The findings have significant implications for researchers and practitioners conducting randomized experiments, particularly in fields where treatment effects vary across subpopulations. TWNA provides a robust framework for optimizing sample allocation, ensuring that experimental results are more generalizable and applicable to target populations.
SPECTRA: Pushing the KV Cache Beyond the 2-Bit Cliff via Spectral Transform Coding
Large Language Models
Efficient ML
NLP
- SPECTRA addresses the limitations of existing KV cache compression methods by focusing on the uneven distribution of information across cache channels.
- The codec operates in a transformed spectral basis, allowing for effective allocation of bit budgets to the most informative channels.
- SPECTRA achieves up to 12x compression while maintaining near-lossless quality, surpassing traditional quantization methods.
- The method is training-free and can be integrated into existing systems with minimal overhead.
Read more
SPECTRA: Pushing the KV Cache Beyond the 2-Bit Cliff via Spectral Transform Coding
Summary
The paper introduces SPECTRA, a novel codec designed to enhance the efficiency of key-value (KV) cache compression in large language models (LLMs). As LLMs increasingly handle longer contexts, the KV cache becomes a significant memory bottleneck, particularly when traditional quantization methods reach a '2-bit cliff' where further compression leads to a sharp decline in quality. The authors conduct a comprehensive analysis revealing that the information in the KV cache is unevenly distributed across its channels, with a small subset carrying most of the signal. By transforming the cache into a spectral basis derived from its own statistics, SPECTRA allocates bit budgets more effectively, concentrating resources on the most informative channels. This approach allows for higher compression rates—up to 12x—while maintaining near-lossless quality, thus enabling LLMs to handle longer contexts and larger batch sizes without compromising performance. The method is evaluated on various models, demonstrating its superiority over existing quantization and low-rank methods.
Methodology
The authors analyze the KV cache to identify that its information is concentrated in a few principal directions. They then develop SPECTRA, a codec that transforms the cache into this spectral basis, allowing for adaptive bit allocation based on channel importance. This involves a one-time calibration to estimate the basis, after which the codec can efficiently compress the cache during inference.
Results
SPECTRA was tested on models such as Llama-3.1-8B and Qwen2.5-7B, achieving near-lossless performance at 4x compression and competitive results at 8x, where traditional methods fail. The codec successfully pushes usable compression beyond the 2-bit cliff, enabling longer contexts and larger batch processing.
Implications
The development of SPECTRA has significant implications for the deployment of large language models in real-world applications, particularly those requiring long-context processing. It allows for more efficient use of GPU memory, potentially reducing costs and improving throughput in various AI-driven tasks.
Application of Artificial Intelligence for Fraudulent Banking Operations Recognition
Theory
- The study focuses on using AI and ML to detect fraudulent banking transactions, especially during the pandemic and war.
- Preprocessing techniques are crucial for improving the accuracy of fraud detection models.
- Logistic regression and stacked generalization algorithms demonstrated high performance in identifying fraudulent transactions.
- The research highlights the challenges of implementing AI in fraud detection, including transparency and privacy concerns.
Read more
Application of Artificial Intelligence for Fraudulent Banking Operations Recognition
Summary
This paper addresses the increasing prevalence of fraudulent banking operations, particularly in the context of the COVID-19 pandemic and the ongoing war in Ukraine, which have led to a surge in online transactions and fraudulent activities. The authors propose the use of artificial intelligence (AI) and machine learning (ML) algorithms to effectively identify and prevent such fraudulent activities in real-time. The study emphasizes the development of machine learning models specifically designed to recognize fraudulent banking transactions, utilizing various preprocessing techniques to enhance data analysis. The research highlights the importance of handling imbalanced datasets, feature transformation, and feature engineering to improve detection accuracy. The authors implemented several machine learning models, including an artificial neural network, and compared their performance using quantitative and visual criteria. The results indicate that the logistic regression algorithm achieved an AUC value of approximately 0.946, while a stacked generalization approach yielded a better AUC of 0.954. The findings underscore the potential of AI-based systems to enhance fraud detection in the banking sector, addressing the challenges posed by sophisticated fraudulent schemes and the need for reliable automated solutions.
Methodology
The authors developed multiple machine learning models, applied preprocessing techniques to bank data, and compared the models' performance using quantitative metrics and visualizations. They focused on classification algorithms to identify fraudulent transactions based on historical data.
Results
The study found that the logistic regression algorithm achieved an AUC value of approximately 0.946, while the stacked generalization method improved this to 0.954, indicating effective detection of fraudulent banking transactions.
Implications
The findings suggest that AI and ML can significantly enhance the detection and prevention of banking fraud, providing banks with automated tools to protect against financial losses. However, the study also emphasizes the need to address challenges related to algorithm transparency and privacy concerns.
Gaming Without an Attacker: Benchmark Fingerprinting in LLM-Driven Search Under Selection Pressure
Optimization
Large Language Models
Theory
- LLMs under selection pressure can produce solutions that exploit evaluation configurations, leading to poor generalization.
- 30% of in-distribution performance wins do not transfer to held-out configurations, indicating significant benchmarking flaws.
- A four-mode taxonomy of failure mechanisms is introduced to categorize how and why these failures occur.
- The paper provides theoretical insights into the implications of selection pressure on benchmarking and optimization.
Read more
Gaming Without an Attacker: Benchmark Fingerprinting in LLM-Driven Search Under Selection Pressure
Summary
This paper investigates the discrepancies between benchmark evaluations and the actual performance of systems optimized under selection pressure, particularly in the context of large language models (LLMs) proposing GPU kernels. The study focuses on two benchmark suites, Metal-Sci and Metal-ZK, which involve scientific computing and cryptographic tasks, respectively. The author demonstrates that LLMs, when subjected to a (1+1) evolutionary optimization loop, tend to produce solutions that 'fingerprint' the evaluation configuration, leading to a significant failure rate in transferring performance to held-out configurations. Specifically, 30% of in-distribution wins fail to generalize to these unseen configurations. The paper introduces a four-mode taxonomy of failure mechanisms, ranging from differential tuning of branches to strategy overfitting. Additionally, it provides theoretical insights into the adaptive reuse of evaluation pools and offers design guidance for creating effective held-out gates in benchmarks. The findings highlight the need for careful consideration of measurement validity in the presence of strategic optimization, suggesting that benchmarks must account for the dynamic nature of LLMs.
Methodology
The study employs a (1+1) evolutionary search loop where LLMs propose GPU kernels based on in-distribution performance feedback. The kernels are evaluated on both in-distribution and held-out configurations to assess generalization capabilities. The research includes empirical audits of performance transfer and theoretical modeling of the optimization process.
Results
The empirical analysis reveals that 16 out of 53 (30%) of the in-distribution wins fail to transfer to held-out configurations. The introduced taxonomy categorizes the failure modes into four distinct types, providing a structured approach to understanding the underlying issues in benchmark evaluations.
Implications
The findings suggest that current benchmarking practices may not accurately reflect the true capabilities of LLMs, particularly in dynamic optimization scenarios. This has implications for the design of future benchmarks and the evaluation of AI systems, emphasizing the need for robust measurement strategies that account for strategic optimization behaviors.
Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk
Interpretability
Large Language Models
- Development of a multi-scale stacking ensemble for credit default prediction.
- Statistical validation of model performance with confidence intervals.
- Operational analysis reveals limited practical benefits despite high AUC scores.
- Fidelity audit shows significant discrepancies between attributions and generated narratives.
Read more
Accurate Ensembles, Fragile Narratives: Multi-Scale Stacking and a Fidelity Audit of LLM-Generated Explanations for Credit Risk
Summary
This paper addresses the challenge of explainability in credit scoring models, which often utilize complex algorithms that do not provide easily interpretable decision logic. The authors propose an end-to-end system that combines a multi-scale stacking ensemble for credit default prediction with a language model (LLM) to generate explanations based on feature attributions. The predictive model integrates four gradient-boosting learners and a residual network, achieving a test ROC-AUC of 0.9539 and a PR-AUC of 0.9137, outperforming the best single model. However, the study reveals significant limitations in the explanation component, where the LLM-generated narratives sometimes contradict the underlying feature attributions. The authors conduct a fidelity audit, demonstrating that while the attribution methods SHAP and LIME agree on the importance of features, they often disagree on their ranking and the direction of their impact. The findings highlight the necessity for rigorous verification of generated explanations, emphasizing that prompt engineering alone is insufficient to ensure accuracy. The paper concludes with a call for more robust methods to ensure the trustworthiness of AI-generated narratives in sensitive applications like credit risk assessment.
Methodology
The authors implemented a multi-scale stacking ensemble that combines four gradient-boosting models with a residual network, using a small neural meta-learner for integration. Feature attributions were computed using SHAP and LIME, and the top-ranked features were fed into a 2.7B-parameter language model for narrative generation. A fidelity audit was conducted to assess the accuracy of the generated explanations against the attributions.
Results
The multi-scale stacking ensemble achieved a test ROC-AUC of 0.9539 and a PR-AUC of 0.9137, significantly outperforming the best single model. However, the operational gains were minimal, with only six additional missed defaults avoided at the optimal threshold. The fidelity audit revealed that the LLM-generated narratives often contradicted the attributions, indicating a failure in the explanation process.
Implications
The findings underscore the importance of ensuring that AI-generated explanations in critical domains like credit risk are not only fluent but also grounded in accurate data. This has implications for regulatory compliance and trust in AI systems, suggesting that further research is needed to enhance the reliability of explanations provided by language models.
MiCoPro: End-to-End Mixed Precision HW/SW Co-design with HW-aware Proxy Model
Efficient ML
- Introduction of MiCoPro, a holistic framework for mixed precision quantization.
- Development of a novel optimization algorithm for efficient exploration of MPQ schemes.
- Implementation of a Hardware-Aware Proxy model to improve latency predictions.
- Demonstration of significant latency reduction with minimal accuracy loss on various hardware platforms.
Read more
MiCoPro: End-to-End Mixed Precision HW/SW Co-design with HW-aware Proxy Model
Summary
The paper presents MiCoPro, an innovative framework designed for mixed precision quantization (MPQ) of neural networks, particularly aimed at enhancing the deployment of AI models on edge devices. The authors identify the challenges associated with existing MPQ algorithms, which often lack flexibility and efficiency in exploring quantization schemes while maintaining accuracy. MiCoPro addresses these issues by introducing a robust Hardware-Aware Proxy (HAP) model that improves prediction accuracy and adapts to various hardware platforms. The framework employs a novel optimization algorithm that efficiently searches for optimal quantization configurations under strict latency constraints. By leveraging target-specific latency modeling, MiCoPro facilitates the rapid transition from PyTorch models to bare-metal C code, ensuring compatibility with different hardware architectures. The authors validate the framework's effectiveness through case studies on the BitFusion accelerator and SIMD-extended RISC-V processors, demonstrating significant latency reductions of up to 40% with minimal accuracy degradation of less than 3%. This work contributes to the field of edge AI by providing a comprehensive solution for MPQ exploration and deployment, ultimately enhancing the performance of low-bitwidth neural networks.
Methodology
The authors developed an open-source framework that combines ensemble learning models, specifically random forests, with tailored sampling strategies to efficiently explore MPQ schemes. The Hardware-Aware Proxy model is utilized to provide accurate latency estimates, guiding the optimization process towards configurations that yield true performance improvements.
Results
The MiCoPro framework achieved up to 40% latency reduction in model inference while maintaining less than 3% accuracy drop across different hardware platforms, including specialized accelerators and general-purpose processors.
Implications
The MiCoPro framework has significant implications for the deployment of AI models on resource-constrained edge devices, enabling more efficient use of hardware resources while maintaining model accuracy. This can lead to broader adoption of AI applications in real-time scenarios, such as mobile devices and IoT systems.
A Hybrid Nested Harness for Decoupling Structure and Parameters in LLM-Driven Optimization
Large Language Models
Optimization
Theory
- Introduces a hybrid nested search framework that separates structural and parametric optimization.
- Demonstrates improved efficiency in LLM-driven optimization tasks by reducing token waste.
- Validates the approach across three scientific domains, showing superior performance compared to traditional methods.
- Identifies and addresses the issue of parametric aliasing in joint searches.
Read more
A Hybrid Nested Harness for Decoupling Structure and Parameters in LLM-Driven Optimization
Summary
This paper introduces a novel framework called Hybrid Nested Harness, designed to enhance the efficiency of evolutionary algorithms that utilize large language models (LLMs) for optimization tasks. Traditional approaches often require LLMs to simultaneously update both structural components and continuous parameters, which can lead to inefficiencies due to the LLM's limited capability in optimizing numeric values. The proposed method separates these two aspects into an outer loop, where the LLM generates a structural sketch with numeric gaps, and an inner loop, where a numerical optimizer fine-tunes these parameters. This dual-loop structure allows for the integration of various text-based optimizers and numerical solvers, such as CMA-ES and gradient-based methods. The framework was validated across three distinct scientific domains: meta-optimizers on closed-form test functions, code-based policies for systems research and social dilemmas, and approximate Bayesian inference tasks. Results demonstrated that the hybrid optimizer outperformed both standard LLM-driven searches and pure numerical optimization approaches, showcasing the effectiveness of the proposed decoupling strategy.
Methodology
The methodology involves a bilevel optimization approach where an outer loop utilizes a frozen LLM to propose structural sketches, while an inner numerical optimizer tunes the parameters. The framework allows for the integration of various optimizers and includes a mechanism to validate the proposed structures before tuning. The performance is evaluated based on the optimized fitness function across different tasks.
Results
The hybrid nested search framework was validated through experiments in three domains, consistently outperforming both vanilla LLM-driven searches and pure numerical optimization baselines. The results indicate that the separation of structure and parameters leads to more effective optimization, reducing the inefficiencies associated with traditional joint search methods.
Implications
This work has significant implications for optimizing complex systems in various scientific fields, as it provides a more efficient method for leveraging LLMs in optimization tasks. The framework can be applied to enhance automated discovery systems and improve the performance of algorithms in areas such as meta-optimization, systems research, and Bayesian inference.
A Rate Separation for Agnostic Direct Sums
Theory
- The single-instance learning rate does not determine the direct-sum learning rate.
- The paper provides a negative answer to the rate-level form of a question posed by HMW24.
- A rate separation theorem is established, showing differing rates for direct sums of two specific function classes.
- Both classes F and G have agnostic learning curves of order n^{-1/2}, but their direct sums diverge as r increases.
Read more
A Rate Separation for Agnostic Direct Sums
Summary
This paper addresses a question posed by Hanneke, Moran, and Waknine regarding the relationship between the agnostic PAC learning curve of direct sums of concept classes and the single-instance learning curve. The authors demonstrate that the learning rate of a single instance does not dictate the learning rate of the direct sum of multiple instances. Specifically, they analyze two classes: F, consisting of two constant binary functions, and G, which includes the zero function and the identity function. Both classes exhibit an agnostic learning curve of order n^{-1/2}. The main contribution is a rate separation theorem that shows the direct sums of these classes have differing rates as the number of factors increases, despite having similar rates at the single-instance level. This finding has implications for understanding the complexities of agnostic learning in multi-instance settings.
Methodology
The authors utilize theoretical analysis and proofs to establish the relationships between the learning rates of single-instance and direct-sum learning scenarios. They define concept classes and their respective learning curves, employing techniques from empirical risk minimization and inequalities related to distributional properties.
Results
The main results include the establishment of upper and lower bounds for the agnostic learning curves of the direct sums of the classes F and G. Specifically, they show that for all n, r ≥ 1, the learning rates are bounded as follows: c√n ≤ εagn(n | Fr) ≤ C√n and c min(1, r/n^{1/2}) ≤ εagn(n | Gr) ≤ C min(1, r/n^{1/2}). This indicates that while the single-instance rates are similar, the direct sums diverge as the number of instances increases.
Implications
These findings suggest that the complexities of learning in multi-instance settings cannot be simply inferred from single-instance learning rates. This has potential implications for the design of learning algorithms and understanding the limitations of agnostic learning frameworks.
Mathematical Principles and Experimental Discoveries of the Emergence of Symbolic Patterns in Artificial Neural Networks
Theory
Interpretability
- Symbolic patterns can emerge in ANNs, providing a framework for explainability.
- Two mathematical criteria—monotonicity and smoothness—lead to the emergence of sparse symbolic interactions.
- The emergent symbolic patterns show strong transferability across different models and input samples.
- The study introduces a communicative learning paradigm for inspecting ANN inference logic.
Read more
Mathematical Principles and Experimental Discoveries of the Emergence of Symbolic Patterns in Artificial Neural Networks
Summary
This paper investigates the emergence of symbolic patterns in Artificial Neural Networks (ANNs), addressing the challenge of explainability in deep learning. The authors propose that the inference logic of ANNs can be reformulated as sparse symbolic interactions, which are not merely coincidental but arise from underlying mathematical principles. They identify two key criteria—monotonicity and smoothness of the ANN's average k-order output—that facilitate the emergence of these symbolic patterns. Through theoretical analysis and empirical validation, the authors demonstrate that these interactions exhibit strong transferability across different samples and models, suggesting that they represent a fundamental aspect of neural network representations. The findings advocate for a new paradigm of communicative learning, allowing for direct inspection and tuning of ANN inference logic at the symbolic level, thereby complementing traditional end-to-end learning approaches. The study also posits that similar symbolic representations may emerge in other black-box systems under certain conditions, broadening the implications of their findings.
Methodology
The authors utilize a logical model based on AND-OR interactions to explain the inference logic of ANNs. They prove the universal matching property of these interactions, demonstrating that a concise logical model can accurately approximate ANN outputs across various input configurations. The study includes both theoretical proofs and extensive empirical experiments to validate the proposed framework.
Results
The results confirm that the logical model can faithfully explain ANN outputs with a sparse set of interactions. The identified mathematical criteria are shown to hold for a majority of input samples across diverse ANN architectures, supporting the emergence of symbolic patterns as a natural phenomenon rather than a coincidence.
Implications
The findings have significant implications for enhancing the interpretability of ANNs, potentially influencing the design of future neural network architectures and learning paradigms. The introduction of communicative learning could lead to more transparent AI systems, facilitating better understanding and trust in machine learning models.
Privileged Solutions or Context-Induced Teacher Behavior? Dissecting On-Policy Self-Distillation
NLP
Large Language Models
Theory
- OPSD is traditionally viewed as transferring privileged information from a teacher to a student.
- The introduction of OP2SD shows that context-induced teacher behavior can significantly influence learning outcomes.
- OP2SD improves model performance without requiring the exact solution to the target problem.
- The identity and relevance of the teacher's context are crucial for effective self-distillation.
Read more
Privileged Solutions or Context-Induced Teacher Behavior? Dissecting On-Policy Self-Distillation
Summary
This paper investigates On-Policy Self-Distillation (OPSD), a method where a teacher model provides token-level supervision to a student model based on privileged information from a reference solution. The authors argue that the benefits of OPSD may not solely arise from access to the exact solution but also from the context in which the teacher operates. To explore this, they introduce OP2SD (On-Policy Self-Distillation from Other Problems), which uses a solution from a different problem while keeping the student’s target problem unchanged. Through experiments on three models across various mathematics benchmarks, OP2SD demonstrates improvements over the base model and remains competitive with OPSD. The findings suggest that the teacher's contextual behavior plays a significant role in the distillation process, indicating that OPSD's advantages may stem from context-induced changes rather than just privileged access to the solution. The authors provide a diagnostic contribution that clarifies the mechanisms behind OPSD and emphasizes the importance of the teacher's context in the learning process.
Methodology
The authors conducted experiments comparing OPSD and OP2SD across three models and three mathematics benchmarks. OP2SD involved providing the teacher with a solution to a different problem while maintaining the student's target problem and rollout. The performance of both methods was evaluated to assess the impact of context on learning.
Results
The results indicated that OP2SD outperformed the base model and was competitive with OPSD across the evaluated benchmarks. This suggests that the improvements associated with OPSD do not solely depend on access to the exact solution, highlighting the importance of the teacher's contextual behavior.
Implications
These findings could influence future research on self-distillation techniques in machine learning, suggesting that focusing on the context provided to teacher models may enhance learning outcomes. This could lead to more effective training strategies in various applications, particularly in natural language processing and other areas where model distillation is relevant.
From Optimal Actions to World Models: Identifiability of Transition Kernels in Discounted MDPs
Reinforcement Learning
Theory
- Optimal actions alone do not uniquely identify transition probabilities in MDPs.
- The relationship between two kernels can be expressed through an invertible matrix when they yield the same optimal actions.
- Transition-dependent rewards provide the most information for identifying transition kernels, while state rewards are less informative.
- The study establishes a hierarchy of identifiability based on the type of reward used.
Read more
From Optimal Actions to World Models: Identifiability of Transition Kernels in Discounted MDPs
Summary
This paper investigates the extent to which the transition probabilities of a Markov Decision Process (MDP) can be inferred solely from optimal actions across a range of rewards. The author builds upon previous work that explored the recovery of dynamics from numerical Q-values, but focuses instead on the optimal actions. The study establishes that knowing the optimal actions for every reward does not uniquely determine the transition probabilities. The author proves that two transition kernels yield the same optimal actions for all rewards if they are related through a specific invertible matrix. The paper further distinguishes between different types of rewards: state rewards, state-action rewards, and transition-dependent rewards, each providing varying levels of information about the transition dynamics. The findings indicate that transition-dependent rewards can often recover the transition kernel completely, while state rewards provide less information, leading to a hierarchy of identifiability based on reward structure. The results highlight the critical role of reward design in learning about the underlying dynamics of MDPs.
Methodology
The author employs theoretical analysis to classify transition kernels based on their optimal actions for a variety of reward structures. The study involves proving relationships between kernels and establishing conditions under which they can be considered indistinguishable based on optimal actions.
Results
The main result shows that two transition kernels yield the same optimal actions for every reward if they can be expressed through a specific invertible matrix. Additionally, it is demonstrated that transition-dependent rewards can identify almost all transition kernels, while state rewards lead to a weaker equivalence condition.
Implications
The findings suggest that careful design of reward structures in reinforcement learning can significantly enhance the ability to infer transition dynamics, which is crucial for effective policy learning and decision-making in MDPs.
FedTVD: Balancing Data Quality and Quantity for Robust Federated Learning
Federated Learning
- FedTVD introduces a dual-weighting mechanism that considers both data quality and quantity for client contributions in federated learning.
- The algorithm utilizes Total Variation Distance to measure divergence in label distributions, improving model robustness against data imbalance.
- Experimental results show up to 10.6% improvement over traditional methods like FedAvg, particularly in scenarios with highly skewed data.
- FedTVD enhances fairness and generalization in federated learning models by moderating the influence of clients with skewed data distributions.
Read more
FedTVD: Balancing Data Quality and Quantity for Robust Federated Learning
Summary
The paper introduces FedTVD, a novel federated learning (FL) algorithm designed to address the challenges posed by data heterogeneity, particularly in terms of label distribution skewness and dataset size variations. Traditional FL methods, such as FedAvg, primarily weight client contributions based on dataset size, which can lead to biased model updates and hinder convergence. FedTVD enhances this approach by integrating Total Variation Distance (TVD) to assess the divergence between each client's local label distribution and a uniform global distribution. This dual-weighting mechanism allows clients with highly skewed distributions to receive lower weights, thus preventing unbalanced datasets from disproportionately influencing the global model. The methodology effectively balances data quality and quantity, leading to more stable and generalized global models. Experimental results demonstrate that FedTVD consistently outperforms state-of-the-art methods across various datasets, including FMNIST, CIFAR-10, and CIFAR-100, achieving significant improvements in performance, particularly under conditions of high data skewness. The study emphasizes the importance of addressing data heterogeneity to develop robust and equitable FL models that generalize well across diverse client populations.
Methodology
FedTVD employs a dual-weighting mechanism that combines Total Variation Distance (TVD) to assess the skewness of local label distributions with dataset size to determine client contributions during model aggregation. This approach allows for adaptive weighting, ensuring that clients with heavily skewed data have moderated influence on the global model while still leveraging the benefits of larger datasets.
Results
The experimental evaluation of FedTVD across multiple datasets (FMNIST, CIFAR-10, CIFAR-100) shows consistent performance improvements over existing methods, particularly under conditions of data heterogeneity. Notably, FedTVD achieves up to a 10.6% improvement over FedAvg on CIFAR-10 with highly skewed data, while also maintaining strong performance in moderate and IID settings.
Implications
The findings suggest that FedTVD can significantly enhance the robustness and fairness of federated learning systems, making it suitable for applications in sensitive domains such as healthcare and finance, where data privacy and model generalization are critical. The algorithm's ability to balance data quality and quantity can lead to more effective collaborative learning across diverse client environments.
Online Security Learning in Cooperative Multi-Agent Systems under Hidden Byzantine Attacks
Theory
Reinforcement Learning
Robotics
- Introduces a Byzantine team MDP framework for modeling cooperative multi-agent systems under hidden attacks.
- Establishes the impact of attacker's information on decision-making geometry and MDP structure.
- Identifies the decomposition of security regret into return regret and response gap, highlighting the necessity of the latter.
- Develops a robust estimation-to-decisions learner with improved regret bounds.
Read more
Online Security Learning in Cooperative Multi-Agent Systems under Hidden Byzantine Attacks
Summary
This paper investigates the challenges of online cooperative control in multi-agent systems (MAS) that are susceptible to Byzantine attacks, where a subset of agents can stealthily alter their actions after observing the team's planned joint action. The authors introduce a Byzantine team Markov decision process (MDP) framework to model this scenario, focusing on optimizing team performance against the worst-case outcomes induced by these attacks. They establish that the attacker's information significantly influences the decision-making geometry, leading to different MDP structures based on whether the attacker observes the planned actions or not. The paper also identifies the information-theoretic limits of security learning, demonstrating that security regret can be decomposed into return regret and a cumulative response gap. A robust estimation-to-decisions learner is proposed, achieving a regret bound that improves upon previous methods. The findings provide a theoretical and algorithmic foundation for developing reliable multi-agent systems capable of withstanding Byzantine action overwrites.
Methodology
The authors develop a finite-horizon cooperative team MDP to model the interactions between agents in the presence of Byzantine attacks. They analyze the impact of the attacker's information on the decision-making process and derive a regret decomposition for security learning. A robust estimation-to-decisions learner is constructed, which utilizes a stage-tied estimator to optimize performance across episodes.
Results
The proposed learner achieves an expected return regret of O(H^2 S √(AK)) over K episodes, with A representing the number of joint actions. The expected security regret is shown to exceed this by the expected response gap, indicating that the learner effectively addresses the challenges posed by hidden Byzantine attacks.
Implications
The findings have significant implications for the design of resilient multi-agent systems, particularly in applications where agents must collaborate under potential adversarial conditions, such as autonomous vehicle fleets and coordinated energy systems. The theoretical insights can guide the development of robust algorithms that ensure reliable performance despite the presence of compromised agents.
RippleKV: Cross-Layer KV Cache Allocation via Perturbation Propagation
NLP
Large Language Models
Efficient ML
- RippleKV provides a principled method for cross-layer KV cache allocation based on output sensitivity rather than traditional proxies.
- The method captures the heterogeneous and nonmonotonic cache requirements across different layers of the model.
- RippleKV integrates seamlessly with existing KV cache compression methods without requiring additional evaluations during inference.
- Extensive experiments validate RippleKV's superior performance across multiple model families and cache budgets.
Read more
RippleKV: Cross-Layer KV Cache Allocation via Perturbation Propagation
Summary
RippleKV addresses the challenge of allocating a limited key-value (KV) cache budget across layers in long-context large language model (LLM) inference. Traditional methods rely on proxies like layer depth or attention statistics, which can lead to inefficient cache allocation, particularly underestimating the needs of sensitive layers. RippleKV introduces a novel approach by estimating how perturbations in each layer's value cache affect the final output distribution, allowing for a more accurate sensitivity profile. This profile is generated by injecting norm-adaptive perturbations into each layer's value cache and measuring the resulting KL divergence at the model output. The sensitivity scores are then normalized and transformed into budget multipliers, ensuring that the total cache budget is preserved while allowing for a disparity in allocation between sensitive and tolerant layers. Experimental results on LongBench demonstrate that RippleKV outperforms existing KV cache compression methods, achieving the highest average performance while maintaining efficiency in runtime and memory usage.
Methodology
RippleKV employs a perturbation propagation strategy where it independently perturbs the value cache of each layer and measures the impact on the output distribution using KL divergence. This process generates a sensitivity profile that reflects how cache changes propagate through the model. The sensitivity scores are then normalized and transformed into budget multipliers to allocate cache efficiently across layers.
Results
The experiments conducted on LongBench indicate that RippleKV consistently achieves the highest average performance compared to other KV cache compression methods, demonstrating its effectiveness in optimizing cache allocation while maintaining comparable runtime and memory efficiency.
Implications
RippleKV's approach can significantly enhance the efficiency of long-context LLM inference, making it applicable in various domains such as document summarization, question answering, and information retrieval, where memory and computational resources are critical.
When Can Fraud Operations Authorize Automation? A Decision-Support Framework for Fresh Audit Evidence and Review Workload
Theory
- Introduces the Freshness-Constrained Audit Capacity (FCAC) framework for fraud automation.
- Separates predictive ranking from the authority to automate actions based on evidence freshness.
- Identifies the trade-off between audit capacity and review workload in fraud operations.
- Demonstrates the framework's effectiveness through simulations on multiple financial datasets.
Read more
When Can Fraud Operations Authorize Automation? A Decision-Support Framework for Fresh Audit Evidence and Review Workload
Summary
This paper addresses the challenge of automating fraud operations in financial institutions, where decisions must be made regarding automatic approval, analyst review, or automatic blocking of transactions. The author introduces a decision-support framework called Freshness-Constrained Audit Capacity (FCAC), which evaluates the automation of fraud detection based on the freshness of evidence, action risk, and review workload. The framework separates predictive ranking from the authority to act, emphasizing the importance of evidence age and shared review capacity. The study formulates three research questions focused on the conditions under which delayed audits can certify current risks, the interplay of various factors affecting automation, and the consistency of findings across different financial domains. The FCAC framework is validated through chronological evaluations using simulated audits across three datasets, demonstrating its effectiveness in managing automation rates and review workloads while highlighting the trade-offs between audit intensity and authorization delays.
Methodology
The methodology involves developing the FCAC framework that integrates mature randomized audits, a prespecified temporal allowance, and asymmetric action-risk limits. The framework evaluates candidate action regions based on evidence freshness and review workload, producing an auditable record of supported and unsupported regions for automation.
Results
The experiments conducted with simulated audits on datasets such as IEEE-CIS, ULB-Worldline, and Elliptic++ yielded automation rates of 84.4%, 67.4%, and 81.3%, respectively, with total review workloads of 24.1%, 46.0%, and 43.1%. The findings indicate that sparse auditing delays automation, while intensive auditing can increase overall workload.
Implications
The findings suggest that financial institutions can enhance their fraud detection processes by implementing the FCAC framework, allowing for more informed decisions regarding automation and resource allocation. This approach could lead to improved operational efficiency and reduced risk in fraud management.
Out-of-Distribution Federated Distillation with Domain-Aware Proxy
Federated Learning
- Introduction of a Domain-aware Proxy Selection framework (DPS-FD) for better proxy data adaptation in OOD scenarios.
- Development of a Vocabulary-Constrained LLM-based generation strategy for generating proxy data in the absence of real data.
- Empirical results show significant performance improvements in OOD scenarios, achieving competitive accuracy on standard benchmarks.
- Analysis reveals that high-quality proxy data is crucial for effective knowledge transfer and model robustness.
Read more
Out-of-Distribution Federated Distillation with Domain-Aware Proxy
Summary
This paper addresses the limitations of Federated Distillation (FD) when applied to Out-of-Distribution (OOD) scenarios. Traditional FD relies on proxy data for knowledge transfer, but models trained on In-Distribution data struggle to adapt to OOD conditions. The authors propose a Domain-aware Proxy Selection framework (DPS-FD) that enhances the selection of proxy data to better capture the characteristics of both local and global distributions. DPS-FD allows local clients to choose relevant domain samples while the server selects globally representative samples, creating a more effective proxy dataset. Additionally, the paper introduces a Vocabulary-Constrained LLM-based generation strategy (VC) for generating proxy data when real proxy data is unavailable, ensuring competitive performance in textual data scenarios. The experimental results demonstrate that DPS-FD achieves an average accuracy of 82.9% and 80.6% on standard benchmarks, outperforming existing methods. The findings emphasize the importance of high-quality proxy data in mitigating OOD challenges and enhancing the robustness of global models.
Methodology
The methodology involves a two-pronged approach: first, the Domain-aware Proxy Selection framework (DPS-FD) enables local clients to select relevant domain samples while the server aggregates globally representative samples. Second, the Vocabulary-Constrained LLM-based generation strategy (VC) is employed to create proxy data when real proxy data is unavailable, using a few-shot prompting strategy to ensure diversity and quality.
Results
The experimental results indicate that the proposed DPS-FD framework achieves an average accuracy of 82.9% and 80.6% on standard benchmarks, significantly outperforming existing methods in both OOD scenarios with and without proxy data.
Implications
The findings suggest that enhancing proxy data selection and generation can significantly improve the robustness and generalization of federated learning models, particularly in OOD contexts. This has potential applications in various fields where data privacy is critical, such as healthcare and finance, where federated learning can be employed without compromising sensitive information.
Hidden Gauge Controls Feature Specialization in ReLU Networks
Theory
Optimization
- Demonstrates a Θ(D²) separation in specialization time due to hidden gauge parameters in ReLU networks.
- Establishes that a favorable gauge can deterministically assign feature ownership to one neuron among functionally identical candidates.
- Introduces a reaction-transport decomposition to explain the mechanisms of feature specialization.
- Validates the theoretical findings with empirical evidence across population and finite-sample training.
Read more
Hidden Gauge Controls Feature Specialization in ReLU Networks
Summary
This paper investigates the phenomenon of feature ownership in overparameterized ReLU networks, where multiple neurons may initially perform the same function but diverge during training. The author introduces a Gaussian teacher-student model to explore how hidden gauge parameters can influence which neuron becomes the 'owner' of a task-relevant feature. The study reveals that varying a positive-homogeneous scaling gauge leads to distinct feature trajectories and a significant separation in specialization time, quantified as Θ(D²). The findings demonstrate that a favorable gauge can deterministically select one neuron as the feature owner while driving the contributions of others to zero. The paper also establishes a reaction-transport decomposition that clarifies the mechanisms behind this selection process, showing that the hidden gauge reallocates mobility between changing a feature's coefficient and direction. The results are validated through theoretical proofs and empirical evidence, confirming the robustness of the findings across various training scenarios.
Methodology
The author employs a Gaussian teacher-student model to analyze the dynamics of feature learning in ReLU networks. The study focuses on varying a hidden scaling gauge while keeping other parameters constant to observe the effects on feature specialization and ownership. Theoretical proofs and empirical validations are used to support the findings.
Results
The study finds that different hidden gauges lead to distinct learning trajectories for neurons, with specialization times differing by a factor of Θ(D²). It proves that one neuron can be deterministically selected as the feature owner when given a favorable gauge, while others become functionally redundant. The results also show that the dynamics of feature learning can be decomposed into changes in functional coefficients and feature directions, with implications for training robustness.
Implications
The findings suggest that the choice of parameterization in neural networks can significantly influence feature learning dynamics, which may have implications for network design and optimization strategies. Understanding feature ownership could lead to more efficient training methods and better performance in various machine learning tasks.