AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
69
Papers today
8h
Update frequency
7
Days of history
Margin-Drop Coordinates for Cross-Budget Robustness Evaluation
Computer Vision
Theory
Efficient ML
- Introduces margin-drop coordinates to evaluate robustness of vision encoders.
- Demonstrates that shallow evaluations can provide critical insights into encoder fragility.
- Shows that median shallow drift coordinates correlate strongly with performance under stronger attacks.
- Distinguishes between different types of margin repairs and their implications for robustness.
Read more
Margin-Drop Coordinates for Cross-Budget Robustness Evaluation
Summary
This paper addresses the limitations of fixed-budget robustness evaluations in selecting appropriate frozen vision encoders, highlighting the fragility of encoders that may appear robust under shallow attacks but fail under stronger evaluations. The authors propose a novel framework that introduces margin-drop coordinates, which consist of clean margin slack, one-step shortfall, and drift. These coordinates are derived from the evaluation of clean-correct samples and provide a more informative representation of an encoder's robustness than traditional survival rates. The study evaluates 42 pretrained frozen vision encoders and demonstrates that the median shallow drift coordinate significantly correlates with performance collapse under stronger attacks, outperforming survival rate rankings. The findings suggest that the proposed margin-drop coordinates can effectively distinguish between encoders with similar fixed-budget outcomes but different robustness characteristics, thereby enhancing the evaluation process for adversarial robustness.
Methodology
The authors developed a framework that computes margin-drop coordinates for each clean-correct sample, capturing clean margin slack, one-step shortfall, and drift. These coordinates are normalized and used to reconstruct the post-attack margin, allowing for a more detailed analysis of encoder performance across varying attack budgets. The methodology involved evaluating 42 pretrained frozen vision encoders under different attack scenarios, including PGD-10 and PGD-200, and comparing the results with traditional survival rate metrics.
Results
The study found that the median shallow drift coordinate ranked the collapse from PGD-10 to PGD-200 with a Spearman correlation of +0.811, while the survival rate showed essentially no correlation (−0.006). The results were consistent across a held-out encoder pool and various evaluation conditions. Additionally, the full coordinate decomposition was able to differentiate encoders with similar fixed-budget outcomes but diverging performance under deeper budgets, revealing distinct repair paths.
Implications
The findings suggest that the proposed margin-drop coordinates can improve the robustness evaluation process for vision encoders, potentially leading to better model selection in applications that require adversarial robustness. This approach could be beneficial in fields such as computer vision, where pretrained models are commonly used in larger systems, enhancing the reliability of these systems against adversarial attacks.
The Dynamics of Quasiregular Neural Learning
Theory
- Neural networks exhibit U-shaped learning dynamics similar to language acquisition.
- Initial acquisition of exceptions can lead to overregularization towards dominant regularities.
- The extent of overregularization is influenced by the prevalence of exceptions in the training data.
- The study isolates the interaction between regularities and exceptions during gradient-based training.
Read more
The Dynamics of Quasiregular Neural Learning
Summary
This paper investigates the dynamics of quasiregular learning in neural networks, particularly in the context of language acquisition where learning exhibits U-shaped trajectories. The study focuses on how neural networks learn a dominant regularity while accommodating systematic exceptions. By constructing controlled regression problems with known regular and exceptional solutions, the author demonstrates that neural networks can initially acquire exceptions, then regress towards the dominant regularity, and eventually recover the correct exceptional behavior. The research reveals that overregularization is more pronounced when exceptions are rare, highlighting a competitive interaction between regularities and exceptions during the learning process. The findings contribute to understanding the complexities of learning dynamics in neural networks, particularly in scenarios where exceptions to rules exist.
Methodology
The author constructs one-dimensional regression problems that combine a dominant regularity with systematic exceptions. A fully connected neural network is trained using gradient descent, with metrics defined to measure the performance on regular and exceptional inputs. The study varies the prevalence of exceptions to analyze its effect on learning dynamics.
Results
The results indicate a non-monotonic learning trajectory for exceptions, characterized by initial improvement, regression towards the regular solution, and eventual recovery. The strength of overregularization is shown to depend on the frequency of exceptions, with a clear pattern of error trajectories observed during training.
Implications
The findings have implications for understanding learning processes in neural networks, particularly in applications involving language and other domains where exceptions to rules are present. This research could inform the design of more robust learning algorithms that better accommodate exceptions.
Marginally Correct Tool Caches Can Reverse Group-Normalized Policy Updates
Theory
Optimization
Reinforcement Learning
- Tool-result caching can couple rollout randomness, affecting policy updates.
- Sharing one stochastic result per group can reverse expected group-normalized updates.
- A derived finite-group expression highlights the difference between marginal reward agreement and update agreement.
- The paper provides a Bernoulli example demonstrating wrong-direction updates.
Read more
Marginally Correct Tool Caches Can Reverse Group-Normalized Policy Updates
Summary
This paper investigates the impact of tool-result caching on agent training, particularly focusing on how it affects group-normalized policy updates. The author presents a two-action model where independent and shared executions maintain the conditional reward distribution for each rollout. However, it is shown that sharing a stochastic result within a group can reverse the expected group-normalized policy update. The paper derives a finite-group expression demonstrating that the shared update is influenced by the probabilities of winning and losing, rather than the expected reward difference. A Bernoulli example illustrates a scenario where the update direction can be incorrect, leading to a persistent variance floor as the group size increases. The author proposes centering without group standard-deviation scaling as a method to preserve the expected-return direction. The findings are supported by extensive numerical checks and an implementation audit that reproduces the results using a specific tool cache setup. The paper concludes that merely having marginal output validity does not guarantee that a stochastic cache is equivalent for training purposes.
Methodology
The study employs a two-action model to analyze the effects of independent versus shared execution of stochastic tools on policy updates. It derives mathematical expressions for expected updates and conducts exhaustive numerical checks across various configurations. An implementation audit is performed to validate the theoretical findings.
Results
The results indicate that sharing stochastic tool outputs can lead to incorrect policy updates, as evidenced by a derived expression that separates marginal reward agreement from update agreement. The Bernoulli example shows a threshold for preference that can lead to wrong-direction updates. The findings are validated through extensive numerical evaluations and an implementation audit.
Implications
The results suggest that optimizing the use of stochastic tools in training does not guarantee equivalent training outcomes. This has implications for the design of training algorithms that rely on tool caching, indicating the need for careful consideration of how tool outputs are shared among agents.
Minimal Recurrent Behavioral Memory for Imitation under Partial Observability
Robotics
Reinforcement Learning
Theory
- Introduces a framework for minimal recurrent memory needed for imitation under partial observability.
- Establishes a compatibility relation that characterizes the necessary distinctions in behavior.
- Demonstrates that learned memory requirements approach theoretical limits despite increasing hidden modes.
- Highlights the challenges in learning representations due to event-agnostic supervision and predictive surplus.
Read more
Minimal Recurrent Behavioral Memory for Imitation under Partial Observability
Summary
This paper addresses the challenge of imitating an expert's behavior under conditions of partial observability, focusing on the minimal recurrent memory required for such imitation. The authors introduce a framework that characterizes this minimal memory through a compatibility relation, which allows for the preservation of necessary distinctions in behavior that future observations cannot restore. They establish that under transitive compatibility, the conditional entropy of the expert's behavioral quotient provides an exact minimum for the required memory. The study also explores the implications of a sole-carrier measurement protocol that differentiates between behavioral sufficiency and excess code rates. Through various manipulation tasks, the authors demonstrate that the learned memory requirements remain close to theoretical limits, even as the complexity of hidden modes increases. The paper highlights the difficulties in learning this representation, noting that event-agnostic supervision can lead to improved performance but may also introduce predictive surplus. The findings suggest that while the proposed representation can effectively capture necessary behavioral distinctions, achieving optimal learning remains a challenge due to the complexities of temporal dependencies and memory requirements.
Methodology
The authors develop a theoretical framework based on conditional entropy and compatibility relations to characterize minimal recurrent behavioral memory. They conduct experiments across various manipulation tasks to evaluate the performance of their proposed memory representation and analyze the effects of different supervision strategies on learning outcomes.
Results
The experiments reveal that learned code rates remain close to zero and two-bit requirements as the number of hidden modes increases. The architecture demonstrates improved performance in learning representations, achieving a significant increase in success rates in longer-horizon tasks. The results also indicate that while supervision aids in commitment, it can lead to predictive surplus, which can be mitigated through annealing techniques.
Implications
The findings have potential applications in robotics and other fields where imitation learning is crucial under partial observability. The proposed framework can inform the design of more efficient learning algorithms that require minimal memory while maintaining performance in complex environments.
Learning Risk Scores Robust to Unobserved Confounders
Optimization
Theory
- Introduces a method for learning risk scores that accounts for unobserved confounding.
- Utilizes robust optimization techniques to handle uncertainty in propensity scores.
- Demonstrates improved calibration of risk scores compared to traditional methods.
- Applies the method to semi-synthetic data derived from UCI Machine Learning Repository.
Read more
Learning Risk Scores Robust to Unobserved Confounders
Summary
This paper addresses the challenge of learning risk scores from historical observational data that may be affected by unobserved confounding factors. In critical resource allocation scenarios, such as public health and homelessness prevention, risk scores are used to prioritize individuals for limited resources based on their characteristics. Traditional methods like inverse propensity weighting (IPW) rely on the assumption of unconfoundedness, which often fails in practice due to unrecorded information influencing historical decisions. The authors propose a novel method that treats propensity scores as uncertain and formulates a robust optimization problem to maximize the worst-case expected log-likelihood. This approach incorporates sensitivity analysis and Wasserstein distributionally robust optimization to create a robust risk score learning framework. The proposed method is shown to outperform traditional benchmarks and state-of-the-art methods in terms of calibration, demonstrating its effectiveness in mitigating the effects of unobserved confounding.
Methodology
The authors develop a robust optimization framework that treats propensity scores as uncertain. They construct an uncertainty set based on observable data and domain-informed estimates of confounding, combining sensitivity analysis with Wasserstein distributionally robust optimization. The resulting problem is reformulated as an exponential cone program, allowing for efficient computation using existing solvers.
Results
The proposed method improves calibration by up to 29.2% over traditional benchmarks and up to 11.1% over the state of the art, without compromising other performance metrics. This demonstrates the effectiveness of the approach in learning risk scores that are robust to unobserved confounding.
Implications
The findings suggest that the proposed method can be applied in various fields where resource allocation decisions are made based on risk scores, such as healthcare, social services, and criminal justice. By improving the accuracy of risk assessments, the method can lead to better prioritization of resources and interventions.
LAYERSCOPE: A Layerwise Characterization of Video and Multimodal Learned Representations
Multimodal
- LAYERSCOPE provides a label-free framework for evaluating learned representations in video and multimodal settings.
- Intermediate-layer representations often outperform final-layer outputs in various tasks.
- Distinct geometric signatures are observed across different model families, indicating the complexity of representation structures.
- No single geometric metric consistently predicts downstream performance, suggesting a need for a multifaceted evaluation approach.
Read more
LAYERSCOPE: A Layerwise Characterization of Video and Multimodal Learned Representations
Summary
This paper introduces LAYERSCOPE, a novel framework designed to characterize learned representations in video and multimodal contexts without the need for labeled data. Traditional evaluation methods often rely on downstream performance metrics that require extensive labeled datasets and are computationally intensive. LAYERSCOPE addresses these challenges by employing a range of geometric metrics—local, global, distributional, and correspondence-based—to analyze the structure of representations at various layers of a model. The authors evaluate seven diverse models across tasks such as video classification, clustering, and text-to-video retrieval using datasets MVEB and MVEB+. The findings reveal that intermediate-layer representations can outperform final-layer outputs, and that different geometric metrics provide varying insights into model performance. Notably, while no single metric consistently predicts downstream success, distinct geometric signatures emerge across model families. The study highlights the utility of LAYERSCOPE in enabling a systematic evaluation of representations, emphasizing the importance of intermediate layers in understanding model behavior and performance.
Methodology
The authors developed LAYERSCOPE, which utilizes a variety of geometric metrics to analyze the learned representations of models at different layers. This approach allows for a layerwise evaluation without relying on task-specific labels. The framework was tested on seven architecturally diverse models across multiple tasks, including classification, clustering, and text-to-video retrieval.
Results
The evaluation revealed that intermediate-layer representations can be more effective than final-layer outputs. The study found that while no single geometric metric consistently predicts performance across tasks, certain metrics like LID show task-dependent relationships with performance, and RankMe is particularly effective for classification and clustering tasks. Additionally, pairing-aware metrics were found to provide better explanations for retrieval tasks than distributional distances alone.
Implications
LAYERSCOPE offers a systematic approach to understanding and comparing learned representations across different models and tasks, which can lead to improved model design and evaluation strategies in video and multimodal applications. This framework can facilitate deeper insights into the internal workings of models, potentially guiding future research in representation learning.
GeoRVQ: Decoder-aware geometry for residual-token prediction in physiological signals
Time Series
Multimodal
Audio & Speech
- GeoRVQ improves the prediction of physiological signals by incorporating decoder-aware geometry.
- The method uses a coarse-to-fine approach to account for the hierarchical structure of RVQ.
- Significant improvements in token accuracy and waveform preservation were observed in experiments.
- Decoder-induced costs are utilized to create geometry-aware soft targets for better predictions.
Read more
GeoRVQ: Decoder-aware geometry for residual-token prediction in physiological signals
Summary
The paper introduces GeoRVQ, a novel approach to residual vector quantization (RVQ) that enhances the prediction of physiological signals by incorporating decoder-aware geometry. Traditional masked modeling methods treat all token errors equally, which can lead to significant waveform distortions. GeoRVQ addresses this by using a coarse-to-fine masked token model that reflects the local response of a frozen waveform decoder. It defines geometry-aware soft targets based on decoder-induced costs, allowing for more accurate predictions that consider the impact of token substitutions on the waveform. The methodology involves a hierarchical structure where predictions are made from coarse to fine levels, leveraging the residual dependencies inherent in RVQ. The authors evaluate GeoRVQ on datasets including MIMIC-IV Waveform, VitalDB, and CODE-15%, demonstrating improvements in token accuracy, decoded distance, and R-peak F1 scores. The findings suggest that decoder-aware objectives can significantly enhance waveform and event preservation without necessitating a large increase in exact token accuracy.
Methodology
GeoRVQ employs a coarse-to-fine masked token model that utilizes a frozen pretrained codec to provide target tokens and measure the waveform changes caused by codeword substitutions. It constructs soft targets based on decoder-induced costs and follows a quantizer-causal prediction strategy that respects the residual dependencies from coarse to fine levels. The model is evaluated against various loss functions, including one-hot cross-entropy and label smoothing, to assess its effectiveness in preserving waveform characteristics.
Results
GeoRVQ achieved an increase in exact token accuracy from 0.133 to 0.143, a reduction in decoded distance from 0.606 to 0.393, and an increase in R-peak F1 score from 0.784 to 0.837 under matched model and training conditions. The decoder-induced cost showed a Spearman correlation of 0.85 with realized decoded cost, outperforming the Euclidean codeword distance correlation of 0.54.
Implications
The findings suggest that incorporating decoder-aware objectives in token prediction can lead to better preservation of physiological signals, which is crucial for applications in medical diagnostics and monitoring. This approach could enhance the performance of neural codecs in various biomedical signal processing tasks.
Linear RNN Scaling Laws: When Longer Sequences Beat More Sequences
NLP
Large Language Models
Theory
- Introduces a recurrent autoregressive pretraining scaling framework that separates the roles of model size, data size, and optimization compute.
- Identifies two distinct scaling regimes: innovation-dominated and initialization-dominated, each with different implications for sequence length and model performance.
- Demonstrates that longer sequences can outperform simply increasing the number of independent trajectories under certain conditions.
- Provides a theoretical basis for understanding the scaling laws observed in language models, contributing to better pretraining strategies.
Read more
Linear RNN Scaling Laws: When Longer Sequences Beat More Sequences
Summary
This paper investigates the theoretical foundations of empirical scaling laws in autoregressive language models, particularly focusing on the trade-offs between model size, data size, and optimization compute in sequential pretraining settings. The authors introduce a teacher-student model where a stable latent linear RNN generates trajectories, and a sketched linear recurrent student is trained using safeguarded full-batch WSD gradient descent for next-token prediction. The study reveals that the sketch dimension acts as a proxy for model size, while independent trajectories of varying lengths provide the training tokens. The authors derive explicit approximation, optimization, and statistical scaling laws, highlighting the significance of the innovation and initialization covariances. They identify two regimes: innovation-dominated scaling, where longer sequences are interchangeable with more sequences, and initialization-dominated scaling, where longer sequences provide greater benefits beyond a certain threshold. This work contributes to a deeper understanding of the dynamics of autoregressive pretraining and the role of sequence length in model performance.
Methodology
The authors utilize a linear teacher-student model where a latent linear RNN generates trajectories. A Gaussian sketch restricts the observable representation, and a linear recurrent student is trained using safeguarded empirical WSD gradient descent. The analysis decomposes the expected population risk into representation approximation, finite-optimization bias, and statistical variance, allowing for a detailed exploration of the roles of model size, data size, and optimization time.
Results
The study establishes distinct scaling laws based on the relationship between initialization and innovation covariances. In the innovation-dominated regime, the authors recover known approximation and bias rates, indicating that N and P are interchangeable. In the initialization-dominated regime, they identify P-dependent crossovers, demonstrating that longer sequences reduce approximation and bias, thus emphasizing the value of sequence length in training.
Implications
The findings suggest that in designing language models, prioritizing longer sequences may yield better performance than merely increasing the number of training sequences. This insight can inform future research and practical applications in autoregressive model training and optimization strategies.
Signal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection
Time Series
- Introduces a neuro-symbolic framework for explainable anomaly detection in physiological time-series.
- Utilizes symbolic representation of signals to enhance interpretability and transparency in anomaly detection.
- Employs rare itemset mining and Allen interval algebra for scoring and temporal reasoning of anomalies.
- Demonstrates the effectiveness of the approach through evaluations on multiple public datasets.
Read more
Signal2Symbol: Neuro-Symbolic Temporal Reasoning for Explainable Physiological Time-Series Anomaly Detection
Summary
The paper introduces Signal2Symbol, a neuro-symbolic framework designed for explainable anomaly detection in physiological time-series data, particularly focusing on ECG and EEG signals. Traditional deep learning models, while effective in detecting anomalies, often lack transparency in their decision-making processes. Signal2Symbol addresses this gap by converting continuous physiological signals into symbolic sequences using either a learned Vector Quantized Variational Autoencoder (VQ-VAE) or a Symbolic Aggregate approXimation (SAX) method. The framework constructs bigram-enriched token-window transactions and scores anomalies based on rare itemset evidence derived from minimal rare itemset mining. Detected anomalies are then merged into intervals and analyzed using Allen interval algebra, allowing for the identification of complex temporal relationships such as escalation chains and artifact overlaps. Additionally, the authors introduce a rare temporal concept lattice based on Formal Concept Analysis (FCA), which organizes anomalous intervals into interpretable families based on shared symbolic evidence and temporal relations. The framework was evaluated on three public datasets, demonstrating its robustness against noise and perturbations, and highlighting the advantages of neuro-symbolic tokenization for temporal anomaly analysis.
Methodology
The methodology involves converting physiological signals into symbolic sequences using VQ-VAE or SAX, followed by the creation of bigram-enriched transactions. Anomalies are scored based on rare itemset mining, and temporal relationships are analyzed using Allen interval algebra. The final step involves organizing detected anomalies into a Galois lattice using Formal Concept Analysis to provide interpretable explanations.
Results
The results indicate that Signal2Symbol effectively detects anomalies in ECG and EEG datasets, providing robust performance even under noise and perturbations. The neuro-symbolic approach allows for compact and interpretable summaries of local detections, enhancing the understanding of temporal relationships among anomalies.
Implications
The proposed framework has significant implications for e-health and clinical monitoring, as it not only improves the detection of physiological anomalies but also enhances the interpretability of the results, which is crucial for clinical decision-making.
What Changed? Drift Detection with Real, Virtual, and Incomparable Diagnosis
Theory
Time Series
Efficient ML
- Introduces a two-axis monitoring system for drift detection that classifies drift types.
- Critiques existing drift detectors for their inability to accurately diagnose drift types.
- Achieves zero false alarms and misses in extensive benchmarking across different drift scenarios.
- Highlights the importance of distinguishing between real and virtual drift for appropriate system responses.
Read more
What Changed? Drift Detection with Real, Virtual, and Incomparable Diagnosis
Summary
This paper addresses the challenge of concept drift detection in adaptive systems, distinguishing between real drift (changes in the conditional distribution P(Y | X)) and virtual drift (changes in the marginal distribution P(X)). The author critiques existing drift detectors, noting that error-stream detectors tend to misattribute alarms to real drift, while input-space detectors fail to detect real concept changes. To overcome these limitations, the paper proposes a novel two-axis monitoring system based on the conditional Jensen–Shannon discrepancy, which allows for the classification of drift types. The proposed method utilizes a pooled discriminator pair to compute covariate and functional statistics, enabling accurate drift typing. Experimental results demonstrate that this method achieves zero false alarms and misses across various benchmark streams, effectively identifying both real and virtual drifts. The paper also discusses the implications of drift type on system responses, emphasizing the importance of accurate diagnosis for adaptive systems.
Methodology
The proposed method maintains a reference window and a sliding recent window to estimate covariate and functional statistics using a pooled discriminator pair. It outputs three states based on the computed statistics: alarm for real drift, alarm for virtual drift, and report as incomparable when the covariate statistic exceeds a comparability ceiling. The method also incorporates a betting e-process for continuous monitoring.
Results
The two-axis monitoring method demonstrated zero false alarms and zero misses across 40 monitored runs on various benchmark streams, including SEA and STAGGER. It achieved 100% correct typing of drift types, significantly outperforming traditional detectors like ADWIN, DDM, and D3 in both accuracy and reliability.
Implications
The findings suggest that accurate drift detection and diagnosis can enhance the performance of adaptive systems by informing appropriate responses to changes in data distributions. This has potential applications in various domains, including finance, healthcare, and autonomous systems, where timely and accurate adaptation to changing environments is critical.
Brain-Inspired Hierarchical Modularity for General Continual Learning
Robotics
Multimodal
Computer Vision
- Introduces FlyGCL, a brain-inspired framework for general continual learning.
- Addresses the limitations of traditional continual learning methods by focusing on online and uncertain data streams.
- Implements a hierarchical modular structure that combines expert specialization and ensemble integration.
- Demonstrates significant performance improvements across various tasks, especially in embodied vision-language-action learning.
Read more
Brain-Inspired Hierarchical Modularity for General Continual Learning
Summary
This paper presents FlyGCL, a novel framework for general continual learning (GCL) inspired by the hierarchical modularity observed in the Drosophila learning and memory system. Traditional continual learning approaches often operate under simplified conditions with clear task boundaries, which limits their applicability to real-world scenarios characterized by online, uncertain, and evolving data streams. FlyGCL addresses these challenges by employing a brain-inspired modular architecture that facilitates both expert specialization and ensemble integration. The framework utilizes lightweight adaptations of pretrained foundation models, incorporating mechanisms for random expansion to enhance expert routing and diversified modular integration across different spatial and temporal scales. The authors evaluate FlyGCL across various tasks, including visual recognition, vision-language understanding, and embodied vision-language-action learning, demonstrating significant improvements in performance. Notably, FlyGCL achieves success rates that exceed the best replay-free continual learning baselines by substantial margins, particularly in embodied manipulation tasks. These findings suggest that hierarchical modularity can effectively organize learning from dynamic experiences, providing a biologically grounded approach to continual learning.
Methodology
The methodology involves a hierarchical modular design that utilizes pretrained foundation models. It incorporates brain-inspired mechanisms such as random expansion for expert routing and adaptive modules for diversified integration across spatial and temporal dimensions. The framework is evaluated through computational analyses and extensive experiments across multiple continual learning scenarios.
Results
FlyGCL consistently outperforms existing replay-free continual learning methods, achieving final average success rates of 83.1%, 86.1%, 94.5%, and 79.1% in various tasks. The gains over the strongest baselines range from 49.5 to 58.8 percentage points, particularly highlighting its effectiveness in embodied vision-language-action learning.
Implications
The findings suggest that hierarchical modularity can serve as a foundational principle for developing intelligent systems capable of adapting to dynamic environments. This approach may enhance the robustness and efficiency of AI systems in real-world applications, particularly in areas requiring continual learning from diverse and evolving data streams.
I-SplineFlow: Learning Monotone Spline Stochastic Interpolant Schedulers for Few-Step Generation
Generative Models
Efficient ML
Optimization
- I-SplineFlow introduces integrated monotone splines for parameterizing SI schedulers, enhancing flexibility and control.
- The method guarantees monotonicity and differentiability by construction, addressing limitations of previous Bézier-based approaches.
- I-SplineFlow achieves superior few-step FID scores across various models, particularly at low NFEs.
- The approach is computationally efficient, allowing for rapid training compared to traditional methods.
Read more
I-SplineFlow: Learning Monotone Spline Stochastic Interpolant Schedulers for Few-Step Generation
Summary
The paper presents I-SplineFlow, a novel approach to optimize the sampling trajectory for few-step generation in pretrained diffusion and flow models. Unlike existing methods that use globally supported polynomial bases, I-SplineFlow employs integrated monotone splines (I-splines) to parameterize the stochastic interpolant (SI) scheduler. This method decouples the polynomial degree from the number of mixture weights, allowing for greater flexibility in model design. The I-spline basis ensures that the scheduler maintains fixed boundary conditions, a monotone signal-to-noise ratio (SNR), and differentiability by construction. The authors demonstrate that I-SplineFlow significantly improves few-step Fréchet Inception Distance (FID) scores compared to Bézier scheduling, particularly at lower numbers of function evaluations (NFEs), while also being computationally efficient, training in minutes. Ablation studies confirm the necessity of both the degree freedom and monotonicity constraint for optimal performance.
Methodology
The authors replace the Bézier basis with I-splines to parameterize the SI scheduler. This allows for local control over the scheduler's derivative and decouples the polynomial degree from the number of mixture weights. The I-spline basis is compactly supported, ensuring better-conditioned optimization and maintaining the necessary properties of the scheduler by construction.
Results
I-SplineFlow outperforms Bézier scheduling in most settings, particularly at low NFEs, leading to improved few-step FID scores across diffusion and flow models. The training process is significantly faster, taking only minutes, and ablation studies confirm the critical role of monotonicity and degree freedom in achieving these results.
Implications
The findings suggest that I-SplineFlow can enhance the efficiency and effectiveness of generative models, particularly in applications requiring rapid sampling and high-quality outputs. This could have significant implications for real-time applications in computer vision and other fields relying on generative modeling.
Fed-ReMasker: Federated Tabular Imputation under Feature-Level Missingness
Federated Learning
- Fed-ReMasker adapts the ReMasker masked autoencoder for federated learning, enabling effective imputation under feature-level missingness.
- The paper presents the first systematic benchmark for federated tabular imputation, evaluating various scenarios of missingness and client heterogeneity.
- Fed-ReMasker outperforms existing federated imputation methods, achieving the lowest error rates in both value-level and feature-level missingness scenarios.
- The method remains robust under client heterogeneity, with standard federated averaging performing competitively against more complex strategies.
Read more
Fed-ReMasker: Federated Tabular Imputation under Feature-Level Missingness
Summary
The paper addresses the challenges of federated learning in multi-center clinical studies, particularly focusing on feature-level missingness where entire features are unobserved at some centers. To tackle this issue, the authors introduce Fed-ReMasker, an adaptation of the ReMasker masked autoencoder for federated settings. This method allows centers to impute features that are not locally observed by leveraging knowledge from other collaborating centers. The authors conduct a comprehensive benchmark evaluation using both synthetic and real-world datasets, assessing the performance of Fed-ReMasker across various scenarios of client heterogeneity and missingness ratios. The results demonstrate that Fed-ReMasker significantly outperforms existing federated imputation methods, achieving the lowest imputation error in a majority of tested scenarios and showing robustness against client heterogeneity. This work represents a significant advancement in federated tabular imputation, providing a systematic benchmark and demonstrating the effectiveness of federated learning in handling complex missing data situations.
Methodology
The authors adapted the ReMasker masked autoencoder to a federated learning context, allowing for the imputation of features that are entirely unobserved at certain centers. They evaluated the method using a benchmark that included synthetic datasets with varying relationships and real-world clinical data, testing under different client counts, missingness ratios, and heterogeneity conditions. Various federated optimization strategies were compared, including FedAvg, FedProx, and FedAdam.
Results
Fed-ReMasker achieved the lowest imputation error in 93.2% of value-level and 96.7% of feature-level scenarios in a homogeneous benchmark. It outperformed all baselines in all 36 value-level scenarios and in at least 35 of 36 feature-level scenarios, coming within 3.0% of the performance of a centralized model trained on pooled data.
Implications
The findings suggest that Fed-ReMasker can significantly enhance the quality of data imputation in multi-center studies, enabling better model training while complying with data privacy regulations. This could lead to improved collaboration in clinical research and more robust predictive models in healthcare.
COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation
NLP
Large Language Models
Reinforcement Learning
- COPE provides a solution for continual personalization of LLMs with sparse user feedback.
- The framework utilizes learnable personalized embeddings to capture user preferences effectively.
- Self-evaluation generates proxy rewards for continuous model updates without explicit feedback.
- COPE outperforms strong baselines in various experimental settings.
Read more
COPE: Continual Personalization of LLMs under Sparse User Feedback via User Embeddings and Self-Evaluation
Summary
The paper introduces COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel framework designed to enhance the personalization of Large Language Models (LLMs) in real-world scenarios where user feedback is sparse. Traditional LLMs often produce homogenized responses due to their reliance on normative values, which can overlook diverse user preferences. COPE addresses this by assigning learnable personalized embeddings to users, allowing for continuous model updates even in the absence of explicit feedback. The framework integrates three key components: Supervised Fine-Tuning for capturing user preferences, Reinforcement Learning for calibrating self-evaluation, and Reinforcement Learning for optimizing personalized responses. This unified approach enables the model to adapt to evolving user preferences while maintaining general capabilities. Experimental results demonstrate that COPE outperforms existing training-free and training-based methods, showing reliable self-evaluation and meaningful preference patterns, while also remaining effective under shifting user preferences.
Methodology
COPE employs a unified optimization framework that integrates three objectives: Supervised Fine-Tuning for preference capture, Reinforcement Learning for self-evaluation calibration, and Reinforcement Learning for personalized response optimization. This approach allows the model to continuously update its parameters based on both explicit user feedback and proxy rewards derived from self-evaluation.
Results
Experimental evaluations indicate that COPE consistently outperforms several strong training-free and training-based baselines across different feedback probabilities. The analyses confirm that COPE's self-evaluation aligns well with actual user feedback, and the learned embeddings effectively capture user preferences. Additionally, the optimized model retains its general capabilities and demonstrates robustness to changes in user preferences and evaluators.
Implications
The COPE framework has significant implications for the development of personalized AI assistants, enabling them to better align with individual user preferences in real-time. This could enhance user satisfaction and engagement, making LLMs more effective in practical applications across various domains.
What Do Tabular Foundation Models Compute In Context? In-Situ Representation Refinement through Attention-Gated Updates
Theory
Efficient ML
Optimization
- Introduces in-situ representation refinement for tabular foundation models.
- Presents REFINEICL, an attention-gated model that operates without expanded FFNs.
- Achieves state-of-the-art performance on benchmark datasets, surpassing previous models.
- Demonstrates that support representation updates enhance query prediction accuracy.
Read more
What Do Tabular Foundation Models Compute In Context? In-Situ Representation Refinement through Attention-Gated Updates
Summary
This paper investigates the computation performed by tabular foundation models in the context of supervised learning tasks defined by labeled tables. The authors propose a novel approach called in-situ representation refinement, where support labels guide updates to the model's representations, allowing these updates to be transferred to unlabeled queries without modifying model parameters. They introduce REFINEICL, an attention-gated architecture that employs a regularized leave-one-out objective to enhance class separation and improve query predictions. The architecture avoids the use of expanded feed-forward networks (FFNs) while still achieving competitive performance. Experimental results demonstrate that REFINEICL-L24 achieves high accuracy and AUC scores on benchmark datasets, outperforming existing models like TabPFN-3 and TabPFN-v3. The study also shows that updating support representations significantly contributes to the accuracy of subsequent predictions, highlighting the importance of dynamic representation adjustments in tabular data tasks.
Methodology
The authors derive a support update mechanism based on a regularized leave-one-out objective that improves class separation. They implement this mechanism in the REFINEICL architecture, which utilizes attention-gated updates and low-rank feature interactions without relying on expanded FFNs. The model's performance is evaluated on various benchmark datasets, and internal interventions are conducted to assess the impact of support updates on query predictions.
Results
REFINEICL-L24 achieves an overall AUC of 0.93836 and an accuracy of 0.87173 on the AMLB29 dataset. It also scores 1644.8 Elo on the TabArena snapshot, outperforming TabPFN-3 by 31.4 Elo. The model improves all reported metrics over TabPFN-v3 on both TabZilla views. Additionally, skipping support updates during prediction leads to increased cross-entropy in all tested episodes, indicating the importance of these updates.
Implications
The findings suggest that dynamic updates to support representations can significantly enhance the performance of tabular foundation models. This approach could be applied to various supervised learning tasks involving tabular data, potentially leading to more efficient and accurate predictive models.
Learning Collective Dynamics with Differentiable Gaussian Representations
Time Series
- DGD connects a Gaussian response distribution with contact-behavior aggregation and feedback recurrence for predicting collective responses.
- The model allows for joint learning of population representations and dynamics from aggregate counts.
- DGD outperforms existing methods in terms of predictive accuracy and behavioral prediction metrics.
- The framework demonstrates the importance of learning population dynamics and feedback processes from aggregate observations.
Read more
Learning Collective Dynamics with Differentiable Gaussian Representations
Summary
This paper introduces Differentiable Gaussian Dynamics (DGD), a novel framework for learning collective dynamics from aggregate observations. The authors argue that collective responses are influenced by individual differences, contact opportunities, and accumulated experiences. DGD consists of three main components: a Gaussian mixture model that captures heterogeneous response propensities, a differentiable aggregation mechanism that evaluates contact intensity and behavioral probabilities, and a feedback recurrence system that updates future responses based on past observations. The model is trained using reparameterized integration and temporal recurrence, allowing for joint optimization of the distribution, observation functions, and feedback parameters. The authors evaluate DGD on datasets from KuaiRand-Pure and Online Retail II, demonstrating its effectiveness in reducing joint behavioral negative log-likelihood compared to existing methods, particularly a DeepAR adaptation. The results indicate that learning the response distribution significantly improves predictive accuracy and that feedback dynamics play a crucial role in refining predictions over time.
Methodology
The methodology involves a Gaussian mixture model to represent heterogeneous response propensities, a differentiable aggregation process to evaluate contact intensity and behavioral probabilities, and a feedback recurrence mechanism to update future responses based on past observations. The model is trained using reparameterized integration and temporal recurrence to optimize multiple parameters simultaneously.
Results
DGD achieved lower joint behavioral negative log-likelihood than the DeepAR adaptation across all evaluated windows, with a one-day behavioral-count MAE of 4.71 compared to 6.88 for DeepAR. Learning the distribution reduced behavioral negative log-likelihood by 10.82% compared to a fixed Gaussian, and removing feedback dynamics significantly increased joint KL divergence in controlled experiments.
Implications
The findings suggest that DGD can be applied in various domains where collective dynamics are important, such as social media interactions and market transactions. The ability to learn population representations and feedback processes from aggregate data could enhance predictive modeling in these areas.
TNLearn: An Open Source Python Package for Task-based Neurons
Theory
Optimization
- TNLearn enables the creation of task-based neurons tailored for specific problems.
- The package emphasizes user-friendliness and modularity for easy integration into workflows.
- Vectorized Symbolic Regression (VSR) is a key feature for discovering optimal formulas from data.
- TNLearn serves as both a practical tool for researchers and a benchmark for algorithmic designs.
Read more
TNLearn: An Open Source Python Package for Task-based Neurons
Summary
The paper introduces TNLearn, an open-source Python package designed for the automated construction and training of task-based neurons and networks. Unlike traditional neural networks that utilize a single type of neuron, TNLearn advocates for the use of customized neurons tailored to specific tasks, reflecting the biological diversity of neurons in the human brain. This approach allows for the integration of task-specific prior knowledge, potentially enhancing the performance of neural networks. TNLearn is built to be user-friendly, featuring a modular architecture that enables users to easily create task-based networks with minimal code. The package includes a Neuron Discovery Engine that implements Vectorized Symbolic Regression (VSR), allowing for the identification of optimal formulas that fit input data through genetic programming. Comprehensive documentation and examples are provided to facilitate adoption by both beginners and experienced researchers, making TNLearn a valuable tool in the machine learning ecosystem.
Methodology
TNLearn employs a modular architecture that allows users to construct task-based neurons using various search methods. It implements Vectorized Symbolic Regression (VSR) for optimal formula discovery, utilizing genetic programming to create a tree structure representing mathematical operations on input variables. The package is designed for ease of use, requiring minimal code for instantiation and training.
Results
TNLearn provides a robust framework for developing task-based neurons, demonstrating the feasibility of integrating customized neurons into neural networks. The initial implementation shows promising results in enhancing model performance by leveraging task-specific prior knowledge.
Implications
TNLearn has the potential to advance research in artificial neural networks by providing a flexible platform for exploring neuronal diversity. Its applications span various fields, including scientific research and industrial applications, where tailored solutions are crucial for complex problem-solving.
PISCES: Physics-Informed Solar-wind Convolutional autoEncoder for Space-weather Anomaly Detection and Early Warning
Time Series
- PISCES utilizes a physics-informed convolutional autoencoder for unsupervised anomaly detection in solar wind data.
- The model incorporates physical constraints into its training loss, improving anomaly detection accuracy.
- PISCES separates anomaly scores into distinct contributions, enhancing interpretability.
- The trained models show improved precision and consistency in detecting space weather anomalies.
Read more
PISCES: Physics-Informed Solar-wind Convolutional autoEncoder for Space-weather Anomaly Detection and Early Warning
Summary
The paper introduces PISCES, a Physics-Informed Solar-wind Convolutional autoEncoder designed for detecting anomalies in solar wind data to provide early warnings for space weather events. Traditional methods for space weather alerting rely on fixed thresholds, which can overlook complex magnetic and plasma structures. PISCES addresses this limitation by employing an unsupervised learning approach that incorporates physics-informed constraints into the training process, allowing it to learn from unlabeled solar wind measurements. The model's loss function integrates several physical principles, including magnetic field consistency, empirical temperature-velocity relationships, and geometric conditions based on the Parker spiral. During inference, PISCES decomposes the anomaly score into contributions from magnetic and plasma reconstruction errors, physics relations, and residual corrections. The results indicate that the trained models enhance the precision of anomaly detection and provide a more consistent ranking of physical contributions compared to untrained models. After applying a smoothing technique, the alarms generated by PISCES can precede independently observed sudden commencements, demonstrating its potential for timely space weather alerts.
Methodology
PISCES is a convolutional autoencoder trained on OMNI solar wind measurements without labeled data. Its training loss function includes terms for magnetic field consistency, empirical relationships between temperature and velocity, and penalties for changes in reconstructed quantities. The model is designed to report the contributions of different physical signatures during inference.
Results
The trained PISCES models demonstrated improved average precision in anomaly detection and provided a more consistent ordering of physical contributions compared to untrained models. The alarms generated were able to precede observed sudden commencements, indicating the model's effectiveness in early warning scenarios.
Implications
PISCES has significant implications for space weather monitoring and early warning systems, potentially enhancing the ability to protect critical infrastructure from solar wind disturbances. Its physics-informed approach may also inspire similar methodologies in other domains requiring anomaly detection.
False-science induction in autonomous scientific discovery
Theory
Robotics
Optimization
- False-science induction occurs when legitimate measurements are incorrectly paired with physical objects, leading to misleading learned associations.
- Systematic binding errors can redirect experimental efforts towards low-performing areas, even when data integrity is maintained.
- Error coherence is identified as a critical factor influencing budget misallocation in experimental loops.
- The study proposes acquisition-trace monitoring as a method to prevent erroneous experimental proposals.
Read more
False-science induction in autonomous scientific discovery
Summary
This paper investigates the phenomenon of false-science induction in closed-loop autonomous scientific discovery systems, where incorrect pairings of physical objects and measurements lead to misleading associations learned by neural surrogates. The authors demonstrate that such systematic binding errors can redirect experimental budgets towards low-performing outcomes, despite the integrity of individual measurements and marginal distributions. Through experiments involving green fluorescent protein fitness and materials band-gap prediction, the study reveals that error coherence, rather than frequency, is the key factor in budget misallocation. The findings highlight the need for monitoring and feedback mechanisms to intercept erroneous proposals before execution, establishing a record-binding reliability boundary and proposing protocols for provenance-aware closed-loop discovery.
Methodology
The authors conducted experiments in closed-loop discovery systems, specifically focusing on green fluorescent protein fitness and materials band-gap prediction. They systematically isolated the binding-to-budget pathway to analyze the effects of paired binding errors on experimental outcomes. The study employed flexible surrogates to fit the conditional record function and monitored acquisition traces to identify and mitigate false associations.
Results
The results demonstrated that coherent paired misbinding significantly redirected experimental budgets towards low-performing basins, while random swaps had minimal impact. The study established that the binding invariance boundary is exact, indicating that valid measurements can still lead to incorrect scientific conclusions when paired incorrectly. The findings underscore the importance of monitoring mechanisms to detect and correct binding errors before they influence experimental decisions.
Implications
The insights from this research could inform the design of more robust autonomous scientific discovery systems by integrating monitoring and feedback mechanisms to ensure the integrity of experimental records. This could enhance the reliability of automated experiments in fields such as materials science and biophysics, where accurate data interpretation is crucial.
RL Starts before RL: On Policy Distillation for Better Reinforcement Learning
Reinforcement Learning
Large Language Models
NLP
- OPD leads to higher final performance in RL compared to direct RL and SFT followed by RL.
- Pre-RL Pass@k does not fully account for the performance gains achieved through OPD.
- Behavioral alignment with the teacher's distribution is crucial for OPD's effectiveness.
- The choice of distillation objective (reverse-KL vs. forward-KL) affects performance before and after RL.
Read more
RL Starts before RL: On Policy Distillation for Better Reinforcement Learning
Summary
This paper investigates the effectiveness of on-policy distillation (OPD) as a preparatory step for reinforcement learning (RL). The authors demonstrate that models initialized with OPD achieve superior final performance compared to those trained directly with RL or through supervised fine-tuning (SFT) followed by RL. Notably, the benefits of OPD extend beyond mere improvements in initial accuracy, as evidenced by the observation that higher pre-RL Pass@k scores do not consistently correlate with better post-RL performance. The authors propose that OPD enhances alignment with the teacher's distribution, which may facilitate the discovery of higher-quality reasoning paths that RL can refine through feedback. The study also explores how different trajectory sources and divergence objectives impact the effectiveness of distillation for subsequent RL training. The findings suggest that the choice of distillation objective should consider both the trajectory source and the nature of the subsequent training, emphasizing the importance of evaluating OPD in the context of its downstream performance.
Methodology
The authors conducted experiments comparing three training approaches: direct RL, SFT followed by RL, and OPD followed by RL. They analyzed pre-RL Pass@k scores, behavioral alignment with teacher distributions, and the effects of different trajectory sources and divergence objectives on performance.
Results
Models initialized with OPD consistently outperformed those trained with direct RL or SFT in terms of final performance. The study found that OPD's advantages were not solely explained by initial accuracy metrics, and that the choice of distillation method influenced outcomes both before and after RL training.
Implications
The findings suggest that OPD can serve as an effective preparatory stage for RL, potentially leading to more efficient training processes in various applications of reinforcement learning. The insights into trajectory sources and distillation objectives may guide future research and practical implementations in RL.
Theory for groupoid equivariant neural networks: an approach for steerable CNNs on bounded domains
Theory
- Introduces groupoid-equivariant neural networks for bounded domains.
- Establishes a bisection-equivariant kernel theorem linking equivariance to transport constraints.
- Demonstrates application to bounded planar domains with distinct isotropy groups.
- Implements architecture using offline nullspace bases and sparse operations.
Read more
Theory for groupoid equivariant neural networks: an approach for steerable CNNs on bounded domains
Summary
This paper presents a novel framework for groupoid-equivariant neural networks, addressing the limitations of traditional equivariant convolutional neural networks (CNNs) that rely on global group actions. The authors argue that such global actions are often inappropriate for bounded or stratified domains, where symmetry may only apply locally. They develop a theory that incorporates groupoids, local bisections, and representation bundles to define equivariant neural networks. A key contribution is the bisection-equivariant kernel theorem, which establishes that equivariance corresponds to specific transport constraints on the kernel. The authors demonstrate the application of their theory to bounded planar domains, revealing how different isotropy groups can be identified at bulk, edge, and corner points. The architecture is implemented using offline nullspace bases and sparse operations, and numerical experiments validate the kernel constraints and composition theorem, showing significant improvements in accuracy compared to traditional models. The findings suggest that the proposed architectures are particularly effective when global symmetries cannot be applied, achieving an accuracy enhancement of at least one order of magnitude.
Methodology
The authors develop a theoretical framework for groupoid-equivariant neural networks, focusing on local symmetries and their representation through groupoids. They prove a bisection-equivariant kernel theorem and apply this theory to specific bounded planar domains, analyzing isotropy groups and their implications for neural network architecture. The implementation involves numerical methods to validate the theoretical findings.
Results
The proposed architecture demonstrates improved performance in scenarios where global symmetries are not applicable, achieving accuracy enhancements of at least one order of magnitude compared to existing models. The numerical experiments confirm the validity of the kernel constraints and the composition theorem, showcasing the architecture's effectiveness in boundary-aware tasks.
Implications
This work has significant implications for the design of neural networks in computer vision and other fields where local symmetries are prevalent. The ability to incorporate groupoid symmetries could lead to more efficient and accurate models for tasks involving complex geometries and boundaries.
Beyond Imitation: Auditing the Recoverability of Reasoning in Distilled Models
NLP
Large Language Models
Theory
- Prefix recovery measures the ability of student models to complete reasoning tasks based on partial teacher solutions.
- Reverse-KL distillation shows significant improvements in performance, especially for smaller student models.
- As student model size increases, both prefix recovery rates and the gap between robust and fragile recovery decrease.
- The study establishes a link between recoverability of reasoning traces and the objective conflict in supervision signals.
Read more
Beyond Imitation: Auditing the Recoverability of Reasoning in Distilled Models
Summary
This paper investigates the compatibility of reasoning transfer from teacher models to student models in the context of knowledge distillation. The authors introduce the concept of prefix recovery, which measures how well a student model can complete a reasoning task after being provided with partial information from a verified teacher solution. The study connects prefix recovery to the cosine conflict between cross-entropy and reverse-KL gradients, revealing that reverse-KL distillation significantly enhances performance, particularly for smaller student models. The experiments demonstrate that as the size of the student model increases, the average prefix recovery improves from 71.0% to 91.9%, and the gap between robust and fragile recovery diminishes. The findings suggest that the effectiveness of distillation is closely tied to the recoverability of reasoning traces and the objective conflict between supervision signals. The paper provides a framework for understanding the conditions under which knowledge distillation is most effective, highlighting the importance of model capacity and trajectory compatibility.
Methodology
The authors conducted experiments using adjacent Qwen3 teacher-student pairs with varying parameter sizes. They measured prefix recovery by revealing portions of verified solutions and assessing student completions. The study also analyzed the cosine conflict between supervision signals generated by cross-entropy and reverse-KL methods to evaluate compatibility and performance improvements.
Results
The results indicate that the 0.6B student model achieved accuracy improvements of 4.30 and 3.66 points on OlympiadBench and HumanEval, respectively. The prefix recovery rate increased from 71.0% at 0.6B to 91.9% at 4B, with the robust-fragile recovery gap narrowing significantly. Additionally, the objective conflict separation decreased from 0.993 to 0.233 as model capacity increased.
Implications
The findings suggest that understanding the recoverability of reasoning in distilled models can inform the selection of teacher-student pairs in knowledge distillation. This has potential applications in optimizing model training processes and improving the performance of smaller models in various NLP tasks.
A Spectral Theory of Grokking: Weight Decay induces Feature Learning
Theory
Optimization
- Grokking separates the timescales of fitting and generalization in neural networks.
- Weight decay induces residual errors that drive continued evolution of the NTK.
- The timescale for grokking is controlled by the product of learning rate and weight decay.
- Empirical results confirm the predicted phase structure in both MLPs and Transformers.
Read more
A Spectral Theory of Grokking: Weight Decay induces Feature Learning
Summary
This paper presents a quantitative theory of 'grokking', a phenomenon where a neural network initially fits training data but later improves generalization significantly. The authors explore the transition from a fixed neural tangent kernel (NTK) regime to one where task-relevant kernel eigendirections evolve, particularly under the influence of L2 weight decay. They demonstrate that a finite residual error remains after the initial fit, with larger residuals linked to smaller NTK eigenvalues. This residual error influences the NTK dynamics, leading to a competition between residual-driven kernel growth and weight decay. The authors derive that the timescale for grokking is governed by the product of learning rate and weight decay, and they observe logarithmic slowing of feature learning near a critical decay threshold. Empirical tests on modular addition tasks in both homogeneous multi-layer perceptrons (MLPs) and Transformers validate their theoretical predictions, revealing a structured relationship between learning rate, weight decay, and generalization time. The findings connect post-fit feature learning to the onset of generalization and its phase structure in the learning rate and weight decay parameter space.
Methodology
The authors utilize a theoretical framework based on the neural tangent kernel (NTK) to analyze the dynamics of generalization in homogeneous networks trained with squared loss and L2 weight decay. They derive a reduced system that captures the interaction between residual errors and NTK evolution, and they conduct empirical tests on modular addition tasks using both homogeneous MLPs and Transformers to validate their theoretical predictions.
Results
The study finds that after initial training accuracy saturation, task-aligned NTK structure continues to evolve, leading to delayed generalization. The empirical results from a grid of trained networks across varying learning rates and weight decay confirm the predicted scaling of generalization time with these parameters, demonstrating a consistent phase structure in both MLPs and Transformers.
Implications
The findings have significant implications for understanding the dynamics of neural network training, particularly in optimizing hyperparameters like learning rate and weight decay to enhance generalization. This could inform the design of more effective training regimes for various machine learning tasks.
NS-ATTENTION: Newton-Schulz Transformations of Attention Outputs in Vision Transformers
Computer Vision
Optimization
Efficient ML
- Introduction of NS-Attention, a parameter-free transformation for attention outputs in Vision Transformers.
- NS-Attn. reduces spectral concentration and increases effective rank of attention representations.
- Demonstrated improvements in classification accuracy on CIFAR-10 and CIFAR-100 datasets.
- Eigenvalue analysis confirms the intended spectral effects of the transformation.
Read more
NS-ATTENTION: Newton-Schulz Transformations of Attention Outputs in Vision Transformers
Summary
This paper introduces NS-Attention (NS-Attn.), a novel transformation applied to the outputs of attention heads in Vision Transformers. The method leverages Newton-Schulz (NS) iterations to modify the spectral properties of attention outputs, aiming to reduce spectral concentration and enhance effective rank. The transformation is parameter-free and operates on the feature-by-token matrices produced by each attention head, normalizing them before applying a finite NS polynomial step and restoring the original norm. The authors evaluate NS-Attn. on Vision Transformer (ViT) and Swin Transformer architectures using CIFAR-10 and CIFAR-100 datasets, demonstrating improvements in classification accuracy across all tested configurations. The results indicate that NS-Attn. effectively redistributes energy in attention representations, leading to better performance, especially on more challenging tasks. However, the method incurs additional inference latency due to the computational overhead of the transformation.
Methodology
The NS-Attn. method involves normalizing the attention output matrices, applying a finite polynomial transformation inspired by Newton-Schulz iterations, and restoring the original Frobenius norm. This process is applied independently to each head's output before head merging and output projection.
Results
The NS-Attn. method resulted in mean accuracy improvements of 0.25–0.83 percentage points across 12 matched-seed comparisons on CIFAR-10 and CIFAR-100. Eigenvalue analysis showed reduced concentration of leading eigenvalues and increased effective rank compared to standard attention mechanisms.
Implications
The findings suggest that NS-Attn. can enhance the performance of Vision Transformers in image classification tasks, particularly in more complex datasets. This approach could be beneficial for future research in optimizing attention mechanisms in deep learning models.
HARN: Hierarchical Associative Resonance Network for Event-Driven Multi-Timeframe Forecasting
Time Series
- HARN maintains persistent representations across multiple temporal levels, updating only when new information is available.
- The architecture combines several advanced techniques, including causal multi-scale encoding and gated associative memory.
- HARN achieves competitive forecasting accuracy compared to single-timeframe models, highlighting its effectiveness in multi-timeframe scenarios.
- The paper provides a thorough evaluation of the model, including ablation studies to assess the impact of individual components.
Read more
HARN: Hierarchical Associative Resonance Network for Event-Driven Multi-Timeframe Forecasting
Summary
This paper introduces the Hierarchical Associative Resonance Network (HARN), a novel framework designed for event-driven multi-timeframe forecasting in financial time series. Traditional forecasting systems struggle with the asynchronous nature of financial data, where observations at different temporal resolutions do not align. HARN addresses this challenge by maintaining persistent representations across multiple temporal levels and updating them only when new information becomes available. The architecture integrates causal multi-scale temporal encoding, gated associative memory, cross-level resonance, and hierarchical evidence aggregation. Forecasting is conducted in basis-point space and later reconstructed to the original price scale. The model is evaluated on four financial assets (AAPL, EURUSD, USDCHF, and XAUUSD) and demonstrates competitive performance against single-timeframe baselines, PatchTST and TimeXer. The paper emphasizes the importance of an event-driven approach that allows for asynchronous updates and persistent states, rather than claiming universal predictive superiority. Additionally, a code-level audit is performed to ensure consistency with the defined event-driven causal protocol, further validating the framework's robustness.
Methodology
HARN employs a hierarchical structure where each temporal level maintains its own state and updates asynchronously based on completed-bar indicators. The model utilizes causal multi-scale encoding for temporal representation, gated associative memory for information storage, and cross-level resonance for interaction between different temporal levels. Forecasting is performed in basis-point change space and reconstructed for evaluation.
Results
HARN was evaluated on four financial assets and showed competitive performance in reconstructed-price forecasting errors when compared to single-timeframe baselines. The ablation studies provided insights into the contributions of individual components, confirming the effectiveness of the model's architecture.
Implications
The HARN framework can significantly enhance forecasting accuracy in financial markets by effectively managing the complexities of multi-timeframe data. Its event-driven approach may also be applicable in other domains where asynchronous data streams are prevalent.
SR-Fraud: An Outcome-Supervised Reflective LLM Agent Framework for Non-Stationary Payment Fraud Detection
Large Language Models
Time Series
Efficient ML
- SR-Fraud decouples real-time decision-making from offline learning, addressing the adaptation gap in fraud detection.
- The framework utilizes a Hybrid Episodic Window to contextualize transactions against user behavior and velocity signals.
- An outcome-supervised reflection mechanism proposes and verifies hypotheses to enhance the decision agent's knowledge state.
- SR-Fraud outperforms traditional tabular classifiers and other LLM-based methods in detecting non-stationary fraud patterns.
Read more
SR-Fraud: An Outcome-Supervised Reflective LLM Agent Framework for Non-Stationary Payment Fraud Detection
Summary
The paper introduces SR-Fraud, a novel framework designed to enhance real-time payment fraud detection in non-stationary environments. Traditional methods, such as tabular classifiers and rule engines, struggle to adapt to rapidly evolving fraud patterns, especially during localized burst attacks. SR-Fraud addresses this challenge by decoupling real-time decision-making from offline learning processes. The framework employs a stateless decision agent that evaluates transactions using a Hybrid Episodic Window, which combines recent user behavior with velocity signals to detect anomalies. An offline reflection agent analyzes matured errors to propose new hypotheses, which are then verified and integrated into an evolving knowledge state. This architecture allows for efficient adaptation to new fraud patterns without compromising latency requirements. The authors benchmark SR-Fraud against existing methods, demonstrating its superior performance in detecting emerging fraud bursts and improving detection metrics over traditional classifiers.
Methodology
The SR-Fraud framework consists of a stateless decision agent that scores transactions based on a Hybrid Episodic Window, which captures both recent user behavior and velocity signals. An offline reflection agent analyzes matured errors to propose new hypotheses, which are verified by a deterministic harness before being added to an evolving knowledge state. This modular design allows for efficient processing and adaptation without impacting real-time performance.
Results
SR-Fraud demonstrated significant improvements in detection metrics on a production payment-fraud benchmark compared to its frozen decision agent and traditional classifiers like CatBoost. It effectively identified emerging fraud bursts, showcasing its capability to adapt to non-stationary fraud patterns.
Implications
The SR-Fraud framework has the potential to revolutionize payment fraud detection systems by providing a robust solution that adapts to evolving threats in real-time. Its modular design could be applied to various financial platforms, enhancing security and reducing losses from fraud.
Deep Reinforcement Learning on Item-Compatibility Graphs for One-Dimensional Bin Packing
Reinforcement Learning
Graph Learning
Optimization
- Introduces a size-agnostic graph reinforcement learning framework for 1D-BPP.
- Utilizes an item-compatibility graph to represent the packing process as an MDP.
- Achieves strong zero-shot generalization to unseen problem scales.
- Demonstrates improved performance over traditional heuristics and other learned methods.
Read more
Deep Reinforcement Learning on Item-Compatibility Graphs for One-Dimensional Bin Packing
Summary
This paper addresses the one-dimensional bin packing problem (1D-BPP), a well-known NP-hard combinatorial optimization problem, by proposing a novel end-to-end, size-agnostic graph reinforcement learning framework. The authors formulate the packing process as a Markov decision process (MDP) on an item-compatibility graph, where actions involve merging compatible bins. A graph neural network (GNN) actor-critic policy is employed to extract relational features from this graph representation, enabling the model to generalize to instances of any size without retraining. The paper conducts a comprehensive empirical study comparing various graph encoders and reinforcement learning algorithms, demonstrating that the proposed method significantly reduces the optimality gap compared to traditional heuristics and other learned methods. The results indicate that the model is particularly effective on structured instances and outperforms state-of-the-art learned solvers in certain scenarios, while also providing insights into the residual gaps in performance against a grouping genetic algorithm. Overall, this work connects classical combinatorial optimization with modern reinforcement learning techniques, offering a robust solution for practical applications in logistics and resource management.
Methodology
The authors define the 1D-BPP as an MDP on an item-compatibility graph, where nodes represent bins and edges represent compatible item pairs. A GNN actor-critic policy is trained through reinforcement learning, with actions involving the merging of edges in the graph. The model is evaluated using various graph encoders and reinforcement learning algorithms, with a focus on in-distribution validation for design choices.
Results
The proposed model reduces the mean optimality gap of a constructive heuristic from 2.66% to 2.31% on the BPPLIB benchmark, showing significant improvements on structured instances. It also outperforms a state-of-the-art learned solver that uses column generation and integer programming, while maintaining stability across different instance distributions.
Implications
This research has potential applications in logistics, manufacturing, and cloud resource management, where efficient bin packing is crucial. The framework can serve as a decision-making component in intelligent systems that require scalable and adaptable packing solutions.
CacheDyG: Decoupling Temporal Propagation for Efficient Dynamic Graph Learning
Graph Learning
Efficient ML
Optimization
- CacheDyG decouples temporal propagation from parameter updates, improving efficiency.
- The framework utilizes a Temporal Dependency Cache to store reusable node-time representations.
- CacheDyG achieves better predictive performance with fewer trainable parameters compared to existing methods.
- The proposed method reduces runtime and memory consumption, making it feasible for larger datasets.
Read more
CacheDyG: Decoupling Temporal Propagation for Efficient Dynamic Graph Learning
Summary
The paper presents CacheDyG, a novel framework designed to enhance the efficiency of dynamic graph learning by decoupling temporal propagation from routine parameter updates. Dynamic graphs are essential for modeling time-evolving relational systems, but existing dynamic graph neural networks often face inefficiencies due to the repeated computation of largely unchanged historical structures during training. CacheDyG addresses this issue by introducing a Temporal Dependency Cache that stores non-trainable, graph-aware node-time representations. This allows the model to read from the cache during training epochs and update only a lightweight cache refiner and a link predictor, significantly reducing the number of trainable parameters and runtime. The authors demonstrate that CacheDyG achieves competitive predictive performance across five dynamic graph benchmarks while maintaining a low computational overhead, thus providing a scalable solution for dynamic graph learning.
Methodology
CacheDyG constructs a time-ordered Temporal Dependency Cache that aggregates current and past information for node-time representations. During training, it selectively reads from this cache and updates only a compact cache refiner and a link predictor, avoiding the need for repeated sparse propagation. A selective cache-refresh mechanism ensures that the cached representations remain aligned with the supervised objectives.
Results
CacheDyG was evaluated on five dynamic graph benchmarks for one-step future-link prediction. It achieved the best average precision on all datasets while using only 19.106K trainable parameters. Additionally, it demonstrated the lowest runtime and maintained feasibility on larger datasets where traditional memory-intensive methods failed. Ablation studies confirmed the contributions of various components of the framework to its performance.
Implications
The findings suggest that CacheDyG can be applied to various real-world applications involving dynamic graphs, such as social networks, recommender systems, and biological interaction networks, where efficient learning from evolving data is crucial.
TinyUDE: Solver-Free Universal Differential Equations on Microcontrollers via Lie-Taylor Jet Matching
Robotics
Efficient ML
Time Series
- Introduces a solver-free training framework for Universal Differential Equations (UDEs).
- Utilizes Lie-Taylor jet matching for direct fitting of hybrid vector fields to time-derivatives.
- Achieves significant memory savings, making real-time training feasible on microcontrollers.
- Demonstrates robustness against sensor noise through noise-adaptive mechanisms.
Read more
TinyUDE: Solver-Free Universal Differential Equations on Microcontrollers via Lie-Taylor Jet Matching
Summary
The paper introduces TinyUDE, a novel framework for training Universal Differential Equations (UDEs) on microcontrollers without relying on traditional numerical ODE solvers. The authors propose a method called Lie-Taylor jet matching, which fits a hybrid vector field directly to the first and second time-derivatives of observed system states. This approach utilizes Savitzky-Golay filtering to estimate these derivatives online, allowing for the computation of fully analytic gradients without the need for automatic differentiation. The study evaluates the accuracy of this solver-free method against conventional baselines, demonstrating that while naive derivative matching suffers under sensor noise, the proposed noise-adaptive mechanisms effectively mitigate this issue. The results show that TinyUDE matches or exceeds the accuracy of traditional methods while significantly reducing memory requirements, making it feasible for real-time on-device training on resource-constrained microcontrollers. The experiments conducted on systems like a damped pendulum and chaotic double pendulum validate the effectiveness of the proposed method, achieving a geometric-mean relative field error of 0.65 times that of the baseline, with a memory footprint of only 108 kB.
Methodology
The methodology involves using Lie-Taylor jet matching to fit hybrid vector fields to the first and second time-derivatives of observed states, estimated via Savitzky-Golay filtering. This allows for the computation of analytic gradients without automatic differentiation. The approach incorporates noise-adaptive mechanisms to enhance robustness against sensor noise, including full-rate phase-shifted sampling, reservoir buffers, and polynomial-misfit quality gating.
Results
The proposed method achieved a geometric-mean relative field error of 0.65 times that of the baseline methods across noise levels from 0% to 5%. On an ESP32 microcontroller, it reached a field error of 0.0020 while recovering the damping coefficient accurately within a memory footprint of 61.3 kB and an update time of 7.24 ms.
Implications
TinyUDE has significant implications for deploying machine learning models in edge computing environments, particularly in applications involving physical systems that require real-time adaptation and training without the computational overhead of traditional solvers. This could enhance the capabilities of robotic systems, UAVs, and other embedded AI applications.
Minimal-Norm Univariate Two-Layer ReLU Classification: Exact Solutions and Global Optimality with Skip Connections
Optimization
Theory
- Exact geometric characterizations of minimal-norm interpolators and regularized loss minimizers are established.
- The presence of skip connections improves the optimization landscape, making all KKT points globally optimal.
- Unique minimizers with specific structural properties arise depending on bias penalization.
- Sparsity-like restrictions are identified in the context of unpenalized biases.
Read more
Minimal-Norm Univariate Two-Layer ReLU Classification: Exact Solutions and Global Optimality with Skip Connections
Summary
This paper investigates minimal-norm interpolation and â„“2-regularized logistic-loss minimization for binary classification using univariate two-layer ReLU networks. The authors provide comprehensive geometric characterizations of optimal classifiers, detailing how solutions vary based on the inclusion of hidden-layer biases in the parameter norm. When biases are unpenalized, minimal-norm interpolators are continuous piecewise-affine functions that align with label switches. Conversely, penalizing biases results in unique minimizers with a single kink per same-label segment, leading to sparsest positive-margin classifiers. The introduction of affine skip connections does not alter function-space solutions but significantly enhances the parameter-space landscape, ensuring that all KKT points of the constrained problem are globally optimal. The paper also explores the implications of weak â„“2-regularization on logistic loss minimization, revealing additional sparsity-like restrictions in the unpenalized-bias scenario. Numerical experiments validate the theoretical findings across various dataset complexities and network configurations.
Methodology
The authors employed geometric analysis to characterize optimal classifiers in function space, examining the effects of bias penalization and skip connections. They also conducted numerical experiments to validate their theoretical findings across different dataset complexities and network widths.
Results
The study found that minimal-norm interpolators exhibit distinct geometric properties based on bias treatment, with skip connections enhancing the global optimality of KKT points. The results indicate that most minimal-norm interpolators cannot be derived as limits of small-regularization logistic-loss minimizers, highlighting the complexity of the relationship between these two approaches.
Implications
The findings have significant implications for the design and training of neural networks, particularly in understanding how architectural choices like skip connections influence optimization and generalization. This work could inform future research on network architectures and regularization strategies in machine learning.
A Scaling Study for fMRI Foundation Models
Multimodal
- The study provides a controlled empirical analysis of scaling in fMRI foundation models, utilizing extensive data and compute resources.
- Performance improvements are linked to the combination of pretraining data size and model size, particularly for larger models.
- Increasing pretraining data generally yields more consistent performance benefits across various tasks compared to increasing model size.
- The research highlights that compute alone does not determine performance; the interaction of data, model size, and training duration is crucial.
Read more
A Scaling Study for fMRI Foundation Models
Summary
This paper investigates the scaling relationships among data, model size, and compute in the context of functional magnetic resonance imaging (fMRI) foundation models. The authors conducted a comprehensive empirical study utilizing pretraining data from over 200 source datasets and more than 10,000 GPU-hours of experiments. By maintaining a fixed pretraining framework and downstream protocol, they varied pretraining data size, model size, and training duration. The findings indicate that while downstream performance generally improves with increased compute, models with similar compute can yield significantly different results. Notably, larger models benefit more from additional pretraining data, suggesting that data and model size should be scaled together. The study also reveals that at matched compute, increasing pretraining data tends to provide more consistent performance gains across tasks compared to increasing model size, which varies more by task. The authors selected optimal configurations based on in-distribution performance and evaluated them on out-of-distribution tasks, achieving superior performance with less pretraining compute. Overall, the results emphasize the need to consider the interplay of pretraining data, model size, and training duration in fMRI scaling, rather than relying solely on compute as a performance metric.
Methodology
The authors conducted a controlled scaling study by fixing the pretraining objective, architecture, and downstream protocol while varying pretraining data size, model size, and training duration. They evaluated multiple configurations across 12 downstream tasks using frozen representations and tracked performance over time.
Results
The study found that downstream performance generally improves with additional training, but models with similar compute can perform differently. Larger models showed greater gains from additional pretraining data, especially later in training. At matched compute, increasing pretraining data provided more consistent performance improvements across tasks than increasing model size.
Implications
The findings suggest that in fMRI model development, careful consideration of the interplay between pretraining data, model size, and training duration can lead to more effective models. This has implications for future research and applications in neuroimaging and related fields, where optimizing model performance is critical.
Label-Efficient Learning for Ground-Based Sky-Image Classification: A Benchmark of Transfer Learning, Active Learning, and Pseudo-Labeling on GCD
Computer Vision
Efficient ML
- Supervised transfer learning significantly improves label efficiency for cloud classification.
- Active learning and pseudo-labeling provide limited additional gains compared to a strong supervised baseline.
- Accepted pseudo-labels are highly reliable but biased towards easier cloud types.
- Uncertainty sampling targets challenging samples but yields modest performance improvements.
Read more
Label-Efficient Learning for Ground-Based Sky-Image Classification: A Benchmark of Transfer Learning, Active Learning, and Pseudo-Labeling on GCD
Summary
This paper addresses the challenge of accurate ground-based cloud classification, which is crucial for various applications such as atmospheric monitoring and solar energy forecasting. The authors benchmark three label-efficient learning strategies—supervised transfer learning, uncertainty-based active learning, and high-confidence pseudo-labeling—using the Ground-based Cloud Dataset (GCD). They utilize a frozen ImageNet-pretrained ResNet50 as a common backbone and evaluate the performance across different label budgets ranging from 1% to 100%. The findings reveal that supervised transfer learning is highly label-efficient, with test accuracy improving significantly as more labels are used. Active learning and pseudo-labeling show competitive performance but do not yield substantial aggregate gains over the supervised baseline. The study highlights the reliability of accepted pseudo-labels and the targeted nature of uncertainty sampling, which focuses on challenging samples. Overall, the results indicate that while transfer learning can greatly reduce annotation needs, simple active and semi-supervised strategies provide limited additional benefits.
Methodology
The authors benchmark three learning strategies: supervised transfer learning, uncertainty-based active learning, and high-confidence pseudo-labeling, using a frozen ImageNet-pretrained ResNet50. They conduct experiments across five random seeds with varying label budgets from 1% to 100% of the training labels, evaluating performance using overall and class-balanced metrics.
Results
Supervised transfer learning achieved a test accuracy of 0.635 ± 0.018 with 1% labels, improving to 0.730 ± 0.002 with 40% labels, and approaching full-label performance of 0.735 ± 0.003. Active learning and pseudo-labeling provided small improvements for some metrics but did not consistently outperform the supervised baseline. Pseudo-label accuracy ranged from 0.946 to 0.977, while uncertainty sampling focused on challenging groups but resulted in modest gains.
Implications
The findings suggest that transfer learning can significantly reduce the need for labeled data in cloud classification tasks, which is crucial for practical applications in atmospheric monitoring and related fields. The study also emphasizes the need for careful consideration of label-efficient strategies in scenarios with limited annotation budgets.
DCRL: Decoupling and Coupling Reinforcement Learning via Policy-Reward Manifold Alignment
Reinforcement Learning
Large Language Models
Theory
- DCRL framework addresses misalignment issues in existing RL reward systems.
- Introduces a dynamic reward rubric refinement mechanism based on syllogistic logic.
- Implements a policy-reward re-coupling mechanism for consistent model updates.
- Demonstrates superior performance of DCRL over traditional rule-based and reward-model systems.
Read more
DCRL: Decoupling and Coupling Reinforcement Learning via Policy-Reward Manifold Alignment
Summary
This paper introduces the Decoupling and Coupling Reinforcement Learning (DCRL) framework, which addresses the limitations of existing reinforcement learning (RL) reward systems, particularly those based on rules and learned models. The authors conceptualize the reasoning process of large language models (LLMs) as a geometric structure composed of three interdependent sub-manifolds: logical deduction, evaluation, and representation. They argue that traditional reward systems often suffer from misalignment between policy and reward manifolds, leading to issues like unstable optimization and reward hacking. The DCRL framework proposes two innovative mechanisms: a syllogistic logic-based prompt evolution mechanism that dynamically refines reward rubrics, and a policy-reward re-coupling mechanism that ensures consistent updates between reward and policy models. Theoretical analyses and extensive experiments demonstrate that DCRL significantly outperforms both rule-based and reward-model baselines, showcasing its effectiveness in enhancing the reasoning capabilities of LLMs.
Methodology
The authors employ a geometric perspective to analyze the reasoning capabilities of LLMs, conceptualizing it as a coupled manifold of logical deduction, evaluation, and representation. They develop the DCRL framework, which includes a dynamic reward rubric refinement mechanism and a policy-reward re-coupling mechanism. The framework is evaluated through theoretical analysis and extensive experiments across various reasoning domains.
Results
The DCRL framework consistently outperformed both rule-based and reward-model baselines in multiple reasoning tasks. Notably, a Qwen3-4B model trained under DCRL surpassed a Qwen3-32B baseline and approached the performance of a Qwen3-235B model, indicating significant improvements in effectiveness and generalization.
Implications
The findings suggest that the DCRL framework could lead to more robust and adaptable RL systems, particularly in applications involving LLMs. By addressing the limitations of existing reward systems, DCRL may enhance the quality of model-generated responses and improve the overall reasoning capabilities of AI systems.
Scalable Subgraph Sampling via Resistance Curvature
Graph Learning
- Introduction of ERC-LG, an efficient approximation method for effective resistance curvature in large-scale graphs.
- Utilization of Johnson-Lindenstrauss projections and multi-GPU batched conjugate gradient solvers to enhance computational efficiency.
- Development of sampling methods that integrate edge-level geometric information for improved GNN training.
- Demonstrated superior performance in node classification tasks across multiple real-world datasets.
Read more
Scalable Subgraph Sampling via Resistance Curvature
Summary
This paper presents a novel framework for subgraph sampling in large-scale graph neural networks (GNNs) that leverages effective resistance curvature (ERC) to improve sampling efficiency and accuracy. The authors introduce ERC-LG, a method that approximates ERC using Johnson-Lindenstrauss (JL) projections combined with regularized multi-GPU batched conjugate gradient solvers. This approach avoids the computational burden of explicit Laplacian pseudoinverse calculations and the need for full embedding storage, making it scalable for large graphs. The proposed sampling methods incorporate edge-level geometric information to enhance the construction of training subgraphs for GNNs. Experimental results demonstrate that ERC-LG-based sampling variants achieve superior mean accuracy on six out of seven real-world datasets in downstream node classification tasks, while also showing significant reductions in runtime compared to traditional methods.
Methodology
The methodology involves approximating effective resistance curvature using JL projections to reduce the number of linear systems needed for resistance estimation. The authors employ a regularized multi-GPU batched conjugate gradient approach to avoid explicit pseudoinversion and full embedding storage, thus enhancing scalability for large graphs. The resulting edge geometry is then used to inform node and edge sampling probabilities for constructing subgraphs for GNN training.
Results
The experiments reveal that ERC-LG-based sampling variants achieve the highest mean accuracy on six out of seven real-world datasets in downstream node classification tasks. Additionally, the proposed method demonstrates a significant reduction in runtime compared to traditional curvature-based computation methods, validating its efficiency and effectiveness.
Implications
The findings suggest that the ERC-LG framework can significantly enhance the training of GNNs on large-scale graphs, making it a valuable tool for applications in various domains that rely on graph data, such as social network analysis, recommendation systems, and biological network modeling.
Learning Local Heterogeneity and Cross-Region Context for Large-Scale Traffic Forecasting
Graph Learning
Time Series
Optimization
- LoReST effectively models local spatial dependencies and cross-region context for traffic forecasting.
- The framework incorporates relation-aware local aggregation to address heterogeneous dependencies.
- Inter-region attention allows for efficient long-range context exchange without high computational costs.
- Experiments show substantial reductions in forecasting errors across multiple metrics.
Read more
Learning Local Heterogeneity and Cross-Region Context for Large-Scale Traffic Forecasting
Summary
This paper addresses the challenge of large-scale traffic flow forecasting by proposing a novel framework called LoReST (Local-Region Spatial Temporal network). The authors highlight the importance of capturing local spatial dependencies and cross-region context in traffic forecasting, as these factors significantly influence the accuracy of predictions. Traditional methods often struggle with the computational costs associated with modeling all-pairs node interactions, especially in large networks. LoReST tackles this by employing a two-tiered approach: it first aggregates local spatial dependencies through relation-aware local aggregation, which considers the unique characteristics of road identity and travel direction. Then, it constructs region representations using mean pooling and inter-region attention to facilitate long-range context exchange. The integration of these two components allows LoReST to efficiently learn spatial dependencies across large-scale road networks. The effectiveness of the proposed method is validated through experiments on four datasets from the LargeST benchmark, demonstrating significant improvements in forecasting accuracy compared to existing methods.
Methodology
The authors developed LoReST, which consists of two main components: a relation-aware local aggregation module that captures heterogeneous dependencies among geographically neighboring nodes, and a cross-region interaction mechanism that constructs region representations and facilitates long-range context exchange through attention mechanisms. This dual approach allows for effective spatial dependency learning while managing computational efficiency.
Results
The experiments conducted on four datasets from the LargeST benchmark revealed average relative reductions of 4.78% in Mean Absolute Error (MAE), 3.60% in Root Mean Square Error (RMSE), and 5.75% in Mean Absolute Percentage Error (MAPE), indicating that LoReST outperforms existing traffic forecasting models.
Implications
The findings suggest that LoReST can significantly enhance traffic forecasting accuracy in intelligent transportation systems, potentially aiding in congestion management, traffic control, and route planning. The methodology could also be adapted for other applications requiring spatial-temporal analysis in large-scale networks.
An open benchmark for machine learning-based polymer property prediction
Graph Learning
- PolyBench26 is the first open benchmark combining homopolymers and copolymers for polymer property prediction.
- The dataset includes nearly 250,000 datapoints across eight physical properties, enhancing the scope of polymer informatics.
- Graph-based models demonstrate superior performance in property prediction compared to language and descriptor-based models.
- The benchmark supports various evaluation tasks, allowing for systematic assessment of ML models in polymer property prediction.
Read more
An open benchmark for machine learning-based polymer property prediction
Summary
The paper introduces Polymer Benchmark 2026 (PolyBench26), a comprehensive open dataset designed to facilitate the rigorous comparison of machine learning (ML) methods for polymer property prediction. Existing benchmarks in polymer informatics have been limited, primarily focusing on homopolymers and lacking coverage of more complex architectures such as copolymers. PolyBench26 addresses this gap by providing nearly 250,000 polymer-property datapoints across eight physical properties, sourced from experimental measurements, density functional theory, and molecular dynamics. The benchmark supports four evaluation tasks: in-distribution property prediction, dataset-size scaling, repeat-unit complexity, and transfer learning to held-out polymer architectures. The authors compare various modeling approaches, including language model, graph-based, and descriptor-based methods, finding that graph-based models consistently outperform others in terms of prediction accuracy and robustness across different training set sizes and complexities. This benchmark aims to establish a reproducible foundation for advancing polymer design through machine learning.
Methodology
The authors developed PolyBench26 by aggregating existing polymer datasets and extracting new data from molecular dynamics simulations. They evaluated three classes of polymer vectorization schemes (language model, graph-based, and descriptor-based) across four benchmark tasks to assess model performance in different scenarios, including in-distribution prediction and transfer learning.
Results
Graph-based models achieved the lowest prediction errors and maintained their advantages across varying dataset sizes and complexities. The evaluation tasks revealed insights into model performance, particularly in low-data regimes and with increasing structural complexity of repeat units. The results indicate that graph-based approaches are particularly effective for both homopolymers and alternating copolymers.
Implications
PolyBench26 provides a standardized framework for the polymer informatics community, enabling researchers to benchmark and improve ML models for polymer property prediction. This can accelerate the discovery and design of advanced polymer materials with tailored properties for various applications.
MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning
Reinforcement Learning
Large Language Models
Graph Learning
- MAGIC allows local granularity selection for functional roles in multi-agent systems, improving collaboration efficiency.
- The framework uses dense-reward reinforcement learning to optimize the construction of agent graphs incrementally.
- Potential-based reward shaping provides intermediate feedback, enhancing learning from limited feedback scenarios.
- MAGIC outperforms state-of-the-art baselines across multiple benchmarks, demonstrating its effectiveness.
Read more
MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning
Summary
The paper introduces MAGIC, a novel framework for generating mixed-granularity agent graphs in multi-agent systems (MAS) using dense-reward reinforcement learning. Traditional approaches to collaboration topology in MAS either fix the granularity at the agent level or group level, which fails to accommodate the varying collaboration needs of different subtasks. MAGIC addresses this limitation by allowing local selection of granularity for each functional role, enabling a combination of fine-grained control and reusable collaboration patterns. The framework constructs agent graphs incrementally, optimizing the selection of roles and their connections based on returns from trajectories sampled under the current policy. It employs potential-based reward shaping to provide intermediate feedback during the construction process, enhancing the learning efficiency. The authors demonstrate that MAGIC significantly outperforms existing state-of-the-art methods across eight benchmarks and exhibits strong inference efficiency, indicating its potential for practical applications in task-specific MAS design.
Methodology
MAGIC employs a dense-reward reinforcement learning framework to incrementally construct mixed-granularity agent graphs. At each step, it selects a functional role, determines whether to instantiate it as a single agent or a reusable group, and connects it to existing units. The optimization of the construction policy is based on returns from current-policy trajectories, supplemented by potential-based reward shaping to provide intermediate feedback.
Results
MAGIC was tested across eight benchmarks, where it consistently outperformed existing state-of-the-art methods. The framework demonstrated strong inference efficiency, indicating that it can generate effective collaboration topologies with reduced execution costs.
Implications
The findings suggest that MAGIC can be applied to enhance the design of multi-agent systems in various domains, particularly where task complexity varies significantly. Its ability to adaptively allocate collaboration resources could lead to more efficient and cost-effective solutions in real-world applications.
Limiting-Kernel Q(λ): Bridging Short and Long Horizons
Reinforcement Learning
Theory
Robotics
- Introduction of Limiting-Kernel Q(λ) (LKQL) as a new off-policy value estimator.
- Theoretical analysis proving improved convergence rates for LKQL compared to n-step truncation.
- Demonstration of LKQL's effectiveness on long-horizon tasks in reinforcement learning.
- Integration of LKQL into both on-policy and off-policy actor-critic algorithms.
Read more
Limiting-Kernel Q(λ): Bridging Short and Long Horizons
Summary
This paper addresses the limitations of existing value estimators in reinforcement learning, particularly those relying on n-step truncation, which are constrained to short evaluation horizons. The authors introduce Limiting-Kernel Q(λ) (LKQL), an off-policy value estimator that combines n-step truncation with a long-horizon approximation based on the limiting kernel (LK). LKQL maintains the computational efficiency of n-step estimators while improving the convergence rate of policy evaluation. The paper provides theoretical guarantees for LKQL's convergence to optimal values in finite Markov decision processes (MDPs) and demonstrates its effectiveness on the MuJoCo continuous-control benchmark, showing significant improvements over traditional n-step baselines, especially in long-horizon tasks. The work emphasizes the importance of accurate policy evaluation in enhancing downstream policy optimization performance.
Methodology
The authors develop LKQL by combining n-step truncation with a limiting kernel approximation. They analyze the convergence properties of the truncated operator and introduce a limiting-kernel operator that improves the convergence rate. The LK term is estimated alongside the value function using temporal-difference learning, allowing for a double backup mechanism that integrates seamlessly into existing reinforcement learning frameworks.
Results
The paper demonstrates that LKQL converges almost surely to optimal values under a fixed behavior policy and i.i.d. initial state-action pairs. Empirical results on the MuJoCo continuous-control benchmark show that LKQL outperforms n-step baselines in most settings, particularly excelling in long-horizon tasks.
Implications
The findings suggest that LKQL can significantly enhance the performance of reinforcement learning algorithms, particularly in environments with long-horizon tasks. This could lead to more efficient training and better policy optimization in complex decision-making scenarios.
CS-WCP: Robust Conformal Sets for LLM-Judge Traffic Shifts with Uncertain Group Proportions
Large Language Models
Theory
Optimization
- CS-WCP addresses uncertainties in group proportions during traffic shifts in LLM applications.
- The method constructs simultaneous confidence intervals for source and target group masses.
- CS-WCP significantly reduces under-coverage failures compared to traditional methods.
- The approach is computationally efficient, operating in linear time relative to group and label sizes.
Read more
CS-WCP: Robust Conformal Sets for LLM-Judge Traffic Shifts with Uncertain Group Proportions
Summary
The paper introduces Confidence-set Weighted Conformal Prediction (CS-WCP), a novel method designed to address the challenges posed by traffic shifts in deployment scenarios where the proportions of task or policy groups are uncertain. Traditional weighted conformal prediction (WCP) relies on known density ratios to maintain coverage under covariate shifts, but in practice, these ratios must be estimated, leading to potential coverage failures. CS-WCP constructs simultaneous exact intervals for both source and target group masses, allowing for the union of weighted conformal sets across all compatible ratio vectors. This method effectively separates the uncertainties related to group masses and covariate shifts, providing a more robust prediction set. The authors demonstrate that CS-WCP achieves high coverage rates while maintaining computational efficiency, with a linear endpoint rule that operates in O(G|Y|) time. Empirical evaluations show that CS-WCP significantly reduces point under-coverage failures compared to traditional source conformal prediction, while also providing a safeguard against uncertain mixture weights. The method is validated across various traffic shifts, showcasing its effectiveness in maintaining coverage and providing conservative tail protection.
Methodology
CS-WCP constructs simultaneous Clopper–Pearson intervals for source and target group masses and maps these to a confidence set of compatible density-ratio vectors. It then computes the union of weighted conformal sets over this confidence set, effectively propagating uncertainty rather than compressing it into a single estimate. The method employs a linear endpoint rule for efficient computation.
Results
In experiments involving 336 constructed shared-support traffic shifts, CS-WCP achieved a mean coverage of 0.973 with only 13 point failures, compared to 0.954 and 44 failures for source conformal prediction. In natural cross-task transfers, coverage improved from 0.882 to 0.962, with a mean set size of 1.87, indicating a robust performance in maintaining coverage under varying conditions.
Implications
CS-WCP provides a reliable framework for ensuring coverage in predictive models deployed in dynamic environments, particularly in applications involving large language models. Its ability to handle uncertainties in group proportions makes it valuable for real-world scenarios where traffic characteristics may shift unpredictably.
MSA-CITE: A Co-Adapted LoRA Specialist Ecology for Fixed-Budget Small-Model Inference
NLP
Large Language Models
Efficient ML
- MSA-CITE treats multiple discarded checkpoints as valuable assets for deployment.
- The method utilizes a fixed-budget allocation across differentiated LoRA branches.
- Results indicate improved accuracy over single-branch baselines in specific tasks.
- The approach does not involve joint training, maintaining the independence of branches.
Read more
MSA-CITE: A Co-Adapted LoRA Specialist Ecology for Fixed-Budget Small-Model Inference
Summary
This paper introduces MSA-CITE, a novel approach for deploying compact language models by utilizing multiple discarded post-training checkpoints as composable assets. The authors propose a method that preserves four frozen LoRA branches from a single Qwen3-4B backbone, allowing for a fixed four-generation budget to be allocated across these branches. Instead of sampling multiple generations from one branch, MSA-CITE samples one completion from each branch, processes the results by grouping terminal answers into equivalence classes, and selects a representative answer based on calibrated scoring. The methodology emphasizes the benefits of differentiated adaptation paths without joint training, demonstrating that even without co-training, multiple post-training branches can enhance deployment outcomes. The results show that the four-path portfolio achieves higher accuracy on held-out mathematics items compared to the strongest single-branch baseline, indicating the potential advantages of this approach in small-model inference scenarios.
Methodology
The methodology involves preserving four frozen LoRA branches from a single model backbone and allocating a fixed four-generation budget across these branches. MSA-CITE processes the outputs by grouping answers into equivalence classes and scoring them using calibration-derived source priors. The readout mechanism selects a representative answer based on deterministic tie-breaking rules, without introducing additional generations or reranking steps.
Results
On a set of 200 held-out mathematics items, the four-path portfolio achieved an accuracy of 65.5%, outperforming the strongest single-branch baseline, which had an accuracy of 62.0%. In a 100-item subject-disjoint shift, the portfolio attained 42.0% accuracy compared to 40.0% for the baseline. The improvements were robust under in-distribution conditions, although results against the strongest baseline under shifted conditions were inconclusive.
Implications
The findings suggest that leveraging multiple post-training branches can enhance the performance of compact language models in deployment scenarios. This approach may lead to more efficient use of computational resources and improved accuracy in various applications, particularly in settings where model size and inference budget are constrained.
hyperbolix: Hyperbolic Deep Learning in JAX
Theory
Optimization
- First comprehensive hyperbolic deep learning library for JAX.
- Includes six manifolds with a unified interface for various hyperbolic geometries.
- Implements a wide range of neural network layers and Riemannian optimizers.
- Achieves higher numerical accuracy than existing libraries, particularly on the hyperboloid.
Read more
hyperbolix: Hyperbolic Deep Learning in JAX
Summary
The paper introduces hyperbolix, an open-source library designed for hyperbolic deep learning in JAX, utilizing Flax NNX. It is the first comprehensive library of its kind for JAX, featuring six different manifolds: Euclidean space, the Poincaré ball, the hyperboloid, the κ-stereographic model, mixed-curvature product spaces, and the proper velocity space. The library includes a variety of layer families such as linear layers, convolutions, attention mechanisms, normalization, positional encoding, regression, and vector quantization, enabling the implementation of both traditional hyperbolic neural networks and recent architectures like Hypformer and Lorentzian ResNet. hyperbolix also provides Riemannian optimizers as optax transformations, wrapped distributions, and hyperbolic dimensionality-reduction techniques. The API is designed to be idiomatic to JAX, ensuring stateless manifold operations and efficient batch processing. Notably, hyperbolix addresses numerical stability issues by implementing cancellation-free formulas for two-point operations on the hyperboloid, maintaining accuracy in float32 even at large distances where other implementations fail. The library is available under the MIT license.
Methodology
The authors developed hyperbolix by implementing a unified interface for six hyperbolic manifolds and creating a comprehensive set of neural network layers compatible with JAX's functional programming paradigm. They also focused on enhancing numerical stability through cancellation-free formulas for operations on the hyperboloid.
Results
hyperbolix demonstrates superior numerical precision compared to existing libraries, particularly in float32 operations on the hyperboloid, where it maintains accuracy at distances that typically cause NaN errors in other implementations. The library's API facilitates easy conversion from Euclidean to hyperbolic models, showcasing its practical utility.
Implications
hyperbolix opens up new possibilities for researchers and practitioners working with hierarchical data structures, such as those found in natural language processing and taxonomic classification, by providing robust tools for hyperbolic deep learning in the JAX ecosystem.
Exact Quantile Balancing and Load-Error Injection for Mixture-of-Experts
Optimization
Large Language Models
Efficient ML
- Introduction of Exact Quantile Balancing (EQB) for precise global load balancing in MoE training.
- Development of Load-Error Injection (LEI) to enhance local load balance through direct error injection.
- EQB and LEI together provide complementary control over global and local load balance in MoE models.
- Experimental results show significant improvements in performance and balance over existing methods.
Read more
Exact Quantile Balancing and Load-Error Injection for Mixture-of-Experts
Summary
This paper addresses the challenges of load balancing in Mixture-of-Experts (MoE) training, which is crucial for optimizing the performance of these models. The authors propose two novel methods: Exact Quantile Balancing (EQB) and Load-Error Injection (LEI). EQB computes exact global-batch BF16 quantiles with minimal communication overhead, improving global load balance compared to existing methods that rely on approximate quantiles. LEI enhances local load balance by directly injecting local load errors into router-score gradients, outperforming the GShard loss while maintaining comparable model quality. The proposed methods were tested on 7.5B-parameter MoEs trained on up to 500B tokens, demonstrating significant improvements in both global and local load balancing, which are essential for efficient expert utilization and performance in MoE architectures.
Methodology
The authors implemented EQB using a two-pass BF16 radix selection technique to compute exact global-batch quantiles with minimal communication. LEI was developed to inject local load errors directly into the router-score gradients, allowing for more adaptive load balancing across microbatches. The methods were evaluated on large-scale MoE models, focusing on their impact on load balancing and downstream performance.
Results
The experiments demonstrated that EQB significantly improved global balance and downstream performance compared to naive Quantile Balancing methods. LEI also showed improvements in local balance and outperformed the GShard loss while maintaining similar model quality. Overall, the combination of EQB and LEI led to enhanced efficiency in MoE training.
Implications
The proposed methods can be applied to large-scale MoE architectures, potentially leading to more efficient training processes and better utilization of expert resources. This could have significant implications for various applications in natural language processing and other fields that leverage MoE models.
Spread and Scale: What Determines Whether Test-Time Budget Allocation Pays
Optimization
Theory
Efficient ML
- The effectiveness of budget reallocation depends on the variability of instance difficulty within a workload.
- A budget-aware policy can recover most of the improvement available when instance difficulty information is assumed free.
- The study employs rigorous pre-registered experiments to ensure the validity of results.
- Findings highlight the need for careful measurement to avoid biases in evaluating allocation strategies.
Read more
Spread and Scale: What Determines Whether Test-Time Budget Allocation Pays
Summary
This paper investigates the effectiveness of reallocating a fixed computational budget during test-time in neural combinatorial optimization solvers, particularly in the context of the traveling salesman problem (TSP). Previous studies indicated that reallocating resources towards harder instances could enhance solution quality, but the measurement of this improvement was flawed due to biases in evaluation methods. The author conducts pre-registered experiments to determine the properties of workloads that influence the benefits of budget reallocation. The findings reveal that the variability in difficulty among instances within a workload, rather than the average difficulty, is the key factor that dictates whether reallocation is advantageous. A budget-aware policy that assesses instance difficulty can recover a significant portion of the potential improvement, although it does not achieve the same level of performance as when the information is assumed to be free. The paper emphasizes the importance of empirical validation and provides a framework for future research in adaptive computation allocation methods.
Methodology
The author conducted pre-registered confirmatory experiments across three independently trained solvers and two methods of constructing harder workloads for the traveling salesman problem. The experiments were designed to assess the impact of instance difficulty variability on the benefits of budget reallocation, with careful tracking of amendments and corrections to ensure transparency in the analysis.
Results
The results indicate that workloads with a diverse range of instance difficulties allow for significant gains from budget reallocation, while uniformly easy or hard workloads do not. A budget-aware policy that assesses instance difficulty recovers a majority of the potential improvement, although not as effectively as when the information is assumed to be free. The study also clarifies that the level of instance difficulty does not independently affect the gains from reallocation.
Implications
The findings suggest that researchers and practitioners in combinatorial optimization and adaptive computation should focus on the variability of instance difficulty when designing budget allocation strategies. This could lead to more efficient resource utilization in various applications, enhancing the performance of neural solvers in real-world scenarios.
Probabilistic and Geometry Aware Neural Surrogate of Scrape Off Layer Plasma Simulations
Generative Models
Computer Vision
Theory
- Introduces a probabilistic model to capture multiple steady states in plasma simulations.
- Develops a geometry-aware representation of the SOLPS-ITER mesh for improved data handling.
- Utilizes conditional flow-matching to model the dynamics of plasma states effectively.
- Demonstrates the model's capability to recover distinct plasma behaviors during regime transitions.
Read more
Probabilistic and Geometry Aware Neural Surrogate of Scrape Off Layer Plasma Simulations
Summary
This paper presents a novel approach to creating a fast surrogate model for tokamak boundary-plasma simulations, specifically addressing the challenges posed by the scrape-off layer (SOL) plasma dynamics. Traditional deterministic models struggle to capture the bifurcating behavior of plasma states near the divertor detachment transition, leading to unreliable point estimates. The authors propose a probabilistic model based on conditional flow-matching that allows for the sampling of multiple plausible steady states, thus avoiding the averaging inherent in deterministic models. Additionally, the paper introduces a geometry-aware representation of the SOLPS-ITER mesh, unrolling it into three fixed-size image tensors that preserve cell adjacency. This representation enables the use of convolutional neural networks (CNNs) while maintaining the geometric structure of the data. The resulting surrogate model is efficient and scalable, capable of capturing significant uncertainties in plasma behavior, particularly during regime transitions. The model's effectiveness is validated through synthetic data tests, demonstrating its ability to recover distinct branches of plasma states rather than averaging them.
Methodology
The authors developed a conditional flow-matching model that learns to sample possible steady states from a given input vector. They transformed the curvilinear mesh of the SOLPS-ITER simulator into three adjacency-preserving image tensors, allowing a CNN to operate on the data while retaining its geometric structure. The model was trained on a dataset of 7221 converged steady states, employing a loss function that accounts for the unique characteristics of the plasma dynamics.
Results
The model successfully captured the bifurcating behavior of plasma states, particularly during a gas-puff scan where the predictive distribution split into distinct hot and cold modes. Validation on synthetic data confirmed that the model could recover both branches of a known bifurcation, demonstrating its robustness and accuracy.
Implications
This work has significant implications for the design and operation of future tokamak devices like ITER and DEMO, as it provides a fast and reliable method for simulating plasma behavior under various conditions. The ability to capture uncertainties and multiple outcomes can enhance decision-making in plasma physics research and engineering.
Efficient Cost-Aware LLM Evaluation via Bayesian Bandit Gittins Indices
Large Language Models
Reinforcement Learning
Efficient ML
- GittinsEval models LLM configuration selection as a cost-aware Bayesian bandit problem.
- The framework incorporates heterogeneous evaluation costs and prior information, enhancing decision-making.
- GittinsEval achieves significant performance improvements over traditional Bayesian optimization methods.
- The method often reaches near-zero simple regret using only 1%-2% of the exhaustive evaluation cost.
Read more
Efficient Cost-Aware LLM Evaluation via Bayesian Bandit Gittins Indices
Summary
This paper addresses the challenge of efficiently evaluating large language model (LLM) configurations across various benchmarks without incurring excessive costs. The authors propose GittinsEval, a framework that formulates the configuration selection process as a cost-aware Bayesian bandit problem. By leveraging the Gittins policy, GittinsEval determines which configuration to evaluate next and when to stop, incorporating both performance estimates and evaluation costs. The framework is designed to be computationally efficient, requiring only lightweight online updates after an initial offline precomputation. Additionally, it introduces an anytime recommendation rule that allows for recommendations based on both fully and partially evaluated configurations. The authors demonstrate that GittinsEval consistently outperforms existing methods, particularly in scenarios with large numbers of configurations and benchmarks, achieving near-zero simple regret while utilizing only a small fraction of the exhaustive evaluation cost.
Methodology
The authors develop GittinsEval by adapting the classical Gittins policy to the context of LLM evaluation. This involves modeling the configuration selection as a Bayesian bandit problem, where each configuration is treated as an arm with associated costs and noisy performance scores. The framework employs an anytime recommendation rule that allows for flexible decision-making based on both fully and partially evaluated configurations, utilizing an LCB-style score to manage uncertainty.
Results
GittinsEval was tested across multiple benchmarks, including GSM8K, PIQA, AlpacaEval, and MMLU. The results show that it consistently outperforms configuration-level Bayesian optimization, especially in large-example benchmarks, and significantly surpasses cost-unaware frequentist baselines in large-candidate tasks. The framework achieves near-zero simple regret while requiring only a small fraction of the exhaustive evaluation costs.
Implications
The proposed GittinsEval framework has significant implications for the efficient evaluation of LLMs, enabling researchers and practitioners to identify high-performing configurations with reduced computational resources. This approach can be particularly beneficial in scenarios where evaluation costs are high or where rapid decision-making is required.
VCMM: Variance-Calibrated Momentum for Multimodal Learning
Multimodal
Optimization
- VCMM addresses modality imbalance in multimodal learning by adapting gradient memory to modality-specific dynamics.
- The method estimates minibatch noise and temporal drift online to derive modality-specific momentum coefficients.
- VCMM incorporates bias correction for improved optimization without extra computational cost.
- Experiments show consistent performance improvements across multiple multimodal benchmarks.
Read more
VCMM: Variance-Calibrated Momentum for Multimodal Learning
Summary
The paper introduces Variance-Calibrated Momentum (VCMM), a novel optimization technique designed to address the issue of modality imbalance in multimodal learning. Traditional methods often focus on balancing modality learning through gradient modulation or training strategies, primarily addressing the current optimization step. However, these approaches overlook the accumulated gradient information from previous steps, which is crucial when using momentum-based optimizers. VCMM adapts gradient memory to the dynamics of each modality by estimating minibatch noise and temporal drift online. It employs a Kalman-inspired controller to derive modality-specific momentum coefficients based on the drift-to-noise ratio, allowing for more effective optimization across different modalities. The method also includes bias correction for the time-varying first moment, enhancing the adaptive nature of the gradient memory without requiring additional network passes or explicit learning-rate adjustments. Experiments conducted on four multimodal benchmarks demonstrate that VCMM consistently outperforms competitive baselines with only modest increases in training overhead.
Methodology
VCMM utilizes a Kalman-inspired controller to estimate modality-specific gradient dynamics, including minibatch noise and temporal drift. It calculates a drift-to-noise ratio to determine appropriate momentum coefficients for each modality. The method integrates these coefficients into the optimization process while applying bias correction to the first moment, ensuring effective learning across different modalities.
Results
The experiments conducted on four multimodal benchmarks indicate that VCMM achieves significant performance improvements over existing methods, demonstrating its effectiveness in addressing modality imbalance with minimal additional training overhead.
Implications
The findings suggest that VCMM can enhance multimodal learning applications, particularly in scenarios where different modalities exhibit varying levels of noise and drift. This could lead to better performance in tasks such as image and text classification, audio-visual integration, and other areas requiring multimodal data fusion.
Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following
NLP
Large Language Models
Reinforcement Learning
- CC-OPD introduces a leave-one-out teacher scoring method to derive per-constraint signals for instruction following.
- The method retains the full instruction context during training, improving the clarity of constraint contributions.
- A 1.5B student model trained with CC-OPD surpasses its 7B RL-trained teacher on multi-dimensional constraint benchmarks.
- CC-OPD achieves the highest average performance across multiple instruction-following benchmarks compared to existing methods.
Read more
Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following
Summary
This paper introduces a novel approach called Counterfactual Constraint-Conditioned On-Policy Distillation (CC-OPD) aimed at improving multi-constraint instruction following in large language models (LLMs). Traditional methods struggle with the simultaneous application of multiple constraints, often leading to violations in instruction adherence. Existing techniques either rely on reinforcement learning (RL) with external verifiers or on-policy distillation (OPD) that dilutes the constraint signals when multiple constraints are active. CC-OPD addresses these issues by employing a leave-one-out strategy that evaluates the impact of each constraint individually, allowing for the extraction of per-constraint signals from a frozen teacher model. This method retains the full instruction context during training while generating a more informative reward signal for the student model. The authors demonstrate that a 1.5 billion parameter student model trained with CC-OPD outperforms its 7 billion parameter RL-trained teacher on the MulDimIF benchmark, showcasing the effectiveness of their approach in achieving better adherence to complex instruction sets.
Methodology
The CC-OPD method involves a counterfactual approach where each constraint is ablated from the teacher's conditioning in turn. The resulting per-token probability differentials are computed to create per-constraint signals. These signals are summed, clipped, and integrated into the vanilla OPD reward, allowing the student model to learn effectively from the frozen teacher without external verification.
Results
The results indicate that the 1.5B student model trained with CC-OPD outperformed its 7B RL-trained teacher by 2.8 percentage points on the overall MulDimIF score and by 6.3 percentage points on the hardest Level-4 split. CC-OPD achieved the highest average performance across seven instruction-following benchmarks, demonstrating its effectiveness compared to other training methods.
Implications
The findings suggest that CC-OPD could significantly enhance the performance of LLMs in multi-constraint scenarios, making them more reliable for complex instruction following tasks. This has potential applications in areas requiring precise adherence to multiple guidelines, such as automated content generation, customer service, and interactive AI systems.
The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence
NLP
Large Language Models
Theory
- Identification of the Proximity Trap phenomenon in long-context LLMs.
- Introduction of LYRA, a t-distributed directional matching mechanism for improved context retrieval.
- Empirical evidence showing that masking proximal background context enhances model performance.
- Development of ProxBench for evaluating distant evidence utilization under proximal interference.
Read more
The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence
Summary
This paper addresses a critical challenge in long-context large language models (LLMs), which is the Proximity Trap, where task-irrelevant proximal background context overshadows distant but relevant evidence. The authors propose a novel mechanism called LYRA (Long-context heavY-tailed Relevance Alignment), which utilizes a t-distributed directional matching approach to reshape context retrieval distributions. This mechanism directs more attention towards task-relevant evidence while maintaining the relative positional information. The study reveals that reducing attention to proximal background can enhance long-context understanding, as evidenced by improved performance on various benchmarks. Additionally, the authors introduce ProxBench, a benchmark designed to evaluate distant evidence utilization amidst increasing proximal background interference. The findings suggest that attention allocation in LLMs should consider not only distance but also the competition posed by proximal context.
Methodology
The authors conducted extensive experiments using LYRA to reshape the context retrieval distribution in LLMs. They analyzed the impact of masking proximal background tokens on model performance and introduced ProxBench as a controlled benchmark to assess distant evidence utilization. The LYRA mechanism modifies the conventional query-key scoring function while preserving other components of the attention mechanism.
Results
The experiments showed that masking proximal background context consistently improved accuracy on benchmarks such as LongBench-v2, RULER, and LongBench, particularly in long in-context learning scenarios. The results indicate that the Proximity Trap significantly affects attention allocation, and LYRA effectively mitigates this issue, leading to enhanced model performance across different context lengths and task categories.
Implications
The findings suggest that LLMs can benefit from strategies that reduce the influence of proximal context, potentially leading to more effective utilization of distant evidence. This has implications for the design of future LLM architectures and training methodologies, emphasizing the need to balance attention between proximal and distant contexts.
Learning to Defer with Guidance on Real World Medical Data
Computer Vision
- Introduction of Learning to Defer with Guidance (L2D-G) for medical image interpretation.
- First empirical evaluation of L2D on real-world medical imaging data with human annotations.
- Demonstrated that L2D-G outperforms classic L2D and baseline models.
- Utilized simpler loss functions compared to existing literature.
Read more
Learning to Defer with Guidance on Real World Medical Data
Summary
This paper addresses the challenges of medical image interpretation, which is often time-consuming and prone to errors when relying solely on AI for autonomous predictions. The authors propose a novel framework called Learning to Defer with Guidance (L2D-G), which selectively routes cases between AI predictions and human experts based on input features and performance metrics. This study is the first to validate the L2D approach on real-world medical datasets with human annotations, specifically using the Collab-CXR dataset. The L2D-G framework extends the decision-making process to three options: autonomous prediction, deferring to a human expert, or deferring with AI guidance. The authors compare various rejector architectures and loss functions, demonstrating that L2D-G outperforms traditional L2D setups and baseline models, achieving better performance with simpler loss functions. This work highlights the potential of integrating AI guidance in medical decision-making, particularly in scenarios where timely interpretations are critical.
Methodology
The study implemented a two-stage predictor-rejector framework where the rejector model learns to decide whether to defer to a human expert or make an autonomous prediction. The authors compared multiple rejector architectures, including tree-based models and multi-layer perceptrons, and conducted an ablation study on input features. The L2D-G setup was validated on the Collab-CXR dataset and reproduced on CheXpert and VinDr-CXR datasets.
Results
The results indicated that the L2D-G framework significantly outperformed both the classic two-stage L2D setup and various baseline models, including human-alone and AI-alone approaches. The performance improvements were achieved using simpler loss functions, suggesting that the proposed method is both effective and efficient.
Implications
The findings suggest that integrating AI guidance into the medical decision-making process can enhance the efficiency and safety of image interpretation, particularly in high-volume clinical environments. This approach could lead to faster patient care and better utilization of human expertise, especially in situations where specialists are not immediately available.
Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning
NLP
Large Language Models
Efficient ML
- QAD effectively restores short-form performance but struggles with long-form reasoning under low-bit quantization.
- The proposed OPD method provides on-policy supervision, addressing the training-inference gap caused by quantization-induced deviations.
- The framework shows significant improvements in reasoning performance, with retention rates increasing from 35% to 70% on MATH-500 and from 66% to 91% on HumanEval.
- Combining QAD and OPD allows for a comprehensive solution that maintains broad capabilities in quantized models.
Read more
Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning
Summary
This paper addresses the limitations of quantization-aware distillation (QAD) in restoring reasoning capabilities in low-bit quantized models, particularly in long-form question answering tasks. The authors identify that QAD suffers from quantization-amplified exposure bias, where the model is trained on fixed corpus prefixes, leading to performance degradation in reasoning tasks due to the model's autoregressive nature. To mitigate this issue, the authors propose an on-policy distillation (OPD) method that allows the student model to generate outputs through the quantized path used during deployment, receiving feedback from a frozen full-precision teacher on its own generated prefixes. This two-stage recovery framework first utilizes QAD for stable low-bit initialization and then applies OPD to enhance reasoning capabilities by providing targeted supervision on the model's actual trajectories. The proposed method significantly improves performance retention on reasoning tasks while preserving short-form performance, demonstrating the effectiveness of combining QAD and OPD for low-bit quantization.
Methodology
The authors introduce a two-stage framework that first employs quantization-aware distillation (QAD) for stable low-bit initialization, followed by on-policy distillation (OPD) where the student model generates outputs through the quantized path and receives feedback from a frozen full-precision teacher. This method incorporates both token-level guidance and task-verifier rewards to enhance reasoning capabilities.
Results
The implementation of OPD alongside QAD leads to a substantial increase in average BF16 performance retention, achieving 70% retention on MATH-500 and 91% on HumanEval, while also maintaining short-form performance. The results indicate that OPD contributes significantly to recovery, especially at lower bit widths.
Implications
This research has implications for deploying large language models on resource-constrained hardware, as it provides a method to maintain reasoning capabilities even at extreme low-bit quantization levels. The findings could enhance the efficiency of AI applications in various domains requiring real-time processing and lower memory usage.
Reliable Fusion of Conflicting Experts
Large Language Models
NLP
Optimization
- Introduction of a probabilistic-circuit-based framework for expert opinion fusion.
- Dynamic aggregation of expert responses based on context-specific credibility estimates.
- Empirical validation shows significant performance improvement over static ensemble methods.
- Framework is agnostic to the underlying expert models and does not require retraining.
Read more
Reliable Fusion of Conflicting Experts
Summary
This paper addresses the challenge of aggregating opinions from multiple black-box experts in environments where expert reliability varies and conflicts arise. Traditional static aggregation methods, such as majority voting, fail to account for the dynamic nature of expert reliability, leading to unreliable outcomes. The authors propose a novel probabilistic-circuit-based fusion framework that dynamically combines expert responses using context-specific credibility estimates. This framework is designed to work without requiring access to the internal workings of the experts or any retraining. The authors validate their approach through experiments on multiple-choice question answering tasks, utilizing various large language models (LLMs) as experts. The results demonstrate that the proposed method consistently outperforms static ensemble methods, improving predictive performance and yielding more reliable decisions in the presence of conflicting expert opinions. This highlights the importance of context-aware credibility modeling in enhancing the robustness of multi-expert fusion systems.
Methodology
The authors developed a lightweight framework that utilizes probabilistic circuits to dynamically aggregate conflicting opinions from black-box experts. The framework computes context-specific credibility scores for each expert based on the input query, allowing for instance-level reliability assessments and improved decision-making.
Results
The proposed framework was tested on the MMLU question answering benchmark, where it consistently outperformed static ensemble methods. The results indicated that the dynamic aggregation approach effectively resolved expert disagreements and improved overall predictive accuracy.
Implications
This research has significant implications for AI systems that rely on multiple expert opinions, particularly in dynamic environments such as healthcare and decision support systems. The ability to dynamically assess expert reliability can lead to more informed and reliable decision-making processes.
A Shared Encoder Is Not a Shared Task: Conditional Comparison for Deep Expert Pools
Theory
- Critiques the reliance on shared encoders for assessing task compatibility in continual learning.
- Introduces conditional comparison as a more reliable method for evaluating task mechanism changes.
- Demonstrates that existing methods can inflate compatibility scores under input shifts.
- Proposes a two-axis gating mechanism that enhances decision-making in mixture-of-heads systems.
Read more
A Shared Encoder Is Not a Shared Task: Conditional Comparison for Deep Expert Pools
Summary
This paper addresses the limitations of using shared encoders in deep continual learning systems, particularly in the context of task compatibility assessment. The author critiques existing methods that rely on head exchange and representation-level comparisons, arguing that these approaches can lead to misleading conclusions about task compatibility. The study introduces a new method called conditional comparison, which utilizes the conditional Jensen–Shannon discrepancy to differentiate between changes in task mechanisms and mere input shifts. Through empirical experiments, the paper demonstrates that traditional methods can inflate compatibility scores under certain conditions, while the proposed method maintains accuracy across various transformations. The findings indicate that a two-axis gating mechanism, which incorporates both functional and covariate axes, significantly improves decision quality in mixture-of-heads systems. The results show that the new method outperforms existing approaches in distinguishing between semantic novelty and photometric shifts, achieving high accuracy rates in various scenarios.
Methodology
The study employs a mixture-of-heads system with a frozen encoder and evaluates various task-comparison scores, including the deep exchange score and Mahalanobis-style representation novelty. The proposed conditional comparison method uses two small MLP discriminators to compute the conditional Jensen–Shannon discrepancy in the embedding space, allowing for a more nuanced understanding of task compatibility.
Results
The paper reports that the deep exchange score can inflate significantly under input shifts, while the conditional comparison method maintains stability across transformations. In experiments, the two-axis gating mechanism achieved a false-spawn rate of 0.044 with an average of 2.7 heads, compared to 11.3 heads with the deep exchange rule. The method also successfully distinguished between semantic novelty and photometric shifts with high AUROC scores (0.99 for CIFAR-10).
Implications
The findings suggest that the proposed conditional comparison method can improve the robustness of continual learning systems by providing more accurate assessments of task compatibility. This has implications for the design of adaptive learning systems that need to manage multiple tasks effectively, particularly in dynamic environments.
A JEPA Recipe for Tabular Foundation Models
Theory
Optimization
Efficient ML
- Introduces a recipe for training JEPA without latent term collapse.
- Proposes an open-horizon protocol for training that stops at convergence plateaus.
- Demonstrates the performance of JEPA against a value-only arm across 147 datasets.
- Finds that JEPA requires more training steps and time to reach convergence compared to the value-only arm.
Read more
A JEPA Recipe for Tabular Foundation Models
Summary
This paper presents a novel approach to training joint-embedding predictive architectures (JEPA) for tabular foundation models, addressing the issue of latent term collapse observed in previous attempts. The authors propose a recipe that allows the latent term to converge alongside the value objective by modifying the way the value head interacts with the encoder and introducing an exponential moving average (EMA) difference as the target. They also introduce an open-horizon training protocol that halts training at a plateau, contrasting with earlier methods that relied on fixed budgets. The study evaluates the JEPA arm against a value-only arm across 147 real datasets, revealing that while the JEPA arm does not outperform the value-only arm, it successfully trains the latent term without collapse. This work contributes to understanding the dynamics of self-supervised learning in tabular data contexts and provides a framework for future research in this area.
Methodology
The authors utilize a shared transformer architecture for both the JEPA arm and the value-only arm, with modifications in the value head's input and the target prediction method. The JEPA arm incorporates an EMA difference as the target for the latent term, while both arms are trained until they reach a plateau, allowing for a fair comparison of their performance across various datasets.
Results
The JEPA arm achieved a performance ratio of 32:70 wins to losses in classification tasks and 8:24 in regression tasks against the value-only arm across 147 datasets. The JEPA arm required 1.42 times as many training steps and 1.66 times the wall-clock time to reach its plateau compared to the value-only arm.
Implications
This research has implications for the development of more robust tabular foundation models, particularly in self-supervised learning contexts. The findings suggest that careful architectural choices and training protocols can mitigate issues of latent collapse, potentially leading to improved performance in real-world tabular data applications.
Geometry-Aware Hyperbolic Residual Quantization
Generative Models
Audio & Speech
Computer Vision
- Introduces a geometry-aware hyperbolic residual quantization method to address inconsistencies in hyperbolic space.
- Restores consistent residual aggregation and provides stable gradient routing using novel techniques.
- Demonstrates improved stability and structural organization of hyperbolic residual codes compared to naive methods.
- Evaluates the method across multiple domains, highlighting its effectiveness in hierarchical representation learning.
Read more
Geometry-Aware Hyperbolic Residual Quantization
Summary
This paper presents a novel approach to residual vector quantization by leveraging hyperbolic geometry, which is particularly suited for hierarchical data representations. The authors identify geometric inconsistencies that arise when traditional residual quantization methods are applied in hyperbolic space, specifically during the forward and backward passes of the quantization process. To address these issues, they propose a geometry-aware hyperbolic residual quantization method (GHRQ-VAE) that stabilizes training and enhances the representation of latent hierarchical structures. The method includes Hyperbolic Residual Aggregation for consistent residual aggregation and a discounted Hyperbolic Straight-Through Estimator for stable gradient routing. The authors evaluate their approach across various tasks, including hierarchical prediction, recommendation, image tokenization, and neural audio coding, demonstrating that GHRQ-VAE outperforms naive hyperbolic baselines in terms of stability and structural organization, although Euclidean methods still excel in pure compression tasks.
Methodology
The authors develop a geometry-aware hyperbolic residual quantization framework that includes Hyperbolic Residual Aggregation for the forward pass and a discounted Hyperbolic Straight-Through Estimator for the backward pass. This approach ensures consistent operations in hyperbolic space, addressing the non-associative nature of hyperbolic addition and improving gradient flow during training.
Results
The proposed GHRQ-VAE method shows significant improvements in training stability and the structural organization of hyperbolic residual codes across various tasks. While it does not outperform Euclidean methods in pure compression scenarios, it excels in tasks requiring hierarchical organization of latent spaces.
Implications
This work suggests that hyperbolic geometry can be effectively utilized in machine learning models that require hierarchical representations, potentially benefiting applications in generative modeling, recommendation systems, and other domains where data exhibits inherent hierarchical structures.
Discover, Falsify, Revise: Auditing Input-Use Claims from Source Code to Predictive Contribution in Agent-Discovered Cell Models
Theory
Interpretability
- CELLAUDIT framework audits input-use claims in AI-generated cell models.
- Addresses the prediction–claim gap where models may not utilize perturbation information effectively.
- Demonstrates that many models change predictions under input replacement, but few show significant predictive contributions.
- Falsification-guided revisions can enhance predictive performance and input contributions.
Read more
Discover, Falsify, Revise: Auditing Input-Use Claims from Source Code to Predictive Contribution in Agent-Discovered Cell Models
Summary
The paper introduces CELLAUDIT, a novel framework designed to audit input-use claims in AI-generated cell models that predict cellular responses to interventions. The authors highlight the critical issue of the prediction–claim gap, where models may perform well without effectively utilizing the specified perturbation information. CELLAUDIT addresses this gap by evaluating three key questions: whether the input can be processed by the model, whether predictions depend on it, and whether this dependence enhances prediction accuracy. The framework was tested on a benchmark involving morphology and transcriptomics perturbations (BBBC047). Results showed that an agent-selected predictor achieved a mean held-out Global Pearson correlation coefficient (PCC) of 0.3153 but remained invariant to compound replacement, indicating a lack of effective input use. Further audits revealed that while many generated models changed predictions under input replacement, only a subset demonstrated meaningful predictive contributions. The authors also found that revisions guided by falsification improved predictive performance and input contributions. Overall, CELLAUDIT adds a necessary layer of verification in agentic model discovery, promoting a shift from merely generating and scoring models to a more rigorous process of discovery, falsification, and revision.
Methodology
The methodology involves a three-question audit framework that assesses input processing, prediction dependence, and predictive contribution. It combines source code inspection with matched input replacement tests on frozen model checkpoints, allowing for the evaluation of both model dependence and target-relevant contributions.
Results
The agent-selected predictor achieved a mean held-out Global PCC of 0.3153, but was invariant to compound replacement. In a broader audit of 48 candidates, 47 changed predictions under input replacement, yet only 20 showed significant predictive contributions. Falsification-guided revisions improved mean contributions while retaining predictive gains over control models.
Implications
The findings suggest that rigorous auditing of input-use claims is essential for ensuring the reliability of AI-generated models in biological contexts. This could lead to more accurate predictions in cellular response modeling, ultimately aiding in intervention design and biological research.
Protocol before progress: leakage-aware evaluation of AIS trajectory prediction
Graph Learning
Time Series
Theory
- The proposed leakage-aware evaluation protocol reveals significant error inflation in AIS trajectory predictions.
- Oracle decoding can reduce prediction error by a factor of 2.1 to 3.2 compared to greedy decoding.
- Vessel-sharing splits inflate error rates significantly for larger models, while smaller models are less affected.
- Region-disjoint splits can drastically increase prediction error due to unseen training contexts.
Read more
Protocol before progress: leakage-aware evaluation of AIS trajectory prediction
Summary
This paper addresses the evaluation of vessel trajectory prediction using Automatic Identification System (AIS) data, emphasizing the importance of a leakage-aware evaluation protocol. The authors argue that reported improvements in prediction accuracy are often inflated due to data leakage from improper dataset splits. They propose a new evaluation protocol that includes vessel-, time-, and region-disjoint splits, applied to two distinct AIS traffic corpora: Danish national traffic and US Gulf coast traffic. The study audits several existing models, including TrAISformer and GATransformer, revealing significant discrepancies in reported error rates due to the evaluation method. Key findings include that the best-of-16 oracle decoder substantially reduces error rates, while vessel-sharing splits inflate errors for larger models. The paper also highlights that architectural features like graph attention do not provide measurable benefits, whereas engineered features do improve performance. The authors conclude that their leakage-aware protocol is crucial for accurate evaluation in maritime deep learning and provide their dataset splits and code for reproducibility.
Methodology
The authors developed a leakage-aware evaluation protocol with three types of dataset splits (vessel-, time-, and region-disjoint) and applied it to two AIS traffic corpora. They audited existing models, including TrAISformer and GATransformer, to measure the impact of these splits on prediction accuracy.
Results
The study found that the best-of-16 oracle decoder significantly lowers error rates, with reductions of 2.1 to 3.2 times compared to greedy decoding. Vessel-sharing splits resulted in a 23-28% error reduction for larger models, while region-disjoint splits increased error rates dramatically. The architectural feature of graph attention provided no measurable benefit, whereas engineered features improved accuracy by 12-22%. The effects of time-disjoint splits varied across datasets.
Implications
The findings underscore the necessity of rigorous evaluation protocols in machine learning, particularly in maritime applications. The proposed leakage-aware protocol can enhance the reliability of AIS trajectory predictions and inform future research and development in this field.
From Reasoning Strings to Partial Orders: Verifier-Certified Rule Transport through Quotient Policy Optimization
Reinforcement Learning
Optimization
Theory
- Introduces Verifier-Certified Rule Transport (VCRT) for reinforcement learning.
- Establishes commutation certificates and anti-diamonds to manage operation order.
- Implements a quotient policy objective to optimize learning based on verified trajectories.
- Demonstrates significant performance improvements in reasoning tasks across multiple environments.
Read more
From Reasoning Strings to Partial Orders: Verifier-Certified Rule Transport through Quotient Policy Optimization
Summary
This paper introduces Verifier-Certified Rule Transport (VCRT), a novel approach in reinforcement learning that addresses the challenges of execution order in computations with independent subgoals. Traditional reinforcement learning with verifiable rewards often misinterprets serialization choices as logical dependencies, leading to incorrect learning. VCRT enhances the learning process by replaying operation pairs with native verifiers to establish commutation certificates for pairs that can be executed in either order without affecting the outcome. The method also identifies anti-diamonds for pairs that violate prerequisites, ensuring that genuine dependencies are preserved. The authors propose a quotient policy objective that leverages these certificates to optimize policy learning while constraining post-swap consistency and policy drift. The evaluation of VCRT is conducted using a leave-one-environment-out approach across three reasoning environments: ProofWriter, CLRS, and Lean, demonstrating its effectiveness in transferring learned policies to unseen environments. The results show a significant improvement in performance, with VCRT achieving a macro pass rate of 77.60%, outperforming the strongest baseline by 13.06 points.
Methodology
The methodology involves replaying pairs of operations with native verifiers to obtain certificates for commutation and anti-diamonds. The VCRT framework uses these certificates to form a quotient policy objective that optimizes the total policy mass of certified trajectories while applying constraints to ensure consistency and prevent policy drift. The evaluation is conducted in a controlled manner, freezing training configurations before testing on held-out environments.
Results
VCRT achieved a macro pass rate of 77.60%, significantly higher than the strongest matched baseline of 64.53%, resulting in a paired gain of 13.06 points. The majority of this improvement was attributed to the Lean environment, which accounted for a gain of 33.49 points, while ProofWriter and CLRS showed an average improvement of 2.85 points. Mechanism tests indicated that anti-diamond supervision consistently outperformed other methods.
Implications
The findings suggest that VCRT can enhance the efficiency and accuracy of reinforcement learning in environments where execution order matters, potentially benefiting applications in automated reasoning, algorithmic problem-solving, and other domains that require structured decision-making.
Shared Global KV with Layer-Specific Local History
NLP
Large Language Models
Efficient ML
- Shared global KV can reduce storage but limits representation diversity across layers.
- Incorporating layer-specific local history improves model performance, reducing perplexity.
- Independent projections with shared local inputs maintain the benefits of historical content.
- The proposed architecture allows for efficient cache construction and resource management.
Read more
Shared Global KV with Layer-Specific Local History
Summary
This paper investigates the design of decoder-only Transformer language models that utilize shared global keys and values (KV) alongside layer-specific local history. The authors highlight the trade-off between storage efficiency and representation diversity when sharing KV across layers. They propose a method to retain historical content while allowing for independent local inputs at each layer. Through an eight-seed factorial study, the paper demonstrates that incorporating local history results in approximately 1.4% lower held-out test perplexity compared to using a current-token local branch. The study also explores the effects of shared local inputs and independent projections, revealing that this configuration maintains the benefits of historical content while reducing construction dependencies. The findings suggest that the proposed architecture can improve performance on tasks while managing resource costs effectively. The paper concludes with a discussion on the implications of these design choices for future Transformer architectures.
Methodology
The authors conducted an eight-seed factorial study to evaluate the impact of local history and shared KV on model performance. They varied the content and input source of local memory while maintaining a shared global KV. The study included comparisons of different configurations, such as adjacent-layer sharing and independent projections, to assess their effects on perplexity and resource costs.
Results
The study found that models utilizing local history achieved about 1.4% lower held-out test perplexity compared to those using a current-token local branch. Additionally, configurations that allowed adjacent layers to share local inputs while retaining independent projections showed persistent benefits in performance. The proposed architecture also demonstrated improved same-source likelihood and efficient cache construction.
Implications
The findings suggest that optimizing the retention of historical content in Transformer models can lead to better performance in language generation tasks. This approach may inform future designs of language models, particularly in balancing storage efficiency and representation diversity.
A Lightweight Plastic-Memory Framework for Graph Few-Shot Class-Incremental Learning
Graph Learning
- Introduces a lightweight plastic-memory framework for GFS-CIL.
- Employs an evolving micro-clustering structure for efficient memory management.
- Integrates meta-learning to enhance adaptability while preserving stability.
- Demonstrates superior performance on benchmark datasets compared to existing methods.
Read more
A Lightweight Plastic-Memory Framework for Graph Few-Shot Class-Incremental Learning
Summary
This paper addresses the challenges of Graph Few-Shot Class-Incremental Learning (GFS-CIL), where models must learn from limited labeled data while retaining previously acquired knowledge. The authors propose a novel Lightweight Plastic-Memory Framework that incorporates a dynamic plastic-memory module designed to evolve over time. This module utilizes an innovative micro-clustering structure to manage class prototypes and their interactions efficiently. The framework integrates a memory-driven meta-learning approach, balancing adaptability to new tasks with stability for older tasks. Extensive experiments on benchmark datasets demonstrate that the proposed framework outperforms existing methods in terms of efficiency and effectiveness, particularly in scenarios with scarce labeled data. The results highlight the framework's ability to mitigate catastrophic forgetting while maintaining a lightweight design, making it suitable for real-world applications involving dynamic graph data.
Methodology
The proposed framework consists of a plastic-memory module that evolves dynamically, utilizing a hierarchical micro-clustering representation to group nodes into class prototypes and sub-prototypes. The framework employs a dual-loop meta-learning strategy, where the inner loop focuses on adapting to new tasks and the outer loop ensures stability for previously learned tasks. This approach optimizes training efficiency by replacing raw data with prototypes and local adjustments instead of global structural updates.
Results
The framework was evaluated on four benchmark datasets, showing significant improvements in balancing stability for old knowledge and adaptability to new knowledge. The experiments confirmed that the proposed method effectively mitigates catastrophic forgetting while being computationally efficient, outperforming traditional methods in scenarios with limited labeled data.
Implications
The Lightweight Plastic-Memory Framework has potential applications in various fields that utilize dynamic graph data, such as social networks, recommendation systems, and biological networks. Its ability to learn incrementally from few labeled examples while retaining knowledge makes it suitable for real-world scenarios where data is often scarce and evolving.
Memory Attention
NLP
Large Language Models
Efficient ML
- Introduction of Memory Attention (MA) that combines token memory with contextual keys to replace traditional value projections.
- Development of MA-Offload for efficient memory management by utilizing CPU storage and prefetching techniques.
- Evaluation shows improved language modeling perplexity and downstream task accuracy under matched training conditions.
- Analysis of trade-offs between parameter capacity, value construction, and GPU memory storage.
Read more
Memory Attention
Summary
The paper introduces Memory Attention (MA), a novel approach to enhance language models by integrating token-indexed memory with contextual keys, thereby replacing the traditional value projection in attention mechanisms. MA aims to improve the efficiency and quality of language models under limited computational resources. By utilizing layer-specific token memory, MA allows for the construction of values through a combination of this memory and contextual keys, preserving the context dependence while reducing the computational burden during inference. The proposed MA-Offload method enables the storage of memory tables in CPU memory, facilitating prefetching and reducing GPU memory requirements. The paper evaluates MA across various attention configurations, demonstrating improved language modeling performance and downstream task accuracy, while also addressing the trade-offs between parameter capacity and computational efficiency. The results indicate that MA can effectively enhance model performance while optimizing resource utilization.
Methodology
The methodology involves the design of Memory Attention (MA) which integrates token-specific representations from memory with contextual keys. This is achieved by replacing the dense value projection with an embedding lookup and addition operation. The MA-Offload technique is developed to store memory tables in CPU memory, allowing for efficient retrieval and reducing GPU memory usage. The paper also includes experiments to evaluate the performance of MA in language modeling and downstream tasks.
Results
The experiments indicate that Memory Attention significantly improves language modeling perplexity and downstream task accuracy compared to standard attention mechanisms. The MA-Offload method successfully reduces GPU memory requirements while accommodating larger parameterizations. Inference measurements show that the proposed methods maintain efficiency despite the increased parameter storage.
Implications
The findings suggest that Memory Attention can lead to more efficient language models that leverage memory effectively, potentially allowing for larger models to be deployed in resource-constrained environments. This approach could influence future designs of attention mechanisms in NLP applications, optimizing both performance and resource utilization.
Tail-Aware Geometry Learning for Conformal Ellipsoids
Theory
Optimization
- Introduces a tail-aware geometry learning framework for conformal ellipsoids.
- Decouples tail sensitivity from the final coverage guarantee in CP.
- Utilizes Conditional Value-at-Risk (CVaR) for learning geometry, prioritizing high-residual samples.
- The proposed method is convex and allows for independent tuning of tail sensitivity and coverage levels.
Read more
Tail-Aware Geometry Learning for Conformal Ellipsoids
Summary
This paper addresses the challenges in multivariate conformal prediction (CP), particularly focusing on the geometry of prediction sets and the impact of tail residuals on their efficiency. Traditional methods for learning geometry in CP often rely on quantile thresholds that overlook the severity of tail residuals, which can lead to suboptimal prediction regions. The author proposes a novel tail-aware geometry learning framework for conformal ellipsoids that separates the sensitivity to tail residuals from the final coverage guarantees. This is achieved through a two-split design where a metric matrix is learned via volume minimization under a Conditional Value-at-Risk (CVaR) constraint on an estimation split, followed by standard conformal calibration on a separate calibration split. The proposed method is convex and emphasizes high-residual samples, allowing for a more effective geometry learning process. The paper also provides a theoretical characterization of the trade-off between ellipsoidal volume and tail severity, and extensive experiments validate the effectiveness of the proposed approach in enhancing the performance of multivariate CP.
Methodology
The methodology involves a two-split design where a metric matrix is estimated using volume minimization under a CVaR constraint on an estimation dataset, followed by conformal calibration on a separate calibration dataset. This approach allows for the independent tuning of tail sensitivity and coverage levels, resulting in a convex optimization problem that prioritizes high-residual samples.
Results
The experimental results indicate that the proposed tail-aware geometry learning framework significantly improves the efficiency of multivariate conformal prediction sets compared to traditional methods. The results demonstrate better handling of tail residuals, leading to more accurate and reliable prediction regions.
Implications
The findings suggest that the proposed method can enhance uncertainty quantification in various applications involving multivariate data, such as anomaly detection and time series prediction, by providing more accurate prediction sets that account for tail behavior in residuals.
EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
Time Series
- EMGBlend addresses the challenges of heterogeneous EMG datasets through a self-supervised learning framework.
- The framework includes mechanisms for shared channel patches, geometry-aware attention, and source-exposure correction.
- Pretraining on diverse EMG sources significantly improves performance in gesture recognition and force decoding tasks.
- Ablation studies confirm the importance of each mechanism in enhancing transfer learning capabilities.
Read more
EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
Summary
The paper introduces EMGBlend, a self-supervised framework designed to address the challenges posed by heterogeneous surface electromyography (EMG) datasets, which vary in electrode layout, channel count, frequency support, and size. The authors argue that simply mixing these datasets for pretraining can lead to misalignment of channel semantics and overexposure of certain datasets. EMGBlend incorporates three key mechanisms: a shared channel patch interface with geometry-aware attention, spectral targets restricted to each recording's supported frequency band, and a source-exposure correction to balance the influence of different datasets. The framework was evaluated using a 109M-parameter model pretrained on 11 public EMG sources, demonstrating superior performance in gesture recognition, continuous-force regression, and contact classification compared to matched random initialization and waveform reconstruction controls. The results indicate that multi-source pretraining enhances gesture recognition and remains competitive for force decoding, with ablation studies confirming the contributions of geometry, band-aware targets, and source balancing. Overall, EMGBlend showcases a novel approach to effectively combine heterogeneous EMG datasets through explicit mechanism design rather than simple concatenation.
Methodology
The methodology involves pooling multiple EMG sources while employing a shared channel patch interface, geometry-aware attention, and spectral targets restricted to each recording's frequency band. The model is pretrained using a self-supervised approach, where task labels are only utilized downstream. The training process includes filtering, resampling, and normalizing recordings, followed by the application of a shared Conv1D encoder and attention mechanisms to learn representations that accommodate varying sensor layouts and channel counts.
Results
The pretrained model consistently outperformed matched random initializations and waveform reconstruction controls across various tasks, including gesture recognition, continuous-force regression, and contact classification. Multi-source pretraining led to improved performance in gesture recognition and maintained competitiveness in force decoding. Ablation studies indicated that the geometry-aware mechanisms, band-aware targets, and source balancing each contributed positively to the model's transfer capabilities.
Implications
The findings suggest that EMGBlend can facilitate the development of more robust and generalizable EMG-based applications in gesture recognition and force estimation. This framework could be applied to enhance the performance of assistive technologies, rehabilitation devices, and human-computer interaction systems that rely on EMG signals.
Topological Signal Processing With Unoriented Operators
Graph Learning
Theory
Optimization
- Introduction of Unoriented Topological Signal Processing (UTSP) framework.
- Establishment of spectral properties for unoriented incidence and Laplacian matrices.
- Development of interaction-order decomposition for analyzing higher-order signals.
- Creation of order-aware regularizers that improve signal reconstruction.
Read more
Topological Signal Processing With Unoriented Operators
Summary
This paper introduces an Unoriented Topological Signal Processing (UTSP) framework that addresses the limitations of traditional Topological Signal Processing (TSP) when applied to higher-order signals that lack orientation. The authors argue that many real-world signals, such as those found in co-authorship networks or biochemical interactions, do not possess a natural orientation, making the application of oriented operators problematic. The UTSP framework replaces oriented boundary operators with unoriented incidence matrices, allowing for the analysis of higher-order signals without the complications introduced by orientation. The paper presents three main contributions: (1) it establishes that unoriented incidence and Laplacian matrices exhibit spectral properties similar to those of graphs, enabling the application of graph signal processing techniques to unoriented simplicial signals; (2) it introduces an interaction-order decomposition that quantifies how much of a higher-order signal can be explained by lower-order signals, compensating for the loss of the Hodge decomposition; and (3) it develops regularizers for signal reconstruction that penalize each interaction order separately. Experimental results demonstrate that these order-aware regularizers outperform traditional oriented methods, particularly when signal energy is unevenly distributed across different orders.
Methodology
The authors replace oriented boundary operators with unoriented incidence matrices and analyze the resulting spectral properties. They introduce an interaction-order decomposition to quantify the contribution of lower-order signals to higher-order signals. Regularizers for signal reconstruction are derived based on this decomposition, allowing for separate penalization of different interaction orders.
Results
The study finds that the proposed order-aware regularizers significantly outperform traditional oriented methods in signal reconstruction tasks, especially when the energy of the signals is unevenly distributed across different interaction orders. The spectral properties of the unoriented matrices align with those of graphs, facilitating the use of graph signal processing techniques.
Implications
The UTSP framework has the potential to enhance the analysis and processing of complex systems characterized by higher-order relationships, such as social networks, biological systems, and collaborative environments. It opens new avenues for research in topological signal processing and its applications in various fields.
Transferable Evidence Reconstruction for Longitudinal Glucose Representations
Time Series
- Introduces Transferable Evidence Reconstruction (TER) for self-supervised learning in glucose monitoring.
- Focuses on reconstructing structured evidence from unlabeled data rather than relying on downstream features.
- Achieves significant improvements in performance metrics over existing methods on multiple tasks.
- Demonstrates the effectiveness of a CGM encoder that organizes historical glucose data over time.
Read more
Transferable Evidence Reconstruction for Longitudinal Glucose Representations
Summary
This paper introduces Transferable Evidence Reconstruction (TER), a novel self-supervised learning approach designed to enhance the representation of longitudinal glucose data from continuous glucose monitoring (CGM). The authors argue that traditional methods often overlook the significance of rare events and recurring patterns in physiological data. TER focuses on reconstructing evidence from unlabeled recordings, allowing a low-capacity reader to learn from one group of recordings and apply that knowledge to another without refitting. This method emphasizes the importance of structured signal evidence, which guides self-supervision without being used as a downstream feature. The proposed CGM encoder effectively organizes glucose data over time, binding glucose levels to recorded times and enabling the model to leverage up to seven days of historical data. The results demonstrate that TER significantly outperforms existing methods on a 14-task leaderboard, achieving new best metrics on 12 out of 14 tasks, thus showcasing its potential for improving glucose monitoring and management.
Methodology
The methodology involves a self-supervised learning framework where a low-capacity reader is trained on one group of recordings to reconstruct evidence, which is then applied unchanged to another group. The model differentiates through cross-group testing to learn transferable evidence-decoding rules. The CGM encoder integrates observation-aware daily content with a clock-aware memory to bind glucose levels to time, facilitating the organization of historical data.
Results
TER achieved overall PR-AUC, ROC-AUC, and Macro-F1 scores of 64.31, 71.13, and 62.70, respectively, surpassing the strongest prior methods by 5.51, 4.43, and 2.80 percentage points. It set new best metrics on 12 out of 14 tasks in the evaluation, demonstrating substantial improvements in performance.
Implications
The findings suggest that TER can significantly enhance the predictive capabilities of glucose monitoring systems, potentially leading to better management of diabetes and other metabolic disorders. The approach may also be applicable to other domains where longitudinal data is prevalent, emphasizing the importance of structured evidence in self-supervised learning.
Repurposing Pre-trained LLMs as High Fidelity Continuous Text Autoencoders
NLP
Large Language Models
Generative Models
- LLMAE provides a novel architecture for transforming pre-trained LLMs into effective text autoencoders.
- The framework achieves near-perfect reconstruction of text sequences, significantly improving upon previous methods.
- LLMAE enables the training of a latent text diffusion model for high-quality image captioning.
- The approach mitigates issues of catastrophic forgetting associated with traditional fine-tuning.
Read more
Repurposing Pre-trained LLMs as High Fidelity Continuous Text Autoencoders
Summary
This paper introduces LLMAE, a novel framework for repurposing pre-trained decoder-only language models (LLMs) into high-fidelity continuous text autoencoders. The authors argue that while autoregressive models excel in fluency, they struggle with global structure and downstream adaptation. LLMAE addresses these limitations by creating a fixed-length continuous latent representation from the internal activations of a lightweight 270M parameter Gemma 3 model. The method employs structured attention masks, LoRA adaptation, and KL regularization to facilitate effective text reconstruction. The authors demonstrate that LLMAE achieves near-perfect reconstruction for text sequences of up to 1024 tokens, significantly outperforming existing text autoencoders. Additionally, they showcase the utility of LLMAE by training a latent text diffusion model for detailed image captioning, which generates fluent captions that outperform larger autoregressive models. This work highlights the potential of LLMAE as a substrate for multimodal integration and complex downstream tasks, offering an alternative to traditional fine-tuning methods that often lead to catastrophic forgetting.
Methodology
The authors developed LLMAE by utilizing structured attention masks to create a fixed-length latent bottleneck from the internal activations of a pre-trained LLM. They incorporated LoRA adaptation and KL regularization to optimize the autoencoding process, allowing the model to reconstruct text sequences effectively. The framework was instantiated using the 270M parameter Gemma 3 model, and a lightweight diffusion model was trained on the learned latents for image captioning.
Results
LLMAE demonstrated near-perfect reconstruction capabilities for text sequences up to 1024 tokens, outperforming existing text autoencoders. The latent text diffusion model trained on LLMAE latents produced fluent, high-quality captions, surpassing larger autoregressive baselines and narrowing the performance gap with state-of-the-art vision-language models.
Implications
The LLMAE framework offers a promising alternative for text representation and generation, facilitating multimodal tasks and improving the efficiency of downstream adaptations. Its ability to maintain fluency while providing a continuous latent space representation could lead to advancements in various applications, including text generation, image captioning, and beyond.
Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning
NLP
Large Language Models
Theory
- Introduces a controlled rule-shuffling task to analyze mathematical reasoning.
- Defines permutation signal-to-noise ratio (SNR) to measure representation distinctiveness.
- Finds a positive correlation between accuracy and representation distinctiveness across models.
- Distinguishes between answer invariance and representation invariance in mathematical reasoning.
Read more
Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning
Summary
This paper explores the relationship between answer invariance and representation invariance in mathematical reasoning tasks. Specifically, it investigates whether models that accurately solve mathematical problems presented in different orders also maintain similar internal representations. The study employs synthetic multi-step function-composition problems, where the same correct answer is derived from various orderings of mathematical rules. The author introduces a novel metric called permutation signal-to-noise ratio (SNR) to quantify how distinctly different rule orderings are represented within the models. The analysis is conducted across 16 language models with parameter sizes ranging from 1B to 8B. The findings reveal a consistent pattern: models that demonstrate higher accuracy in solving reordered problems also exhibit greater distinctiveness in their internal representations of different rule orderings. This suggests that successful mathematical reasoning can coexist with varied internal representations, challenging the assumption that answer invariance necessitates representation invariance. The study emphasizes the importance of understanding representational dynamics in mathematical reasoning beyond mere answer accuracy.
Methodology
The study constructs synthetic multi-step function-composition problems with fixed correct answers but varying rule orderings. It measures model performance using accuracy and the newly defined permutation SNR, which quantifies the separation of representations for different orderings relative to within-group variation.
Results
The analysis shows that across 16 language models, there is a strong positive correlation (Spearman correlation reaching 0.86) between the layer-averaged permutation SNR and the accuracy of the models in solving reordered problems. Models that perform better at these tasks also represent different rule orderings more distinctly.
Implications
The findings suggest that understanding the representational aspects of models can provide deeper insights into their reasoning capabilities, particularly in mathematical contexts. This could inform future research on model design and evaluation, emphasizing the need to consider internal representation dynamics alongside accuracy.
Resource-Efficient Distributed Recursive Gaussian Processes
Efficient ML
Robotics
Theory
- Development of two distributed recursive GP algorithms: ADMM-RGP and PDMM-RGP.
- Analysis of stability and convergence with parameter selection strategies for faster convergence.
- Validation on a real-world multi-output wind dataset showing effectiveness in learning unknown functions.
- Significant reduction in communication compared to existing methods while maintaining accuracy.
Read more
Resource-Efficient Distributed Recursive Gaussian Processes
Summary
This paper addresses the challenges of applying Gaussian processes (GPs) in multi-agent systems, particularly focusing on the need for efficient distributed algorithms that can operate with local measurements and limited communication. The authors propose two novel distributed recursive GP algorithms, ADMM-RGP and PDMM-RGP, designed for multi-output GP regression. These algorithms leverage communication-efficient methods, specifically the alternating direction method of multipliers (ADMM) and the primal-dual method of multipliers (PDMM), to enhance the performance of GPs in distributed settings. The paper provides a thorough analysis of the stability and convergence of both algorithms, along with parameter selection strategies to accelerate convergence and reduce communication overhead. The proposed methods are validated using a real-world multi-output wind dataset, demonstrating their effectiveness in maintaining estimation accuracy while significantly reducing communication compared to existing state-of-the-art methods. The results indicate that ADMM-RGP and PDMM-RGP can effectively operate in sparse communication networks, making them suitable for applications in areas such as environmental monitoring and sensor networks.
Methodology
The authors developed two algorithms, ADMM-RGP and PDMM-RGP, for distributed recursive multi-output GP regression. ADMM-RGP utilizes an alternating direction method of multipliers for efficient communication, while PDMM-RGP employs a primal-dual method. Both algorithms were analyzed for stability and convergence, and parameter selection strategies were introduced to enhance performance. The methods were tested on a real-world dataset to assess their effectiveness.
Results
The proposed algorithms demonstrated a significant reduction in communication needs compared to the state-of-the-art Consensus-RGP, while achieving comparable estimation accuracy and consensus across the network. ADMM-RGP showed faster convergence due to the implemented parameter selection strategies, and both algorithms were effective in various communication graph scenarios.
Implications
The findings suggest that ADMM-RGP and PDMM-RGP can be effectively utilized in real-world applications involving multi-agent systems, such as environmental monitoring, smart grids, and logistics, where efficient data fusion and communication are critical. The algorithms' ability to operate under sparse communication conditions enhances their practical applicability in resource-constrained environments.
Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
Computer Vision
Audio & Speech
Efficient ML
- ACORN optimizes expert review allocation based on ecological goals rather than classifier accuracy.
- The method effectively combines ML predictions with expert reviews to improve occupancy modeling.
- ACORN significantly reduces the number of expert reviews needed while maintaining ecological inference quality.
- The approach is applicable to various ecological data sources, including camera traps and bioacoustic sensors.
Read more
Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
Summary
This paper addresses the challenges of using machine learning (ML) in ecological studies, particularly in biodiversity surveys where accurate species identification is crucial. The authors introduce Active Continuous-Score Occupancy Modeling (ACORN), a novel method that integrates ML predictions into occupancy models while strategically selecting samples for expert review. The goal is to optimize the allocation of expert review efforts based on ecological objectives rather than purely on classifier accuracy. The method is benchmarked against traditional review policies using two datasets from camera traps and bioacoustic sensors. Results indicate that ACORN can recover ecological conclusions similar to those derived from fully human-labeled data while significantly reducing the number of required expert reviews. This approach emphasizes the importance of aligning review efforts with downstream ecological inference, thereby enhancing the efficiency of large-scale biodiversity monitoring.
Methodology
The authors developed ACORN, which employs Bayesian experimental design to prioritize expert reviews based on the information they provide for ecological inference. The method combines noisy ML classifications with targeted review, focusing on maximizing information gain for occupancy parameters and detection factors. A stopping criterion is also proposed to determine when further expert review is no longer beneficial.
Results
ACORN demonstrated the ability to recover ecological conclusions comparable to those from fully human-labeled datasets while requiring fewer expert reviews. The method was benchmarked against five different review policies across two ecological datasets, showing superior performance in terms of ecological inference metrics.
Implications
The findings suggest that ML-assisted workflows can enhance biodiversity monitoring by optimizing expert review efforts, making it feasible to analyze large datasets efficiently. This approach could lead to more timely and accurate ecological insights, ultimately supporting conservation efforts and policy-making.
Gap-Free Streaming PCA Beyond Rank-One Updates: Near-Optimal Rates and Applications to Differential Privacy
Theory
Efficient ML
- Introduces a gap-free analysis of Oja's algorithm for streaming PCA.
- Achieves near-optimal convergence rates without eigengap assumptions.
- Extends results to approximate PCA and differentially private PCA.
- Utilizes a second moment bound on stochastic updates for convergence guarantees.
Read more
Gap-Free Streaming PCA Beyond Rank-One Updates: Near-Optimal Rates and Applications to Differential Privacy
Summary
This paper presents a novel analysis of Oja's algorithm for streaming principal component analysis (PCA) in a gap-free setting, where no eigengap assumptions are made on the underlying mean matrix. The authors demonstrate that Oja's algorithm can achieve near-optimal convergence rates under weaker regularity conditions than previously required. They establish a new convergence guarantee that relies solely on a second moment bound of the stochastic updates, circumventing the need for almost sure bounds used in prior analyses. Additionally, the paper extends the results to approximate PCA using a Rayleigh quotient approach. A significant application of this work is in differentially private PCA, where the authors provide guarantees for sub-Gaussian data, addressing a conjecture from previous literature. The results indicate that Oja's algorithm can effectively recover leading eigenvectors in streaming settings without the limitations of eigengaps or rank restrictions, thus broadening its applicability in various domains.
Methodology
The authors analyze Oja's algorithm under a general formulation of streaming PCA, focusing on the correlation PCA (cPCA) objective. They provide theoretical guarantees based on a matrix variance bound, avoiding the need for more stringent conditions used in previous works. The analysis includes a lower bound for convergence rates and explores extensions of Oja's algorithm to block matrices. The paper also discusses the implications of their findings in the context of differentially private PCA.
Results
The main results include a convergence guarantee for Oja's algorithm that holds with constant probability, using a sample complexity that scales with the matrix variance and other parameters. The authors also derive a nearly-matching lower bound for their convergence rates. For differentially private PCA, they provide a sample complexity bound that matches previous results under the assumption of eigengaps, thus extending the applicability of their findings to gap-free scenarios.
Implications
The findings have significant implications for streaming PCA applications, particularly in scenarios where eigengaps cannot be assumed. The results also enhance the understanding of differentially private PCA, making it feasible to apply these techniques to a broader range of data types while maintaining privacy guarantees.