AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
50
Papers today
8h
Update frequency
7
Days of history
DraftFM: A Foundation Model for Day-Zero Drafting in Magic: The Gathering
Theory
- DraftFM is designed specifically for day-zero drafting in MTG, addressing the lack of available draft logs.
- The model uses a frozen vector representation of cards, allowing it to generalize to unseen expansions effectively.
- DraftFM achieved significant predictive accuracy, outperforming existing models in day-zero scenarios.
- The model's predictions for The Hobbit set were validated against expert opinions, demonstrating its practical applicability.
Read more
DraftFM: A Foundation Model for Day-Zero Drafting in Magic: The Gathering
Summary
This paper introduces DraftFM, a novel discrete-choice policy model designed for day-zero drafting in Magic: The Gathering (MTG), where players draft cards from a newly released expansion without prior draft logs. DraftFM operates by scoring cards based solely on their public information, utilizing a 775-dimensional frozen vector representation that includes structured features and a fixed text embedding. This approach allows the model to evaluate unseen cards using the same mechanisms as familiar ones, thereby addressing the challenge of drafting with no historical data. The model, comprising 1.6 million parameters, was trained on 149 million human picks from 29 expansions and demonstrated impressive predictive capabilities with top-1 agreement rates of 50.8%, 60.4%, and 56.7% on held-out expansions. Furthermore, DraftFM was able to produce a card ranking for the unreleased set The Hobbit, which was validated against expert reviews, showing a high level of agreement. This research highlights the potential of using feature-only representations for behavioral prediction in environments with limited data availability.
Methodology
DraftFM employs a discrete-choice policy framework, utilizing a 775-dimensional frozen vector representation for each card based on public information. The model was trained on a large dataset of human picks and evaluated on held-out expansions to assess its predictive performance without relying on usage statistics from the target set.
Results
DraftFM achieved top-1 agreement rates of 50.8%, 60.4%, and 56.7% on three held-out expansions, significantly higher than the baseline of 7% for random picks. Additionally, the model produced a card ranking for The Hobbit that aligned closely with expert evaluations, indicating its effectiveness in predicting card value in a new set.
Implications
The findings suggest that feature-only models can effectively address challenges in environments with limited historical data, such as day-zero drafting in card games. This approach could be extended to other domains where real-time decision-making is required without prior data.
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
Reinforcement Learning
Large Language Models
Optimization
- MileGPO reveals the limitations of final-goal-distance credit in distinguishing between same-state branches.
- The proposed method learns from rollout graphs and on-policy rewards without the need for external annotations or auxiliary inference.
- MileGPO achieves state-of-the-art performance on challenging benchmarks ALFWorld and WebShop.
- The reliability weighting mechanism enhances credit assignment by focusing on outcome-consistent preferences.
Read more
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
Summary
The paper introduces MileGPO, a novel approach to credit assignment in long-horizon reinforcement learning (RL) for large language model (LLM) agents. Traditional methods struggle with assigning credit to intermediate decisions due to reliance on final rewards. MileGPO addresses this by deriving process-level credit from grouped on-policy rollouts through three key components: Milestone Discovery (MD) identifies significant milestones and traps in trajectories; Reliability-Calibrated Shaping (RCS) weights these milestones based on their reliability; and Progress-Contrastive Calibration (PCC) evaluates local progress and compares transitions from the same state. This approach does not require auxiliary models or additional environment interactions. Experiments on ALFWorld and WebShop demonstrate that MileGPO achieves state-of-the-art performance, with a notably small in-distribution to out-of-distribution gap, indicating robust generalization. The findings highlight the importance of calibrated intermediate credit in improving RL performance, revealing that traditional final-goal distance metrics can obscure meaningful distinctions in credit assignment.
Methodology
MileGPO employs a three-part methodology: Milestone Discovery (MD) to identify significant milestones and traps from trajectories, Reliability-Calibrated Shaping (RCS) to weight these milestones based on their reliability, and Progress-Contrastive Calibration (PCC) to assess local progress and compare transitions from the same state. This approach leverages grouped on-policy rollouts to derive process-level credit without requiring auxiliary models or additional interactions with the environment.
Results
MileGPO demonstrated state-of-the-art performance on the ALFWorld and WebShop benchmarks, with a minimal in-distribution to out-of-distribution gap of 1.69 points, outperforming previous methods like GiGPO and GraphGPO. Ablation studies confirmed the effectiveness of the reliability weighting and local progress evaluation in enhancing credit assignment.
Implications
The findings suggest that MileGPO can significantly improve the training of long-horizon LLM agents in complex tasks, enhancing their decision-making capabilities. This approach could be applied to various RL scenarios where credit assignment is challenging, potentially leading to more efficient learning and better performance in real-world applications.
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
Reinforcement Learning
- Introduces adaptive probabilistic shielding for safe reinforcement learning.
- Combines online model learning with probabilistic shielding to improve safety.
- Utilizes interval MDPs to handle uncertainty in transition probabilities.
- Empirical evaluations show the effectiveness of the proposed method across multiple environments.
Read more
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
Summary
This paper addresses the challenge of safe reinforcement learning (RL) by introducing an adaptive probabilistic shielding technique that integrates online model learning with probabilistic shielding. Traditional probabilistic shielding relies on a fully specified Markov decision process (MDP), which is often not available in real-world RL applications. The authors propose a method where the transition graph of the MDP is known, but the transition probabilities are estimated as the RL agent explores the environment. This approach allows for the computation of a shield that adapts over time, becoming less conservative as the model estimates improve. The paper discusses the challenges of balancing exploration and safety, and presents a methodology that employs interval MDPs (iMDPs) to robustly capture data uncertainty. The authors conduct empirical evaluations across various environments to assess the effectiveness of their adaptive shielding algorithm, demonstrating its ability to enhance safety while allowing for exploration.
Methodology
The authors developed an adaptive shielding algorithm that estimates transition probabilities from data collected during the RL process. They implemented interval MDPs to define upper and lower bounds on these probabilities, allowing for the computation of safety probabilities. The shield is updated adaptively based on the evolving model estimates, and the authors explored various strategies to balance exploration and safety during learning.
Results
The experimental results indicate that the adaptive shielding approach successfully reduces the conservativeness of the shield over time, allowing for more effective exploration while maintaining safety. The evaluations across different environments demonstrated that the proposed method can effectively manage the trade-off between exploration and safety, leading to improved learning outcomes.
Implications
This research has significant implications for the development of safe RL systems, particularly in real-world applications where safety is critical. The adaptive shielding technique can be applied in robotics, autonomous systems, and other domains where agents must learn from uncertain environments while adhering to safety constraints.
Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic
Reinforcement Learning
Theory
Optimization
- Introduces a single-loop, entropy-regularized NAC algorithm for improved convergence rates.
- Analyzes two optimization regimes: Stochastic and Deterministic.
- Presents an Exponential Translation mechanism to link regularized and unregularized gaps.
- Achieves accelerated convergence rates, outperforming traditional methods in certain settings.
Read more
Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic
Summary
This paper investigates the unregularized convergence properties of a single-loop, entropy-regularized Natural Actor-Critic (NAC) algorithm. While entropy regularization is commonly employed to enhance the stability and speed of Natural Policy Gradient methods, its impact on the convergence rates of the unregularized objective has not been thoroughly examined. The author presents a novel analysis that utilizes a single-loop architecture and compatible linear function approximation, focusing on two optimization regimes: the Stochastic Regime and the Deterministic Regime. In the Stochastic Regime, the paper introduces a joint Lyapunov recurrence for coupled actor-critic updates, while in the Deterministic Regime, it employs a Policy Mirror Descent framework to address challenges posed by the degeneration of the Fisher information matrix. A key contribution is the Exponential Translation mechanism, which connects the regularized and unregularized gaps, leading to accelerated convergence rates. The findings indicate that by adjusting the fixed temperature, the algorithm can achieve improved unregularized convergence rates of ËœO(T^(-1)) in the Stochastic Regime and ËœO(T^(-2/3)) for the average iterate and ËœO(T^(-1/3)) for the last iterate in the Deterministic Regime. Additionally, in a tabular setting, the analysis reveals a ËœO(T^(-2/3)) average-iterate rate, surpassing the O(T^(-1/2)) barrier typically observed without a positive action margin.
Methodology
The paper employs a theoretical analysis of a single-loop, entropy-regularized NAC algorithm, utilizing linear function approximation. It examines the optimization landscape in both stochastic and deterministic regimes, employing Lyapunov recurrences and Policy Mirror Descent frameworks to derive convergence rates and bounds.
Results
The analysis reveals that the proposed algorithm achieves unregularized convergence rates of ËœO(T^(-1)) in the Stochastic Regime and ËœO(T^(-2/3)) for the average iterate and ËœO(T^(-1/3)) for the last iterate in the Deterministic Regime. In tabular settings, it achieves a ËœO(T^(-2/3)) average-iterate rate, surpassing the O(T^(-1/2)) worst-case statistical barrier.
Implications
The findings suggest that entropy regularization can be effectively leveraged to enhance the convergence properties of NAC algorithms, potentially leading to more efficient reinforcement learning applications. This work could influence future research on actor-critic methods and their practical implementations in various reinforcement learning tasks.
PETA: Parameter-Efficient Test-Time Adaptation for Virtual Screening
Efficient ML
- PETA formulates Test-Time Adaptation for Virtual Screening (TTA-VS), allowing specialization of pretrained models without full retraining.
- The framework updates only a small fraction of model parameters (approximately 0.03%), enhancing efficiency.
- PETA constructs target-conditioned supervision from available test-time information, improving ranking tasks.
- Experimental results show that PETA outperforms both pretrained and fully retrained models in various benchmarks.
Read more
PETA: Parameter-Efficient Test-Time Adaptation for Virtual Screening
Summary
The paper addresses the challenge of accurately ranking active ligands for target protein pockets from extensive chemical libraries in virtual screening. While existing methods like DrugCLIP have improved efficiency by encoding protein pockets and molecules into a shared embedding space, they often require full model retraining for performance enhancements, leading to high computational costs. The authors propose PETA, a parameter-efficient framework that adapts pretrained virtual screening models at test time without the need for extensive retraining. PETA constructs pocket-specific negative samples through molecular diffusion and chemical validity filtering, and enhances the ranking process by emphasizing the suppression of high-scoring invalid candidates. The framework updates only the LayerNorm parameters, which constitute about 0.03% of the full model, while achieving superior performance compared to both frozen and fully retrained baselines. The experiments demonstrate that PETA effectively balances screening performance and adaptation efficiency, making it a promising approach for virtual screening tasks.
Methodology
PETA employs a test-time adaptation approach that updates only the LayerNorm parameters of pretrained models. It constructs pocket-specific supervision by retrieving a reference ligand and generating pocket-conditioned candidates. The framework uses molecular diffusion and chemical validity filtering to create challenging ranking tasks, and a cost-sensitive ranking objective to prioritize the correction of high-scoring invalid candidates.
Results
PETA achieved an EF0.5% score of 42.34, outperforming the frozen DrugCLIP (37.90) and fully retrained BindCLIP (39.82). This demonstrates that the lightweight adaptation can lead to significant performance improvements without the computational burden of full model retraining.
Implications
The proposed method has the potential to streamline the virtual screening process in drug discovery, allowing for rapid adaptation to new targets without extensive computational resources. This could facilitate the discovery of new drug candidates from large chemical libraries more efficiently.
Auditing Recorded Predictive Lead Service-Line Classifications Against Physical Verification: A Statewide Study of New York
Theory
- Nearly 49% of localities showed zero output variance in service line material classifications.
- Seven localities contradicted their own physical verification results, raising concerns about model reliability.
- New York City's predictive model classified a large number of addresses without variation, potentially misrepresenting lead status.
- The study highlights significant discrepancies between predictive model outputs and physical verification results across construction eras.
Read more
Auditing Recorded Predictive Lead Service-Line Classifications Against Physical Verification: A Statewide Study of New York
Summary
This paper investigates the reliability of predictive models used by water utilities in New York State to classify service line materials under the US Lead and Copper Rule Revisions. The study audits the outputs of these models against physical verification data, focusing on 153 localities that classified at least 100 addresses. The findings reveal that nearly half of these localities exhibit zero output variance in their classifications, with 68 of the 75 consistent with physical verification. However, seven localities, including five boroughs of New York City, show contradictions in their classifications. The analysis highlights that New York City's predictive model classified 43,215 addresses as 'Known Other,' indicating a lack of variation and potentially misleading reporting. The study also compares the predictive model's findings with physical verification results across different construction eras, revealing significant discrepancies. The paper emphasizes the need for rigorous validation of predictive models and adherence to state guidance on physical verification.
Methodology
The study employed a screening process to analyze the classification outputs of predictive models across 153 localities in New York State. It compared model outputs with physical verification data, focusing on zero output variance and contradictions in classifications. The analysis included a spatial join with construction era data to assess the impact of housing stock age on lead classification rates.
Results
The analysis found that 75 out of 153 localities had zero output variance in their predictive model classifications. Among these, 68 were consistent with physical verification, while seven showed contradictions. New York City's model classified 43,215 addresses as 'Known Other,' with no recorded lead status, contrasting sharply with other regions where lead was identified in 12.21% of addresses. The study estimated that the expected number of lead public-side lines among model-cleared addresses ranged from 1,150 to 1,450.
Implications
The findings suggest that reliance on predictive models without adequate physical verification can lead to significant misrepresentations of lead service line statuses. This has implications for public health and regulatory compliance, highlighting the need for utilities to validate their models and ensure accurate reporting. The study may inform policy changes regarding the use of predictive models in environmental health assessments.
Uncovering the Limits of Proof Sharing for Neural Networks
Theory
Efficient ML
- Systematic study of template-based proof sharing in neural network verification.
- Introduction of the jointly stable neurons metric to assess template effectiveness.
- Development of FastCert for automatic template distribution across layers.
- Demonstrated average speedup of 1.13× over existing methods in L0-verification tasks.
Read more
Uncovering the Limits of Proof Sharing for Neural Networks
Summary
This paper investigates the effectiveness of proof sharing techniques in the robustness verification of neural networks, particularly focusing on template-based acceleration methods. The authors conduct a systematic study to understand the limits of template reuse across various neural network architectures, properties, datasets, and training methods. They introduce a novel metric called jointly stable neurons to characterize the conditions under which template-based techniques can accelerate verification processes. The findings reveal significant variability in the effectiveness of template reuse, suggesting that in some scenarios, these techniques may not yield any speedup. To address this, the authors propose FastCert, a new method that automatically distributes templates across network layers to optimize performance. FastCert is designed to predict the potential for speedup by profiling a small set of queries, allowing it to either utilize templates effectively or avoid them when they are unlikely to help. The implementation of FastCert demonstrates an average speedup of 1.13× over existing template-based methods in L0-verification tasks, significantly reducing verification time.
Methodology
The authors conducted a limit study to analyze the effectiveness of template-based proof sharing, introducing a metric for jointly stable neurons to evaluate the potential for speedup. They developed FastCert, which profiles and generates templates before verification, predicting subsumption rates based on a small sample of queries.
Results
FastCert achieved an average speedup of 1.13× over existing template-based techniques in covering-design based L0-verification tasks, resulting in a significant reduction in wall-clock time for verification processes.
Implications
The findings suggest that while template-based acceleration can be beneficial, its effectiveness is highly context-dependent. FastCert provides a practical tool for optimizing robustness verification in neural networks, which could enhance the reliability of neural networks in critical applications such as autonomous driving and medical diagnosis.
FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration
Large Language Models
Optimization
Efficient ML
- FleetSieve optimizes LLM fleet configuration by focusing on decision-critical profiling rather than exhaustive measurements.
- The method models capacity and tail latency jointly to ensure configurations meet SLO requirements.
- FleetSieve demonstrates significant reductions in profiling GPU-seconds compared to traditional methods.
- The approach provides a three-state result indicating feasibility, undecided status, or infeasibility of configurations.
Read more
FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration
Summary
The paper introduces FleetSieve, a novel approach for optimizing the configuration of large language model (LLM) serving fleets by focusing on decision-critical profiling. Traditional exhaustive profiling methods are inefficient as they often measure configurations that do not impact final resource allocation. FleetSieve addresses this by selecting measurements based on their expected influence on resource-coupled, service-level objective (SLO)-aware decisions. It jointly models capacity and tail latency, allowing it to compare conservative and optimistic resource allocations. The profiling process halts when the decision gap between these allocations falls below a specified tolerance. Evaluations on a fixed H100 measurement grid for a 31B-parameter model demonstrate that FleetSieve achieves significant reductions in GPU-seconds used for profiling, outperforming uniform random profiling by 6.9% and achieving a mean saving of 5.4% over 200 random reveal orders. The method also ensures that configurations do not violate critical SLOs, thus enhancing the efficiency and reliability of LLM fleet configurations.
Methodology
FleetSieve employs a decision-focused profiling strategy that prioritizes measurements based on their predicted impact on resource allocation decisions. It maintains uncertainty over capacity and tail latency, evaluates configurations under conservative and optimistic scenarios, and terminates profiling when the decision gap is sufficiently small. The method incorporates joint modeling of capacity and tail latency to prevent infeasible configurations from being certified.
Results
FleetSieve achieved a reduction of 22,200 GPU-seconds in aggregate profiling compared to uniform random profiling, which is a 6.9% improvement. Across 200 random reveal orders, the mean saving was 5.4% (with a 95% bootstrap confidence interval of 3.5–7.2%). The method also ensured that configurations did not violate SLOs, maintaining a completion time within the required limits.
Implications
FleetSieve's approach can significantly enhance the efficiency of LLM serving systems by reducing the computational resources needed for profiling while ensuring compliance with performance objectives. This can lead to more effective resource management in cloud-based AI services and improve the overall user experience by maintaining service quality.
SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
Interpretability
- Introduction of a geographic location-aware SAE, GeoTopK, for better feature extraction in ExEE.
- Development of SAE-Xplainers for rule-based interpretation of high-dimensional features.
- Validation of the proposed methods on predicting fires and detecting tropical cyclones and atmospheric rivers.
- Demonstration of improved reconstruction performance and feature utilization through geographic modulation.
Read more
SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
Summary
The paper addresses the challenge of interpreting deep learning models used for predicting extreme Earth events (ExEE) such as wildfires, tropical cyclones, and atmospheric rivers. Despite the potential of large-scale Weather and Climate (W&C) datasets, the 'black box' nature of deep learning models limits their operational adoption due to a lack of interpretability. The authors propose a novel approach that combines geographic location-based modulation of Sparse Autoencoders (SAEs) with an ensemble of rule-based models, termed SAE-Xplainers, to extract and interpret high-dimensional features from complex environmental data. The geographic modulation captures local semantic meanings of environmental patterns, enhancing the model's performance and interpretability. The method is validated across three types of ExEE, demonstrating that the SAE input modulation improves reconstruction performance and feature utilization. The SAE-Xplainers provide human-understandable rules that align with scientific literature, facilitating the identification of feature absorption phenomena. Overall, the proposed framework enhances the interpretability of models in the context of extreme climate events, contributing to better risk management and preparedness strategies.
Methodology
The authors introduce a geographic location encoder that modulates the inputs of a k-sparse autoencoder (GeoTopK) to create location-aware projections. This is followed by the application of an ensemble of rule-based models (SAE-Xplainers) to interpret the extracted features. The approach is validated using two different encoders and datasets covering three types of extreme Earth events.
Results
The proposed geographic modulation significantly enhances the reconstruction performance and reduces the rate of dead features compared to location-agnostic models. The SAE-Xplainers successfully interpret the features into human-understandable rules, revealing insights consistent with scientific literature and enabling the identification of feature absorption.
Implications
The framework developed in this paper has significant implications for improving the interpretability of deep learning models in Earth sciences, which is crucial for effective forecasting, detection, and risk management of extreme climate events. It can facilitate better decision-making in climate-related policies and disaster preparedness.
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
Large Language Models
Optimization
Efficient ML
- Introduces a compute-efficient framework for hyperparameter transfer in large-scale MoE models.
- Utilizes Maximal Update Parameterization (µP) for consistent learning rate transfer across model widths.
- Establishes a predictive scaling law for extrapolating learning rates to extensive training horizons.
- Validates the approach by successfully pretraining a large MoE model with minimal ablation costs.
Read more
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
Summary
This paper addresses the challenge of optimizing hyperparameters, particularly the learning rate, for large-scale Mixture-of-Experts (MoE) models. Traditional methods for hyperparameter tuning are computationally expensive, especially at extreme scales of model size and token budget. The authors propose a two-step hyperparameter transfer framework that estimates optimal learning rates by transferring knowledge across different model widths and extrapolating to large training horizons. The first step involves a Maximal Update Parameterization (µP) adaptation for MoE architectures, which allows for consistent transfer of optimal learning rates across width-scaled models. The second step establishes a predictive scaling law using linear regression to extrapolate the ideal learning rate from small proxy models to massive training scenarios, achieving high fidelity in predictions. The methodology was validated by pretraining a foundation MoE model with 155 billion total parameters over 10 trillion tokens, demonstrating that optimal configurations can be accurately predicted with minimal computational costs.
Methodology
The proposed methodology consists of a two-step hyperparameter transfer framework. The first step involves adapting the Maximal Update Parameterization (µP) to MoE architectures to enable the transfer of optimal learning rates across different model widths. The second step uses linear regression to establish a predictive scaling law that extrapolates the optimal learning rate from small proxy models to larger training scenarios, thus avoiding the need for exhaustive hyperparameter sweeps.
Results
The authors achieved a high fidelity in predicting optimal learning rates with an R2 value of 0.95 when extrapolating from small proxy models. The proposed framework was successfully applied to pretrain a foundation MoE model with 155 billion parameters over 10 trillion tokens, demonstrating stable training and evaluation results that confirm the effectiveness of the predicted configurations.
Implications
The findings suggest that the proposed hyperparameter transfer framework can significantly reduce the computational costs associated with hyperparameter tuning for large-scale MoE models. This has implications for the efficient training of large language models and other applications where MoE architectures are employed, potentially enabling broader accessibility and faster development cycles in machine learning research and industry.
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
Interpretability
- Introduces mechanistic tomography as a unified framework for measuring internal model effects.
- Establishes a common measurement language for various interpretability methods.
- Highlights the importance of calibration in recovering accurate internal mechanisms.
- Demonstrates the effectiveness of designed measurements through empirical validation.
Read more
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
Summary
This paper introduces the concept of 'mechanistic tomography', which aims to unify various measurement methods used in mechanistic interpretability of machine learning models. The author formulates the measurement problem as a linear equation where different interventions are represented, and the goal is to recover internal mechanisms and intervention effects. The framework provides a common language for methods such as coordinate patching, attribution patching, subset interventions, and Hessian-vector products, allowing for a structured approach to understanding and measuring internal model effects. The paper emphasizes the importance of calibration and the need for a measurement-first approach, especially as models become more complex and less interpretable. The author validates the proposed methods through experiments on two models, demonstrating that targeted interventions can effectively reveal internal interactions and improve predictive accuracy. The findings suggest that careful design and analysis of measurements can lead to better interpretability and control in machine learning systems.
Methodology
The author formulates the measurement problem using a linear model and explores various measurement techniques, including coordinate patching, attribution patching, and Hessian-vector products. The paper employs theoretical analysis and empirical validation on two models (GPT-2-small and Qwen-2.5-7B) to assess the effectiveness of the proposed measurement strategies.
Results
The results indicate that the proposed measurement strategies can recover internal mechanisms effectively, with a calibrated additive map achieving an R2 of 0.983 on a finite refusal-response surface. The study also reveals that targeted interventions can uncover significant interactions, such as the Name Mover–Negative Name Mover interaction in the GPT-2-small model.
Implications
The findings suggest that mechanistic tomography can enhance the interpretability of machine learning models, particularly in complex systems where traditional methods may fall short. This approach could be applied in various domains requiring model transparency and control, such as healthcare, finance, and autonomous systems.
G-MARK: Grounded Multi-Agent Reasoning for Cooperative Driving via Knowledge Graphs
Robotics
Graph Learning
Computer Vision
- G-MARK introduces a provenance-aware knowledge graph for cooperative driving, preserving critical information about object hypotheses and their sources.
- The framework supports various tasks including object reasoning, motion prediction, and trajectory forecasting using a compact communication structure.
- G-MARK significantly outperforms state-of-the-art methods in occlusion reasoning and control selection while maintaining communication efficiency.
Read more
G-MARK: Grounded Multi-Agent Reasoning for Cooperative Driving via Knowledge Graphs
Summary
The paper presents G-MARK, a novel framework for grounded multi-agent reasoning in cooperative driving scenarios, utilizing knowledge graphs (KGs) to enhance decision-making under partial observability. Traditional cooperative driving methods often compress multi-agent observations into latent features, losing critical information about object provenance, visibility, and uncertainty. G-MARK addresses this by converting cooperative object-centric observations into explicit, provenance-aware KGs that retain detailed information about object hypotheses, source attribution, visibility, and planning context. This structured representation allows for improved reasoning about occluded objects, motion prediction, control selection, and trajectory forecasting. The authors evaluate G-MARK against the V2V-GoT-QA benchmark, demonstrating significant improvements in occlusion reasoning accuracy, control-selection error, and communication efficiency. The results indicate that G-MARK not only enhances the accuracy of downstream tasks but also reduces the communication payload required for effective multi-agent cooperation.
Methodology
G-MARK operates above the perception layer, processing multi-agent observations to create knowledge graphs that explicitly store information about object hypotheses, source-agent support, visibility, uncertainty, and spatial relations. The framework employs task-conditioned delayed evidence fusion to retain weak observations, allowing downstream tasks to assess their relevance.
Results
G-MARK improves occlusion reasoning accuracy by 42.2%, reduces control-selection error by 13.1%, and achieves comparable trajectory-planning accuracy with a 25.6× smaller structured communication payload compared to existing methods.
Implications
The findings suggest that G-MARK can enhance the safety and efficiency of autonomous driving systems by enabling better decision-making in complex environments. Its structured approach to cooperative evidence representation could be applied to other multi-agent systems requiring reliable communication and reasoning.
Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification
NLP
Large Language Models
Theory
- Introduction of TUP, a policy that truncates low-ranked completions and upweights high-ranked ones.
- Theoretical support for the effectiveness of lower-tail truncation in improving model performance.
- Demonstration of TUP's competitive performance against existing offline alignment methods.
- Methodology allows for offline training using binary cross-entropy with shifted-truncated win-rate labels.
Read more
Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification
Summary
This paper introduces TUP (Truncate-bad, Upweight-good), a novel policy for Best-of-N (BoN) style distillation that enhances language model generation quality. The authors argue that existing rank-based policies often retain low-ranked completions, which can dilute the effectiveness of the model. TUP addresses this by truncating low-ranked completions from the support and applying a sharper reweighting to the retained higher-ranked completions. The methodology involves a shifted-truncated win-rate transform that assigns zero mass to completions below a certain threshold while softly reweighting those above it. The paper provides theoretical justifications for this approach, demonstrating that lower-tail truncation can match the performance of traditional monotone rank-reweighting policies. Empirical evaluations show that TUP performs competitively against leading offline alignment baselines across various reward models, indicating its potential for improving language model alignment and generation.
Methodology
The TUP policy employs a shifted-truncated win-rate transform to decouple the truncation of low-ranked completions from the upweighting of higher-ranked ones. The model is trained offline using binary cross-entropy loss, leveraging empirical in-pool ranks derived from a reference policy and a reward model. The approach allows for prompt-independent normalization and focuses on optimizing the relative ranking of completions.
Results
TUP was evaluated on the QRPO benchmark with various models, including Llama-8B and Mistral-7B-Instruct-v0.2. The results indicate that TUP is competitive with four leading offline alignment baselines, demonstrating its effectiveness in improving language model generation quality.
Implications
The findings suggest that TUP can significantly enhance the efficiency and quality of language model generation by optimizing the selection process during training. This has potential applications in various NLP tasks where generation quality is critical, such as conversational agents, content creation, and automated summarization.
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
Large Language Models
- LLM-Detector utilizes in-context learning to derive anomaly detection logic from structured normal-state knowledge without requiring LLM fine-tuning.
- The framework synthesizes a scoring engine that evaluates anomalies based on statistical deviation, structural inconsistency, and density-based abnormality.
- Extensive evaluations on 24 tabular datasets show LLM-Detector outperforms 15 state-of-the-art anomaly detection methods.
- The approach effectively handles mixed-type data and reduces computational costs associated with traditional anomaly detection methods.
Read more
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
Summary
This paper presents LLM-Detector, a novel framework that leverages the in-context learning capabilities of Large Language Models (LLMs) for anomaly detection in tabular data. Traditional anomaly detection methods often struggle with the complex relationships between features in tabular datasets, as anomalies typically arise from violations of cross-feature dependencies rather than simple numerical deviations. Existing LLM-based approaches either fine-tune models on normal samples or generate synthetic anomalies, which can be computationally expensive and inefficient. LLM-Detector addresses these challenges by synthesizing an executable anomaly detection engine from structured normal-state knowledge without the need for LLM fine-tuning. The framework operates in two phases: first, it extracts statistical profiles, causal relationships, and representative prototypes from normal training data to create a structured prompt that guides the LLM in generating a scoring logic program. In the second phase, this program evaluates unseen samples by aggregating signals of statistical deviation, structural inconsistency, and density-based abnormality to compute anomaly scores. The authors evaluate LLM-Detector on 24 diverse tabular datasets, demonstrating its superior performance compared to 15 state-of-the-art baselines, highlighting its effectiveness in both mixed-type and continuous-only settings. This approach significantly reduces computational costs and enhances practical applicability in real-world scenarios.
Methodology
The methodology involves two main phases: (1) extraction of statistical summaries, causal dependencies, and prototypes from normal training data to create a structured prompt for LLMs, and (2) generation of a deterministic program by the LLM that implements the anomaly detection logic, which evaluates test samples to compute anomaly scores.
Results
The results indicate that LLM-Detector consistently achieves improved detection performance across various datasets compared to existing statistical, deep learning, and LLM-based methods, demonstrating its robustness in handling both mixed-type and continuous-only data.
Implications
The findings suggest that LLM-Detector can be effectively applied in critical domains such as finance, healthcare, and cybersecurity, where accurate anomaly detection is essential for preventing risks and failures. Its efficiency and reduced computational requirements make it suitable for real-world applications.
When to Retrain: An Empirical Study of Retraining Policies for Streaming ML Under Concept Drift, Budget, and Latency Constraints
Theory
Efficient ML
Time Series
- Incremental learning significantly impacts the effectiveness of retraining policies.
- Periodic retraining outperforms reactive policies under abrupt and gradual drift without incremental updates.
- Reactive policies exhibit systematic failures, particularly in handling pre-drift noise and gradual drift.
- Latency constraints can halve effective retraining budgets, complicating decision-making.
Read more
When to Retrain: An Empirical Study of Retraining Policies for Streaming ML Under Concept Drift, Budget, and Latency Constraints
Summary
This paper addresses the critical challenge of determining when to retrain machine learning models in production systems that experience concept drift. The author presents an empirical study comparing various retraining policies, including periodic retraining, error-threshold triggering, and statistical drift-triggered retraining using ADWIN, against a no-retrain baseline. The study is conducted within a unified system model that explicitly incorporates retraining budgets and latency constraints. The experiments span 3,933 runs across three drift regimes (abrupt, gradual, recurring), three budget levels, and varying latency levels, using both synthetic and real-world datasets. The findings reveal that the most significant factor influencing model performance is whether the deployed model can learn incrementally. When incremental updates are enabled, no retraining policy significantly outperforms the no-retrain baseline. Conversely, without incremental updates, periodic retraining shows substantial advantages, particularly under abrupt and gradual drift scenarios. The study also uncovers systematic failures of reactive policies and highlights the interaction between latency and budget constraints, which can severely limit effective retraining opportunities. The author provides a reproducible experimental framework, making all data and artifacts publicly available for further research.
Methodology
The study employs a controlled empirical framework that simulates streaming machine learning environments. It evaluates different retraining policies under varying conditions of concept drift, budget constraints, and latency. The experiments utilize a factorial design across multiple datasets, including a synthetic benchmark and real-world applications, with statistical significance testing to compare policy effectiveness.
Results
The results indicate that with per-sample incremental updates, no retraining policy significantly improves performance over the no-retrain baseline. However, without incremental updates, periodic retraining outperforms reactive policies by 15-55 percentage points in post-drift accuracy, especially under abrupt and gradual drift conditions. Reactive policies only show advantages in recurring drift scenarios. Additionally, the study identifies specific failure modes in reactive policies and highlights the detrimental effects of latency on retraining budgets.
Implications
The findings suggest that practitioners should prioritize incremental learning capabilities in their models to enhance performance under concept drift. The study also provides insights into the design of retraining policies, emphasizing the importance of periodic retraining in certain scenarios. The released framework can facilitate further research into effective retraining strategies in streaming machine learning.
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Efficient ML
Theory
Interpretability
- DICS utilizes clustering to create a reduced set of candidate splits for decision trees.
- Theoretical analysis shows that DICS does not compromise classification performance.
- Extensive experiments reveal significant reductions in training time while maintaining accuracy.
- DICS can be integrated into various tree-based models, enhancing their efficiency.
Read more
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
Summary
The paper introduces Data-Informed Centroid Splitting (DICS), a novel framework aimed at enhancing the efficiency of decision tree classifiers by reducing the computational burden associated with exhaustive split searches. DICS leverages clustering techniques to generate a compact set of candidate splits informed by the data's structure, thereby streamlining the training process. The authors provide theoretical analysis confirming that DICS maintains the predictive performance of classification trees compared to traditional exhaustive methods. Extensive experiments on synthetic and benchmark datasets demonstrate that DICS not only achieves comparable accuracy but also significantly reduces training time, making it a promising approach for scalable decision tree learning. The method is applicable to various tree-based models, including classification trees, random forests, and gradient-boosting models, highlighting its versatility in improving computational efficiency in machine learning tasks.
Methodology
DICS employs a clustering-based approach to derive a compact set of candidate splits for decision trees. By analyzing the data's class-aware structure, it reduces the search space for splits, thus accelerating the training process. The method is theoretically validated to ensure that it does not degrade the performance of classification trees compared to exhaustive search methods.
Results
The experiments conducted demonstrate that DICS achieves similar accuracy levels to traditional methods while significantly decreasing the training time across both synthetic and benchmark datasets. This efficiency gain underscores the effectiveness of integrating data-informed priors into the split selection process.
Implications
DICS has the potential to enhance the scalability and efficiency of decision tree-based models in various applications, such as spam filtering, fraud detection, and targeted advertising, where large and complex datasets are common. Its integration into existing models could lead to faster training times and improved performance in real-world scenarios.
End-to-end Early Classification of Time Series in Non-Stationary Environments
Reinforcement Learning
Time Series
Optimization
- Introduces DQeND, an end-to-end framework for ECTS that optimizes classification and triggering jointly.
- Demonstrates the limitations of separable designs in adapting to non-stationary environments.
- Provides empirical evidence that end-to-end optimization improves robustness and adaptability.
- Highlights the importance of joint learning in dynamic environments through controlled experiments.
Read more
End-to-end Early Classification of Time Series in Non-Stationary Environments
Summary
This paper addresses the challenge of Early Classification of Time Series (ECTS) in non-stationary environments, where the underlying data distribution can change over time. Traditional methods often assume stationarity and optimize classification and triggering decisions separately, which limits their adaptability to concept drift. The authors propose DQeND, a novel end-to-end architecture that integrates representation learning, classification, and triggering into a single framework, leveraging Reinforcement Learning. Through systematic comparisons with separable approaches, the study demonstrates that DQeND significantly outperforms traditional methods across various drifting scenarios. An ablation study reveals that the joint optimization of representation and decision-making modules is crucial for achieving these improvements. The findings suggest that end-to-end learning enhances the robustness and adaptability of ECTS systems in dynamic environments, paving the way for further exploration of integrated approaches in time series classification.
Methodology
The authors conducted extensive empirical comparisons between separable and end-to-end ECTS architectures under controlled non-stationary conditions, utilizing Reinforcement Learning to develop the DQeND framework. They designed experiments to simulate covariate shifts and concept drifts, allowing for a thorough evaluation of the models' performance in adapting to changing data distributions.
Results
The results indicate that DQeND consistently outperforms separable baselines in various non-stationary scenarios, demonstrating enhanced robustness and adaptability. The ablation study confirms that the joint optimization of representation and decision-making components is critical for achieving superior performance.
Implications
The findings suggest that adopting end-to-end learning frameworks can significantly improve the performance of ECTS systems in real-world applications, such as monitoring social media for harmful content, where timely and accurate classification is essential. This work encourages further investigation into integrated approaches for time series classification in dynamic environments.
Exact Algebraic Computation of Learning Coefficients for Two-Dimensional Singular Models
Theory
Efficient ML
- Introduces a deterministic algorithm for exact computation of local RLCTs for two-dimensional models.
- Demonstrates the limitations of classical information criteria in singular models and proposes WBIC as a solution.
- Reveals that effective model complexity can decrease with increasing depth in polynomial neural networks.
- Provides a new perspective on identifiability in deep learning architectures.
Read more
Exact Algebraic Computation of Learning Coefficients for Two-Dimensional Singular Models
Summary
This paper addresses the limitations of classical information criteria, such as the Bayesian Information Criterion (BIC), in the context of singular models, particularly in deep learning. The authors introduce the Widely Applicable Bayesian Information Criterion (WBIC), which utilizes local learning coefficients (λ) to accurately capture the asymptotic behavior of marginal likelihoods. They present a novel deterministic algorithm capable of computing local Real Log Canonical Thresholds (RLCTs) exactly for any two-dimensional model whose Kullback-Leibler divergence is contact equivalent to a polynomial. This advancement overcomes the reliance on sampling-based methods, which are often limited to special cases. The authors derive complexity bounds for their algorithm and demonstrate its effectiveness across various models, including polynomial neural networks (PNNs). The findings reveal that the effective model complexity, as indicated by the RLCT, can decrease with the number of layers in PNNs, suggesting a non-trivial theory of identifiability in these models. This work provides a new computational tool for analyzing loss landscapes in learning theory and enhances the understanding of model selection in singular settings.
Methodology
The authors developed a deterministic algorithm to compute local RLCTs for two-dimensional models, deriving complexity bounds for the algorithm. They applied this method to polynomial neural networks and analyzed the results to understand the implications for model complexity and identifiability.
Results
The algorithm successfully computes exact RLCTs for various two-dimensional models, including polynomial neural networks. The results indicate that the effective model complexity can decrease with the number of layers, challenging conventional assumptions about model depth and identifiability.
Implications
This research has significant implications for model selection in deep learning, particularly in singular models where traditional methods fail. The exact computation of learning coefficients can improve model calibration and understanding of loss landscapes, potentially leading to more robust and interpretable machine learning models.
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
Optimization
Theory
- Formulation of localized minibatch influence as a signed, finite-horizon response under paired AdamW trajectories.
- Derivation of a joint parameter-moment ISO operator that characterizes the propagation and expression of minibatch effects.
- Establishment of an exact multistep error decomposition, proving fixed-horizon first-order accuracy.
- Experiments validate the response mechanism and reveal substantial structure in delayed influences.
Read more
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
Summary
This paper investigates the delayed effects of minibatch perturbations in the AdamW optimizer, which retains past gradient information in its states. The authors formulate AdamW as a finite-horizon input-state-output (ISO) system, allowing them to analyze how a localized gradient perturbation influences future loss outcomes. By comparing paired training trajectories that differ by a single gradient update, they isolate the evolution of this perturbation over time. The study derives a signed response operator that captures the dynamics of how optimizer memory affects the magnitude, timing, and sign of the perturbation's impact. Additionally, the authors establish a multistep error decomposition that accounts for nonlinear dynamics and activation changes, proving first-order accuracy under certain conditions. Experiments validate the proposed response mechanism and demonstrate the significant prospective structure of delayed influences, which can be partially approximated using ISO models. This work contributes to a deeper understanding of the temporal dynamics in adaptive optimization and the implications of minibatch ordering on training stability.
Methodology
The authors model the AdamW optimizer as a finite-horizon input-state-output system, analyzing the effects of localized gradient perturbations through paired training trajectories. They derive a tangent model to map initial perturbations to future loss responses and establish a multistep error decomposition to characterize the dynamics involved.
Results
The study demonstrates that the effects of a minibatch perturbation can persist beyond the immediate update, with significant implications for future loss. The derived response operator reveals how optimizer memory influences the timing and magnitude of these effects. Experimental results confirm the validity of the proposed mechanisms and highlight the structured nature of delayed influences.
Implications
This research provides insights into the dynamics of adaptive optimizers like AdamW, potentially informing strategies for improving training stability and efficiency in neural networks. Understanding the temporal dependencies of minibatch effects may lead to better optimization techniques and enhanced performance in various machine learning applications.
Ask Self, Ask Others: Relation Is All You Need
NLP
Large Language Models
Efficient ML
- Introduction of Self–Exchange Relation (SER) and Multi-Head Relation (MHR) as new token mixing operators.
- Full Relation outperforms traditional Multi-Head Attention across various model sizes in terms of validation NLL.
- FlashRelation significantly improves execution speed, making it a practical alternative for real-time applications.
- Hybrid Relation effectively combines Full and Linear layers to optimize language modeling quality.
Read more
Ask Self, Ask Others: Relation Is All You Need
Summary
This paper introduces a novel token-mixing primitive called Relation, which organizes pairwise evidence into explicit Self and Exchange relations, allowing for a more structured approach to information flow in language models. The authors propose several variations of the Relation operator, including Full Relation, FlashRelation, Linear Relation, Hybrid Relation, and a KV-style Relation Cache. The key innovation is the reversal of the traditional attention hierarchy, where flow follows relation rather than the other way around. The authors evaluate the performance of these methods across decoder-only models with varying parameter sizes (10M, 30M, and 100M) and demonstrate that Full Relation consistently achieves lower final validation negative log-likelihood (NLL) compared to Multi-Head Attention (MHA) at all scales. FlashRelation is highlighted for its efficiency, being 3.60–4.41 times faster than the Full Relation implementation, while Hybrid Relation combines the strengths of Full and Linear Relations to maintain high language-modeling quality. The findings support a relation-first perspective on token mixing, emphasizing the importance of structured relational organization in enhancing model performance.
Methodology
The authors developed the Self–Exchange Relation (SER) operator to structure token interactions into Self and Exchange categories. They extended this to Multi-Head Relation (MHR) and evaluated various configurations (Full, Flash, Linear, Hybrid Relations) across decoder-only models with different parameter sizes. Performance was assessed using validation negative log-likelihood (NLL) and execution speed metrics.
Results
Full Relation achieved lower final validation NLL than MHA across all tested model sizes (10M, 30M, 100M). FlashRelation demonstrated a speed increase of 3.60–4.41 times compared to Full Relation, while Hybrid Relation maintained high performance with 75% Linear Relation layers.
Implications
The findings suggest that adopting a relation-first approach in token mixing can lead to more efficient and effective language models. This could have significant implications for the development of faster and more accurate NLP applications, particularly in real-time processing scenarios.
Triangular Fuzzy Rescaling Distance
Theory
- Introduction of the Triangular Fuzzy Rescaling Distance (dTR) for comparing TFNs.
- Integration of Linear Rescaling directly into the distance calculation to ensure normalization.
- Formal proof that dTR satisfies all properties of a metric.
- Demonstration of dTR's applicability in heterogeneous fuzzy data contexts.
Read more
Triangular Fuzzy Rescaling Distance
Summary
This paper addresses the challenge of measuring distances between Triangular Fuzzy Numbers (TFNs) in decision-making processes that involve imprecise or uncertain information. The authors propose a novel metric called the Triangular Fuzzy Rescaling Distance (dTR), which integrates Linear Rescaling (LRE) directly into the distance calculation. This approach ensures that normalization occurs during the comparison of fuzzy numbers, thus mitigating biases introduced by differing scales or units. The paper rigorously proves that dTR satisfies the properties of a metric, including non-negativity, identity, symmetry, and the triangle inequality, while also being bounded, scale-invariant, and origin-invariant. The proposed metric is particularly suitable for applications involving heterogeneous fuzzy data, such as synthetic indicators and distance-based machine learning algorithms. The paper includes mathematical definitions, properties of the proposed distance, and numerical examples to illustrate its application.
Methodology
The authors define the Triangular Fuzzy Rescaling Distance (dTR) mathematically, proving its properties as a metric. They also provide a numerical example and a case study to demonstrate the calculation and application of dTR in real-world scenarios.
Results
The paper successfully establishes that dTR is a valid metric for TFNs, showing that it is bounded, scale-invariant, and origin-invariant. The numerical examples illustrate the effectiveness of dTR in handling heterogeneous fuzzy data.
Implications
The proposed dTR metric can significantly enhance decision-making processes in complex systems by providing a reliable method for comparing imprecise information represented as TFNs. Its applications extend to synthetic indicator construction, multicriteria decision-making, and various machine learning algorithms.
In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
Theory
Generative Models
Robotics
- Introduction of the 4MAS architecture that mimics biological hemispheric specialization for continual learning.
- Implementation of a wake-sleep training cycle to enhance memory consolidation and reduce representational drift.
- Demonstration of competitive performance on Split-MNIST, Split-Fashion-MNIST, and Split-CIFAR-100 datasets.
- Focus on biological mechanisms such as experience replay and REM sleep to inform machine learning strategies.
Read more
In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
Summary
This paper addresses the challenge of catastrophic forgetting in machine learning, particularly in the context of continual learning. The authors propose a novel architecture called 4MAS (4 Module Awake/Sleep), which mimics biological learning processes by incorporating features such as experience replay, REM sleep, and hemispheric specialization. The architecture consists of two asymmetric hemispheres, each with distinct long- and short-term memory mechanisms, allowing for effective learning and memory consolidation. The proposed model undergoes a wake-sleep training cycle, where learning occurs during the awake phase and memory consolidation happens during a simulated sleep phase. The results demonstrate that 4MAS achieves competitive accuracies on various datasets while minimizing representational drift, showcasing its potential to overcome the limitations of traditional continual learning methods.
Methodology
The authors developed the 4MAS architecture, which includes two hemispheres with specialized memory modules. The model alternates between an awake phase for task learning and a sleep phase for cross-hemispheric memory consolidation. This approach leverages biological principles of memory consolidation to mitigate catastrophic forgetting.
Results
The 4MAS architecture achieved accuracies of 98.3% on Split-MNIST, 84.9% on Split-Fashion-MNIST, and 29.29% on Split-CIFAR-100, demonstrating its effectiveness in continual learning scenarios while exhibiting low representational drift across tasks.
Implications
The findings suggest that incorporating biological principles into machine learning architectures can significantly enhance the ability to learn continuously without forgetting previous knowledge. This has potential applications in areas requiring adaptive learning systems, such as robotics, autonomous systems, and personalized AI.
Separating Covariate Shift from Mechanism Change with Two Discriminators: CJSD, a Conditional Discrepancy with an Exact Covariate-Concept Decomposition
Theory
- CJSD effectively separates covariate shift from mechanism change using two discriminators.
- The method provides a clear decomposition of task discrepancy into covariate and functional axes.
- CJSD achieves superior performance in empirical evaluations compared to existing measures.
- The paper establishes several theoretical properties that enhance the reliability of CJSD as a decision-making tool.
Read more
Separating Covariate Shift from Mechanism Change with Two Discriminators: CJSD, a Conditional Discrepancy with an Exact Covariate-Concept Decomposition
Summary
This paper introduces the Conditional Jensen–Shannon Discrepancy (CJSD), a novel method for distinguishing between covariate shift and changes in the underlying mechanism of supervised learning problems. The author highlights the importance of understanding whether two datasets share the same input-output mechanism, which is crucial for continual learning, drift diagnosis, and category discovery. Existing methods for task comparison either focus on input distribution distances, which overlook changes in the conditional distribution of labels given inputs, or exchange-based scores that conflate covariate shift with mechanism change due to extrapolation errors. CJSD utilizes two discriminators to estimate the difference in held-out cross-entropies, effectively separating the total task discrepancy into covariate and functional axes. The paper proves several properties of CJSD, including a covariate-null theorem and a drift-mass law, and demonstrates its effectiveness through empirical evaluations across various datasets. CJSD outperforms other measures in separating concept from covariate shift, achieving an AUC of 1.0 in multiple scenarios, while also providing practical applications in synthetic data fidelity auditing and fairness assessments.
Methodology
The methodology involves using two discriminators to estimate the conditional discrepancies between datasets, specifically calculating the difference in held-out cross-entropies to derive the CJSD. This approach avoids the need for task-specific predictors or generative models, relying instead on mutual information principles to split the total task discrepancy into covariate and functional components.
Results
Empirical results demonstrate that CJSD and a neighborhood plug-in conditional mutual information estimator are the only methods that cleanly separate concept from covariate shift, achieving an AUC of 1.0 across various datasets. CJSD maintains robustness even under controlled dimensionality scaling, outperforming other methods that fail under higher dimensions.
Implications
The findings suggest that CJSD can be a valuable tool for online systems that need to assess the consistency of input-output mechanisms in real-time. Its applications extend to continual learning, drift diagnosis, and fairness auditing, making it relevant for various domains that require reliable model adaptation and evaluation.
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
NLP
Interpretability
- BERT-LER integrates laboratory data representation with explainability in a single framework.
- The model achieves competitive performance on both benchmark and real-world clinical tasks.
- Token-level attributions using Integrated Gradients provide clinically meaningful insights.
- The approach demonstrates the feasibility of combining strong predictive performance with interpretability.
Read more
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
Summary
This paper presents BERT-LER, a BERT-style model designed for clinical prediction tasks using structured electronic health records (EHRs). The model addresses the challenge of integrating quantitative laboratory data into transformer architectures while maintaining interpretability. BERT-LER encodes laboratory test results as discrete tokens through percentile-based binning, allowing it to retain graded information. The model is pretrained and fine-tuned on a dataset of 75 million patients and evaluated on the EHRShot benchmark suite as well as a real-world asthma severity progression study. The authors demonstrate that BERT-LER achieves competitive predictive performance, often exceeding existing benchmark models, particularly in laboratory-related tasks. Additionally, the use of Integrated Gradients for token-level attributions provides clinically relevant explanations that align with known risk factors. This work bridges the gap between predictive performance and interpretability in EHR modeling, offering insights applicable across various therapeutic areas and prediction tasks.
Methodology
The authors developed BERT-LER, a transformer-based model that encodes laboratory test results as discrete tokens using percentile-based binning. The model is pretrained on a large EHR dataset and fine-tuned for specific clinical prediction tasks. Integrated Gradients are employed to generate token-level attributions that link model predictions to input medical events, enhancing interpretability.
Results
BERT-LER demonstrated competitive predictive performance on the EHRShot benchmark suite and outperformed existing models in laboratory-related tasks. The model's attributions aligned with clinically known risk factors, indicating its potential for practical clinical applications.
Implications
The findings suggest that BERT-LER can be utilized in various therapeutic areas for clinical prediction tasks, facilitating the integration of machine learning models into clinical workflows. The model's ability to provide interpretable predictions may enhance clinical decision-making and validation processes.
DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule
Optimization
Efficient ML
Theory
- DELTAMOMENTUM introduces direction-awareness in momentum updates, adapting forgetting rates based on frequency of direction queries.
- The method is proven to be a valid momentum update that tracks gradients effectively without matrix inversion.
- It outperforms traditional EMA momentum in clearing stale directions, enhancing training efficiency.
- DeltaAdamW achieves lower validation loss in fewer training steps compared to AdamW across multiple model sizes.
Read more
DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule
Summary
The paper introduces DELTAMOMENTUM, a novel momentum update method that incorporates direction-awareness into the momentum update rule of optimizers. Traditional momentum methods use an exponential moving average (EMA) of past gradients, which can be inefficient due to the anisotropic nature of training data. DELTAMOMENTUM leverages the key-value structure inherent in gradient updates, where the input acts as a key and the output-side error as a value. This allows for a more adaptive forgetting rate based on how frequently each direction is queried during training. The authors prove that DELTAMOMENTUM is a valid momentum update that effectively tracks gradients without requiring matrix inversions, and it clears stale directions faster than EMA. The method is designed to be a drop-in replacement for existing momentum buffers in optimizers, maintaining compatibility with other optimization techniques. Experimental results demonstrate that AdamW with DELTAMOMENTUM (DeltaAdamW) achieves significantly lower validation loss in fewer steps compared to traditional AdamW across various model sizes and tasks, confirming its effectiveness in improving training efficiency.
Methodology
The authors propose DELTAMOMENTUM, which utilizes a key-value structure in gradient updates to adaptively manage momentum based on the frequency of direction queries. The update rule is derived from the classical delta rule, allowing for efficient momentum updates without the need for matrix inversions. The method is validated through theoretical proofs and empirical testing on language model pretraining tasks.
Results
In experiments, DeltaAdamW (AdamW with DELTAMOMENTUM) achieved validation loss in up to 46.39% fewer steps at 67M parameters and 22.12% fewer steps at 370M parameters compared to standard AdamW. The performance improvements were consistent across different model sizes, including a 1B parameter model, and also held for other optimizers like SGD and architectures such as ResNet-18 and ViT-Tiny on CIFAR-10.
Implications
The introduction of DELTAMOMENTUM could significantly enhance the efficiency of training large neural networks by reducing the number of steps required to reach optimal performance. Its compatibility with existing optimizers allows for easy integration into current training workflows, potentially leading to faster convergence and reduced computational costs in large-scale AI applications.
Quantifying Event Impacts on Time Series via Multiscale Contrastive Learning
Time Series
- Introduces EventTime, a framework for estimating abnormal financial losses due to cybersecurity events.
- Utilizes a multi-resolution approach to capture various temporal scales of market dynamics.
- Implements an event fusion module to integrate event metadata with temporal representations.
- Employs a dynamic contrastive learning objective to address sparse supervision challenges.
Read more
Quantifying Event Impacts on Time Series via Multiscale Contrastive Learning
Summary
This paper addresses the challenge of quantifying the impacts of cybersecurity events on financial time series, particularly focusing on estimating short-term abnormal losses following such events. The authors introduce a novel framework called EventTime, which integrates multi-resolution market context, pre-event dynamics, and event metadata to enhance prediction accuracy. A key innovation is the event fusion module that couples temporal representations with event attributes to identify relevant market patterns. To tackle the issue of sparse supervision due to the rarity of high-impact events, the framework employs a dynamic contrastive objective that generates event- and time-series-aware positive and negative pairs during training. The authors also present SECURE, a dataset that aligns cybersecurity incidents with stock-market time series and semantic features derived from structured and large language models. Experimental results demonstrate that EventTime outperforms existing state-of-the-art models in estimating post-event financial losses, showing improved sensitivity to events, robustness to incomplete metadata, and enhanced interpretability of market impact estimates.
Methodology
The methodology involves a multi-resolution framework that combines long-term market context, short-term pre-event dynamics, and event metadata. The event fusion module integrates temporal representations with event attributes, while the dynamic contrastive learning objective constructs positive and negative training pairs to enhance model robustness against sparse data.
Results
EventTime consistently outperformed state-of-the-art time-series and event-aware baselines in estimating post-event financial losses. The framework showed improved event sensitivity, robustness to incomplete metadata, and provided more interpretable estimates of market impacts following cybersecurity disclosures.
Implications
The findings have significant implications for investors, firms, and regulators by providing a more accurate assessment of cyber-risk and enabling better portfolio protection strategies. The model's ability to quantify the impact of cybersecurity incidents can enhance decision-making processes in financial markets.
Reliable Neural Collapse Approximation for Open-World Test-Time Adaptation
Computer Vision
Theory
Efficient ML
- Introduces the concept of neural collapse to Open-World Test-Time Adaptation (OWTTA).
- Develops the Reliable Neural Collapse approximation (ReNC) method to address label distribution shifts.
- Proposes a filtering mechanism for Out-Of-Distribution (OOD) samples based on prototype similarity.
- Demonstrates superior performance of ReNC on open-world benchmarks compared to existing methods.
Read more
Reliable Neural Collapse Approximation for Open-World Test-Time Adaptation
Summary
This paper addresses the challenges of Test-Time Adaptation (TTA) in open-world scenarios, where label distribution shifts can occur, rendering traditional TTA methods ineffective. The authors introduce a novel method called Reliable Neural Collapse approximation (ReNC) for Open-World Test-Time Adaptation (OWTTA). ReNC leverages the concept of neural collapse as a structural prior for adapting to the target domain. The method justifies that pre-trained classifier weights can serve as prototypes for the source domain and proposes a mechanism to filter out Out-Of-Distribution (OOD) samples based on their similarity to these prototypes. Additionally, ReNC includes a neural collapse approximation mechanism that refines these prototypes, allowing them to adapt to the target domain while maintaining the neural collapse structure. The authors conduct extensive experiments on various open-world benchmarks, demonstrating that ReNC outperforms existing methods by better preserving neural collapse properties in the target domain, thus providing insights for future model design.
Methodology
The methodology involves leveraging neural collapse as a structural prior for reliable target-domain adaptation. The authors propose a two-step process: first, filtering OOD samples based on their similarity to source domain prototypes, and second, refining these prototypes through a neural collapse approximation mechanism that adapts them to the target domain.
Results
The experimental results show that ReNC significantly improves the adaptation process in open-world scenarios, outperforming existing TTA methods. The method effectively preserves neural collapse properties in the target domain, leading to enhanced model performance.
Implications
The findings suggest that ReNC can be applied in various domains where label distribution shifts occur, such as medical imaging and other real-world applications. The insights gained from this study may inform future research and development in model adaptation techniques.
Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay
Reinforcement Learning
Large Language Models
Theory
- Existing step-level credit signals do not reliably identify causally significant steps in LLM agents.
- Causal contribution is sparse and model-dependent, with significant variability in measurability across policies.
- Implicit credit primarily reflects the fluency of the policy rather than its causal impact on outcomes.
- A confidence-only routing rule can reduce costs while maintaining performance at chance levels.
Read more
Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay
Summary
This paper investigates the effectiveness of step-level credit assignment signals used in training large language model (LLM) agents by auditing them against causal ground truth derived from executed replay in a single-agent tool environment (ALFWorld). The study reveals that none of the evaluated credit signals—such as LLM-judge scores, outcome-conditioned log-probability ratios, or the policy's own confidence—can reliably identify causally significant steps better than chance. The authors establish a new ground truth for causal contribution, demonstrating that it is sparse and model-dependent, with only 30.5% of decision points showing measurable effects. The findings indicate that existing evaluations based on annotated step correctness do not align with actual contributions to outcomes. The paper also identifies a failure mode where implicit credit reflects the policy's fluency rather than causal significance. A confidence-only routing rule is proposed, which reduces judge costs while maintaining chance-level pivotal recall. Overall, the research highlights the need for more robust methods to assess credit assignment in LLM training, emphasizing the importance of distinguishing between correctness and contribution.
Methodology
The authors utilize executed replay to measure causal contributions at each decision point in a trajectory. They re-sample alternative actions supported by the policy and assess how these alternatives affect the outcome distribution. This method is implemented in the ALFWorld environment, allowing for a detailed audit of step-level credit signals against a structured ground truth.
Results
The audit reveals that none of the credit signals outperform their own shuffled controls, indicating that they do not effectively capture causal contributions. The study finds that the implicit credit scores correlate strongly with the policy's fluency, and conditioning on outcomes does not provide additional causal information. In a pre-registered training experiment, no credit assignment method significantly outperformed an untrained policy, suggesting that apparent differences in performance are attributable to training dose rather than the content of the credit signals.
Implications
The findings suggest that current methods for credit assignment in LLM training may be inadequate, potentially leading to ineffective training strategies. The proposed confidence-only routing rule offers a cost-effective alternative for managing credit assignment without sacrificing performance. This research could inform future developments in reinforcement learning and LLM training methodologies, emphasizing the need for more accurate measures of causal contribution.
Multi-Source Wasserstein Distributionally Robust Graph Learning
Graph Learning
Optimization
Theory
- Introduces MS-WDRO, a framework for robust graph learning using Wasserstein metrics.
- Addresses the challenges of heterogeneous source data and limited target samples.
- Establishes non-asymptotic performance guarantees for the proposed method.
- Utilizes an efficient ADMM scheme for optimization.
Read more
Multi-Source Wasserstein Distributionally Robust Graph Learning
Summary
This paper addresses the challenge of network topology inference from graph signals, particularly in scenarios where target-domain samples are limited while heterogeneous source-domain data is abundant. The authors propose a novel framework called MS-WDRO (Multi-Source Wasserstein Distributionally Robust Optimization) that utilizes the Wasserstein metric for fusing diverse data sources into a reliable nominal distribution. This approach preserves the intrinsic geometry of each source and mitigates the issues arising from inter-source divergence. The framework constructs an ambiguity ball around the weighted Wasserstein barycenter to manage residual uncertainty, allowing for the minimization of worst-case risk. The authors provide non-asymptotic guarantees, including finite-sample concentration bounds and out-of-sample excess risk bounds. The methodology is implemented using an efficient ADMM scheme, and the hyperparameters are calibrated through an end-to-end differentiable architecture. Experimental results demonstrate that MS-WDRO outperforms seven baseline methods in terms of graph recovery accuracy, sample efficiency, and diagnostic utility, particularly in sample-scarce contexts.
Methodology
The proposed MS-WDRO framework employs the weighted Wasserstein barycenter to fuse heterogeneous data sources, creating a nominal distribution that reflects the intrinsic geometries of the sources. An ambiguity ball is constructed around this barycenter to hedge against uncertainty, and the worst-case risk is minimized using a regularized Laplacian estimator. The optimization is efficiently solved via an Alternating Direction Method of Multipliers (ADMM) approach, with hyperparameters calibrated through a differentiable architecture.
Results
The experiments conducted on synthetic benchmarks and the multi-site ABIDE I neuroimaging dataset reveal that MS-WDRO consistently outperforms seven baseline methods in terms of accuracy in graph recovery, efficiency in sample usage, and overall diagnostic utility. The method shows particularly significant improvements in scenarios where sample sizes are limited.
Implications
The findings suggest that MS-WDRO can be effectively applied in various domains requiring network topology inference, such as neuroimaging, sensor networks, and social network analysis, where data scarcity and heterogeneity are common challenges.
A Standardized Framework for Machine Learning in Power System Protection
Theory
- Introduces a standardized framework for evaluating ML in power system protection.
- Defines seven essential dimensions for consistent evaluation of ML studies.
- Demonstrates the framework's application through a case study on fault classification and localization.
- Highlights the impact of decision windows and observability on ML performance metrics.
Read more
A Standardized Framework for Machine Learning in Power System Protection
Summary
This paper addresses the challenges in evaluating machine learning (ML) approaches for power system protection, which often report high performance metrics without a standardized evaluation framework. The authors propose a comprehensive framework that outlines seven critical study dimensions necessary for consistent evaluation: protection objective, physical system scope, observability and measurements, timing and decision windows, targets and valid samples, training and validation protocol, and evaluation outputs. The framework is applied to a case study using the PROTECT-90 benchmark, which includes 9022 simulated episodes for fault classification and localization in a 90 kV double-line topology. The study demonstrates that a multi-layer perceptron (MLP) achieved a mean macro-averaged F1 score of 0.991 and a localization error of 10.20% under specific conditions. The findings indicate that while extending the decision horizon maintained performance, reduced observability significantly impacted localization accuracy. The proposed framework aims to enhance the reproducibility and comparability of ML research in power system protection, facilitating future certification and auditing processes.
Methodology
The authors developed a standardized evaluation framework and applied it to a bounded case study using the PROTECT-90 electromagnetic-transient benchmark. They utilized a multi-layer perceptron (MLP) for fault classification and localization, assessing performance under various conditions related to decision windows and observability.
Results
The MLP achieved a five-fold mean macro-averaged F1 score of 0.991 ± 0.001 for classification tasks and a mean absolute localization error of 10.20 ± 0.25% of line length. Extending the decision horizon to 50 ms preserved performance asymmetry, while reduced observability approximately doubled the localization error but had minimal effect on classification accuracy.
Implications
The proposed framework could lead to more reliable and comparable evaluations of ML-based protection systems, fostering trust and facilitating certification processes in safety-critical domains such as power systems. This could ultimately enhance the deployment of ML technologies in real-world applications.
DeltaML-Bench: Evaluating Machine Learning Agents on Real-World Research Repositories
Computer Vision
Graph Learning
Time Series
- DeltaML-Bench introduces a benchmark for evaluating ML agents on real-world tasks, focusing on improving baselines in imperfect codebases.
- The search-based ARG scaffolding significantly improves the success rates of GPT-5 in task completion.
- Specification gaming is prevalent in Modular configurations, while ARG configurations do not exhibit this issue.
- The benchmark spans diverse domains including Computer Vision, Graph Learning, and Time Series.
Read more
DeltaML-Bench: Evaluating Machine Learning Agents on Real-World Research Repositories
Summary
DeltaML-Bench is introduced as a novel benchmark designed to evaluate autonomous machine learning agents on real-world research repositories. Unlike existing benchmarks that focus on isolated tasks, DeltaML-Bench comprises 48 tasks derived from research papers, requiring agents to improve upon published baselines within imperfect, open-source codebases. The benchmark addresses the challenges of navigating heterogeneous repositories, debugging training pipelines, and conducting productive experimentation under realistic compute constraints. The authors evaluate two advanced models, GPT-5 and Claude Sonnet 4, using both a standard Modular agent and a proposed search-based ARG scaffolding. The results demonstrate that the ARG scaffolding significantly enhances GPT-5's success rate in completing tasks, achieving a per-run success rate of 49.0% under a 2×12h allocation. The study also highlights the issue of specification gaming, with Modular configurations exhibiting high rates of gaming compared to the ARG configurations, which showed no gaming. The findings underscore the importance of scaffolding design and integrity checks in the deployment of agents for autonomous machine learning experimentation.
Methodology
The authors developed DeltaML-Bench, comprising 48 tasks sourced from research papers, and evaluated two advanced models (GPT-5 and Claude Sonnet 4) using both a standard Modular agent and a search-based ARG scaffolding. They conducted experiments under different compute allocations (4×6h and 2×12h) to assess the agents' performance.
Results
In the 4×6h allocation, the ARG scaffolding raised GPT-5's per-run success rate from 9.4% to 33.9%. Under the longer 2×12h allocation, GPT-5 ARG achieved a 49.0% success rate. Modular configurations showed specification gaming rates as high as 47.9%, while no gaming was detected in ARG configurations.
Implications
DeltaML-Bench provides a valuable framework for evaluating the capabilities of machine learning agents in real-world scenarios, highlighting the importance of effective scaffolding and integrity checks. This benchmark can guide future research in autonomous ML experimentation and improve the reliability of ML agents in practical applications.
Continuous Adversarial MeanFlow Transfer
Generative Models
Computer Vision
Efficient ML
- Introduction of MeanFlow-Transfer (MF-T) for efficient adaptation of pretrained models.
- Development of Continuous Adversarial MeanFlow (CAMF) to enhance few-step generation quality.
- Demonstrated performance improvements in FID and FDD metrics with reduced NFEs.
- Unified adaptation and acceleration processes for diverse pretrained model parameterizations.
Read more
Continuous Adversarial MeanFlow Transfer
Summary
This paper addresses the challenges of training fast generators on new domains with limited data, specifically focusing on the adaptation of pretrained diffusion or flow models. The authors propose MeanFlow-Transfer (MF-T), which maps heterogeneous source outputs into a shared velocity representation, allowing for the initialization of a MeanFlow (MF) generator from source weights and optimizing an MF objective on the target domain. This approach unifies adaptation and acceleration in a single training loop across various pretrained models. Additionally, the authors introduce Continuous Adversarial MeanFlow (CAMF), a post-training stage that extends continuous adversarial flow models to accommodate finite-interval average velocities, contrasting changes in a learned potential between real and predicted interval endpoints. The proposed methods were evaluated by adapting four ImageNet-based source models to five target domains, demonstrating that MF-T with CAMF matches or exceeds the performance of fine-tuned models in terms of Fréchet Inception Distance (FID) and Fréchet Distance Divergence (FDD) while requiring significantly fewer Neural Function Evaluations (NFEs).
Methodology
The authors developed MF-T to map outputs from various pretrained models into a shared instantaneous-velocity representation, allowing for the initialization of an MF generator. CAMF was introduced as a post-training refinement stage that adapts continuous adversarial learning to the finite-interval average velocities predicted by MF models. The methods were evaluated using a range of pretrained models adapted to multiple target domains.
Results
MF-T combined with CAMF achieved performance that matched or exceeded fine-tuned models in FID and FDD metrics, with up to 125 times fewer Neural Function Evaluations required. CAMF alone improved MF-T's few-step FID by an average of 29%.
Implications
The proposed methods could significantly enhance the efficiency and quality of image synthesis in applications requiring fast generation from limited data, making them valuable for real-time interactive systems and scenarios where computational resources are constrained.
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
Generative Models
Theory
Efficient ML
- Investigates flow-matching-based anomaly detection under contaminated training data.
- Highlights the significance of anomaly scoring functions for robustness.
- Demonstrates that trajectory-based scores improve detection performance.
- Shows that Forest-Flow can outperform TCCM in certain scenarios.
Read more
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
Summary
This paper addresses the challenge of unsupervised anomaly detection in tabular data, particularly in financial contexts where training datasets often contain unlabeled anomalies. The authors explore the robustness of flow-matching-based anomaly detection methods, specifically Time-Conditioned Contraction Matching (TCCM) and Forest-Flow, under contaminated training sets. They highlight the critical role of anomaly scoring functions in determining the performance of these methods. The study reveals that TCCM's original Decision score is sensitive to contamination, while trajectory-based scores such as Deviation and Reconstruction provide more stable anomaly signals. The authors conduct experiments on two financial datasets with varying anomaly ratios, demonstrating that Forest-Flow, when paired with suitable scoring functions, can match or outperform TCCM. This work emphasizes the importance of robust scoring strategies in enhancing anomaly detection performance in real-world applications where data contamination is prevalent.
Methodology
The authors evaluate two flow-matching-based models: TCCM, which learns a contraction field for anomaly detection, and Forest-Flow, a generative model for tabular data. They assess three anomaly scoring functions—Decision, Deviation, and Reconstruction—across two financial datasets with different anomaly ratios. The study focuses on how these scoring functions impact the models' robustness under contaminated training conditions.
Results
The results indicate that TCCM's Decision score is highly sensitive to contamination, while the Deviation and Reconstruction scores provide more stable anomaly signals. The experiments show that when using these trajectory-based scores, Forest-Flow becomes competitive with TCCM and, in some cases, outperforms it, particularly in settings with severe class imbalance.
Implications
The findings suggest that careful selection of anomaly scoring functions is crucial for effective anomaly detection in financial applications, where data contamination is common. This research can inform the development of more robust anomaly detection systems that can operate effectively in real-world scenarios.
Time-Uniform Self-Normalized Concentration for Discounted Least Squares: Limits and Corrections
Theory
Reinforcement Learning
- Identifies a counterexample that disproves the claimed time-uniform guarantee for discounted least-squares estimators.
- Establishes a lower bound for valid anytime boundaries in non-stationary problems.
- Clarifies the proof error related to the use of different Gaussian mixing distributions at different terminal times.
- Demonstrates that the weighted inequality remains valid at fixed deterministic times.
Read more
Time-Uniform Self-Normalized Concentration for Discounted Least Squares: Limits and Corrections
Summary
This paper addresses the limitations of self-normalized concentration inequalities in the context of discounted least-squares estimators used in non-stationary problems within bandit and reinforcement learning frameworks. The author presents a counterexample demonstrating that the claimed time-uniform guarantee for these estimators is incorrect, as it can exceed the bounded radius with probability one. The analysis reveals that for fixed discount and regularization parameters, a valid anytime boundary must grow at least as fast as R√log(T/δ) for sufficiently large T/δ when δ ≤ 1/2. The paper identifies a proof error in the existing literature, where different terminal times were incorrectly assumed to use the same Gaussian mixing distributions. The author clarifies that while the time-uniform claim is invalid, the weighted inequality remains valid at fixed deterministic times, and provides corrections for finite and infinite-horizon scenarios. This work has significant implications for the analysis of discounted estimators in sequential decision-making, emphasizing the need for careful consideration of time-uniform validity in theoretical guarantees.
Methodology
The author employs a scalar Gaussian counterexample to illustrate the failure of the claimed time-uniform inequality. The analysis involves deriving lower bounds for valid anytime boundaries and revisiting the proof of classical self-normalized inequalities. The paper also discusses the implications of the identified proof error and provides valid corrections for both finite and infinite-horizon scenarios.
Results
The main results indicate that the claimed time-uniform self-normalized concentration inequality is incorrect, as it can exceed the bounded radius with probability one. The author establishes that any valid anytime boundary must grow at least as fast as R√log(T/δ) under certain conditions. Additionally, the paper clarifies that the weighted inequality is valid at fixed deterministic times and provides corrections for the analysis of discounted estimators.
Implications
The findings have significant implications for the theoretical analysis of discounted least-squares estimators in bandit and reinforcement learning contexts. The corrections and clarifications provided in this paper can enhance the reliability of confidence bounds in non-stationary problems, leading to improved decision-making strategies in sequential learning tasks.
Towards On-Board Implementation of ML-Based Helicopter Weight Estimator
Time Series
Robotics
Efficient ML
- Development of a supervised ML model for helicopter weight estimation.
- Alignment with EASA guidelines and Eurocae ED-324 standards for safety assurance.
- Implementation verified on legacy avionics computers for airborne deployment.
- Accurate weight estimation enhances safety and reduces operational costs.
Read more
Towards On-Board Implementation of ML-Based Helicopter Weight Estimator
Summary
This paper presents the development and implementation of a supervised Machine Learning (ML) model aimed at estimating helicopter weight during takeoff, utilizing extensive datasets from Airbus's global fleet. The authors detail a learning assurance process that aligns with the European Union Aviation Safety Agency (EASA) guidelines and the ongoing Eurocae ED-324 standards. They propose specific Machine Learning Requirements (MLCR) and a Machine Learning Model Description (MLMD) for a Long Short-Term Memory (LSTM) recurrent neural network. The implementation is verified on legacy avionics computers, demonstrating its suitability for deployment in airborne systems for critical functions such as on-board alerting. The paper emphasizes the importance of accurate weight estimation for enhancing safety and reducing operational costs, particularly in Condition-Based Maintenance (CBM) frameworks. The authors also address the challenges of ensuring compliance with safety regulations while integrating ML models into aviation systems, highlighting the need for a robust verification process.
Methodology
The authors employed a W-shape development process to create the ML model, specifying MLCR and Data Processing Requirements (DPR) in accordance with EASA and ED-324 guidelines. They utilized a Long Short-Term Memory (LSTM) architecture for the model and implemented it using Scade modeling tools to generate C code. The verification process included statistical testing and high-level tests to ensure compliance with safety requirements.
Results
The implementation of the ML-based weight estimator was successfully demonstrated on legacy avionics systems, showing that the model can operate effectively within the constraints of existing hardware. The verification process confirmed that the model meets the specified high-level requirements and performs reliably in the target environment.
Implications
The successful implementation of this ML-based weight estimator could lead to improved safety in helicopter operations by providing more accurate weight estimations, which are critical for flight safety and operational efficiency. Additionally, it opens avenues for further integration of ML technologies in aviation systems, potentially enhancing various aspects of flight operations and maintenance.
A Two-Stage Time-Aware Transformer for Short-Horizon AECOPD Risk Prediction
Time Series
- Introduces a two-stage model for AECOPD risk prediction using raw ventilator waveforms.
- First stage classifies high-risk patients, while the second stage estimates time to event.
- Achieves superior performance compared to traditional models with F1 = 0.91 and RMSE = 1.00 days.
- Preserves temporal dynamics of respiratory data, enhancing prediction accuracy.
Read more
A Two-Stage Time-Aware Transformer for Short-Horizon AECOPD Risk Prediction
Summary
This paper addresses the critical need for timely prediction of acute exacerbations of chronic obstructive pulmonary disease (AECOPD) using a novel two-stage framework that leverages raw pressure and flow waveforms from home ventilators. Traditional machine learning approaches often rely on episodically collected clinical data, which can introduce delays in risk detection. The proposed model operates directly on the continuous respiratory data collected over a seven-day period, allowing for both high-risk classification and time-to-event estimation. The first stage of the model identifies patients at high risk of severe exacerbation, while the second stage estimates the number of days until the event occurs. The authors demonstrate that their two-stage model significantly outperforms traditional baseline models, achieving an F1 score of 0.91 for risk classification and a root mean square error (RMSE) of 1.00 days for time-to-event estimation. This approach not only provides early warnings but also actionable lead time for clinical interventions, thereby enhancing the management of AECOPD in home monitoring settings.
Methodology
The methodology involves a two-stage framework where the first stage uses a Time-Aware Transformer encoder to classify patients at high risk of AECOPD based on raw pressure and flow waveforms. The second stage employs a regression model to estimate the time remaining until a severe exacerbation for those identified as high risk. This approach maintains the temporal structure of the data, allowing for effective short-horizon predictions.
Results
The experimental results indicate that the two-stage model outperforms traditional baseline models in both risk classification and time-to-event estimation. The selected Stage 1 classifier achieved an F1 score of 0.91, while the Stage 2 regression model demonstrated an RMSE of 1.00 days and an R2 value of 0.76 on the held-out test set.
Implications
The findings suggest that this two-stage model could significantly improve the management of AECOPD by providing timely risk assessments and actionable insights for clinicians. This could lead to better patient outcomes through early interventions and more effective home monitoring strategies.
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Generative Models
Optimization
- FAR-DPO improves cyclic peptide design by addressing geometric and biophysical constraints.
- The framework utilizes feasibility-aware preference construction and difficulty-aware optimization.
- Significant increases in success rates for feasible designs were achieved on benchmark datasets.
- The method is applicable across different generative architectures, enhancing its versatility.
Read more
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
Summary
Cyclic peptides are gaining attention in drug discovery due to their high binding affinity and structural stability. However, the design of cyclic peptides poses significant challenges due to geometric and biophysical constraints that limit the feasible design space. Existing generative models often struggle with low yields of feasible designs and limited control over multi-objective trade-offs. To address these issues, the authors propose FAR-DPO (Feasibility-Aware and Robust Direct Preference Optimization), an architecture-agnostic framework that enhances the design process by steering generative models towards feasible cyclic peptide designs. FAR-DPO integrates feasibility-aware preference construction with difficulty-aware group-robust optimization. It constructs preference pairs through feasibility-gated multi-objective dominance and adaptively reweights difficulty groups based on current preference losses. Evaluations on the CPSea LNR benchmark demonstrate that FAR-DPO significantly improves the success rate of feasible designs, increasing it from 46.89% to 57.79% for PepGLAD and from 47.96% to 49.57% for PepFlow under a fixed generation budget. These improvements extend to the hardest target quartile and yield better binding scores, showcasing FAR-DPO's effectiveness in enhancing feasibility and robustness in cyclic peptide design.
Methodology
FAR-DPO employs a two-step approach: first, it constructs feasibility-aware preference pairs from generated cyclic peptides using multi-objective dominance; second, it implements difficulty-aware group-robust optimization to adjust the optimization weights of preference groups based on their performance, ensuring a balanced focus on challenging targets.
Results
FAR-DPO increased the overall success rate of feasible cyclic peptide designs from 46.89% to 57.79% for PepGLAD and from 47.96% to 49.57% for PepFlow. These improvements were consistent across challenging targets and resulted in better binding scores, demonstrating the framework's effectiveness in enhancing design feasibility and robustness.
Implications
The advancements presented by FAR-DPO could lead to more efficient and effective cyclic peptide designs, potentially accelerating drug discovery processes and improving therapeutic outcomes. The framework's architecture-agnostic nature allows for broader applications in peptide design across various generative models.
DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
Time Series
Generative Models
Interpretability
- DecoVAE explicitly separates trend and seasonal components for improved forecasting accuracy.
- The framework employs a Hodrick-Prescott-inspired regularization for trend smoothness.
- Seasonality is modeled in the complex frequency domain, enhancing periodic pattern learning.
- DecoVAE shows significant performance improvements over strong baselines across multiple benchmarks.
Read more
DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
Summary
The paper presents DecoVAE, a novel framework for probabilistic time series forecasting that effectively decomposes time series data into trend and seasonal components. Traditional forecasting methods often struggle with accurately modeling these distinct dynamics, leading to inefficiencies and a lack of interpretability. DecoVAE addresses these challenges by employing domain-specific inductive biases to enforce structural smoothness in the trend component and utilizing a complex Gaussian VAE for the seasonal component, allowing for the natural capture of periodic patterns. The framework integrates these components into a variational autoencoder, ensuring well-calibrated probabilistic forecasts. Extensive evaluations on seven real-world datasets demonstrate that DecoVAE outperforms existing baselines, achieving significant reductions in Continuous Ranked Probability Score (CRPS) and Normalized Mean Absolute Error (NMAE) for both short-term and long-term forecasting, while also being highly efficient in terms of model size and computational speed.
Methodology
DecoVAE utilizes a variational autoencoder architecture that incorporates specialized subnetworks for trend and seasonal decomposition. The trend component is regularized using a differential approach inspired by the Hodrick-Prescott filter, while the seasonal component is modeled in the complex domain to effectively capture amplitude and phase information. This dual approach allows for structured temporal dynamics and well-calibrated probabilistic forecasts.
Results
DecoVAE achieves up to 14.96% improvement in CRPS and 23.30% in NMAE for short-term forecasting, and up to 52.68% and 26.51% improvements for long-term forecasting. Additionally, it reduces model weight by up to 93% and accelerates processing speed by up to 74% compared to the second-best method.
Implications
The DecoVAE framework has significant implications for various fields requiring accurate time series forecasting, such as finance, energy, and weather prediction. Its lightweight and interpretable nature makes it suitable for real-time applications where computational efficiency and accuracy are critical.
Kähler landscapes for complex neural network descents and guarantees including a search and destroy of the Calabi-Yau manifold
Theory
Optimization
- Introduces Kähler information metrics for complex neural network optimization.
- Explores the implications of Calabi-Yau manifolds on optimization landscapes.
- Demonstrates the impact of negative curvature on loss landscapes.
- Establishes theoretical results related to initialization and convexity in complex networks.
Read more
Kähler landscapes for complex neural network descents and guarantees including a search and destroy of the Calabi-Yau manifold
Summary
This paper investigates the optimization landscapes of complex-parameterized neural networks through a geometric lens, specifically utilizing Kähler manifolds and information geometry. The author introduces a descent path governed by a Kähler information metric derived from cross-entropy and the Wirtinger Hessian on the log-likelihood potential. The study emphasizes the role of Calabi-Yau manifolds, which present theoretical guarantees but also exhibit ill-curvature-conditioned landscapes that complicate optimization. The paper discusses the implications of negative curvature on the loss landscape, particularly focusing on sectional and Ricci curvature. Theoretical results are presented, including second derivative results, initialization conditions, and convexity arguments, all framed within a geometric analytic context. The findings suggest that complex neural networks can achieve performance improvements over their real counterparts, but their optimization landscapes can be significantly affected by the underlying geometric properties.
Methodology
The methodology involves a geometric analysis of complex neural networks using Kähler metrics and information geometry. The author employs natural gradient descent techniques and examines the effects of curvature on optimization landscapes, particularly focusing on Calabi-Yau metrics and their implications for theoretical guarantees.
Results
The paper presents several theoretical results, including conditions for convexity, the behavior of the loss landscape under negative curvature, and the implications of Ricci curvature on optimization trajectories. It establishes that negative curvature can complicate descent paths and that certain geometric conditions can lead to performance gains in complex neural networks.
Implications
The findings have potential implications for the design and training of complex neural networks, suggesting that understanding the underlying geometric properties can lead to improved optimization strategies and performance. This work may influence future research in deep learning, particularly in the context of complex parameterizations and their optimization landscapes.
CLaST: Context-aware Contrastive VAE for Probabilistic Time Series Forecasting
Generative Models
Time Series
- CLaST introduces a context-aware contrastive loss function for improved probabilistic time series forecasting.
- The concept of Lag-Invariant Non-stationary Time Series (LINTS) is introduced to better capture temporal dynamics.
- CLaST achieves up to 16.4% improvement in CRPS and 14.4% in NMAE for short-term forecasting.
- For long-term predictions, CLaST surpasses the second-best method by up to 48.6% in CRPS and 25.1% in NMAE.
Read more
CLaST: Context-aware Contrastive VAE for Probabilistic Time Series Forecasting
Summary
The paper introduces CLaST, a novel variational autoencoder (VAE) framework designed for probabilistic multivariate time series forecasting. CLaST addresses the limitations of existing generative models in capturing internal temporal dependencies by employing a contrastive loss function that preserves contextual similarity between observations. This approach is particularly relevant for time series data, where contextual relationships can significantly impact forecasting accuracy. The authors define a new class of time series called Lag-Invariant Non-stationary Time Series (LINTS), which allows for a more realistic representation of temporal dynamics. The proposed method enhances the stability of training and the quality of learned representations by replacing conventional mutual information estimators with a contrastive-based penalty. Extensive empirical validation across nine benchmark datasets demonstrates that CLaST consistently outperforms strong baseline methods, achieving significant improvements in both short-term and long-term forecasting tasks.
Methodology
CLaST employs a variational autoencoder framework that incorporates a novel contrastive loss function to maintain contextual similarity in time series data. The authors derive theoretical properties of the loss function under the LINTS process, ensuring that the model captures lag-aware information effectively. The approach replaces traditional mutual information estimators with a more stable contrastive-based penalty, enhancing the learning of informative latent representations.
Results
CLaST demonstrates superior performance in both short-term and long-term forecasting tasks. In short-term tasks, it achieves improvements of up to 16.4% in Continuous Ranked Probability Score (CRPS) and 14.4% in Normalized Mean Absolute Error (NMAE) compared to the second-best method. For long-term predictions, CLaST exceeds the second-best results by up to 48.6% in CRPS and 25.1% in NMAE, confirming its effectiveness across various time series modalities.
Implications
The findings suggest that CLaST can be effectively applied in various domains requiring probabilistic time series forecasting, such as energy systems, finance, and healthcare. The ability to capture contextual relationships in time series data may lead to more accurate predictions and better decision-making in these fields.
Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking
Time Series
- SATS introduces a scale-aware token alignment mechanism for time series pretraining.
- The framework effectively captures multi-scale temporal structures through a hybrid masking strategy.
- SATS achieves state-of-the-art performance on LSTF benchmarks, significantly improving MSE and GIFT-Eval MASE.
- The model demonstrates a substantial increase in efficiency compared to existing methods.
Read more
Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking
Summary
This paper introduces SATS, a novel framework for pretraining time series foundation models that effectively addresses the challenges posed by heterogeneous datasets with varying sampling frequencies. Traditional methods either use dataset-specific patch sizes leading to fragmented representations or enforce a fixed patch size that overlooks temporal variations. SATS proposes a scale-aware token alignment mechanism that treats patch size as a scale, allowing for the alignment of representation spaces across different scales while maintaining distinct modeling capacities. Additionally, a hybrid masking strategy is introduced, combining random and contiguous masking to capture multi-scale temporal structures. Experimental results on LSTF benchmarks show that SATS outperforms competitive baselines, achieving a 9.2% improvement in mean squared error (MSE) and an 8.3% gain in GIFT-Eval MASE, while also increasing model efficiency by 65.6%. These findings highlight SATS's effectiveness and scalability in time series pretraining, establishing it as a strong candidate for zero-shot and in-distribution forecasting tasks.
Methodology
The SATS framework employs a scale-aware token alignment mechanism that aligns representation spaces across different patch sizes using scale-specific feedforward networks (FFNs). It minimizes the distance between mean token embeddings across scales while maximizing the distance between their maximal embeddings. Additionally, a hybrid masking strategy is utilized, combining random and contiguous masking to enhance the model's ability to capture both fine-grained and long-range temporal dependencies.
Results
SATS achieved a 9.2% improvement in mean squared error (MSE) and an 8.3% gain in GIFT-Eval MASE on LSTF benchmarks compared to competitive baselines. Furthermore, it demonstrated a 65.6% increase in model efficiency, underscoring its effectiveness and scalability.
Implications
The proposed SATS framework has significant implications for the development of time series foundation models, particularly in enhancing generalization across diverse datasets and improving model efficiency. Its ability to effectively handle varying temporal structures makes it suitable for applications in finance, healthcare, meteorology, and IoT, where time series data is prevalent.
Inadvertent Context Leakage in Language Models
NLP
Large Language Models
Theory
- Inadvertent context leakage occurs even when models refuse direct extraction of sensitive information.
- An adaptive adversary can reconstruct numeric secrets with high accuracy using benign prompts.
- More capable models exhibit greater leakage, indicating a correlation between model capability and privacy risks.
- The study introduces a suppression mechanism to mitigate leakage and evaluates its effectiveness across different models.
Read more
Inadvertent Context Leakage in Language Models
Summary
This paper investigates the phenomenon of inadvertent context leakage in large language models (LLMs), particularly in scenarios where sensitive user data is present in the model's context window. The authors explore whether the presence of such sensitive information can lead to hidden correlations in the model's outputs, allowing for the reconstruction of secrets even when the model refuses direct extraction requests. They introduce a novel adaptive attack methodology that assumes black-box access to the model, enabling adversaries to engineer prompts that exploit this leakage. Through controlled experiments on eight proprietary models, the authors demonstrate that 2-digit secrets can be reconstructed with near-perfect accuracy, while 4-digit secrets achieve an 82% exact match. The study reveals that more capable models tend to leak more information, suggesting that leakage is a byproduct of model capability rather than a fixable flaw. The paper also discusses practical attacks, including inferring semantic predicates about user memories and extracting sensitive information like Social Security Numbers from benign outputs.
Methodology
The authors formalize a threat model for inadvertent context leakage as a predicate inference game. They conduct controlled experiments on eight proprietary models, using benign prompts to assess the models' ability to leak sensitive information. An adaptive adversary is employed to learn and exploit model-specific channels that carry in-context secrets into benign outputs.
Results
The experiments reveal that 2-digit secrets can be reconstructed with nearly 100% accuracy, while 4-digit secrets achieve an 82% exact match. The study also finds that conditioning prompts on previously recovered digits significantly improves recovery rates. Additionally, the authors demonstrate that an adversary can successfully extract Social Security Numbers from production-style agents with high accuracy.
Implications
The findings highlight significant privacy concerns for LLMs used in sensitive applications, suggesting that even benign outputs can inadvertently leak sensitive information. This research underscores the need for improved privacy-preserving mechanisms in AI systems that handle personal data.
Evaluating Neural Cartographic Relief Shading for Urban Environments: A Downtown Calgary Study Using High-Resolution DEM and DSM Data
Computer Vision
- Comparison of analytical and neural-based hillshading methods in urban environments.
- Investigation of parameter tuning in Eduard for urban relief shading.
- Identification of strengths and weaknesses of neural shading in urban contexts.
- Recommendations for future neural model training specifically for urban morphology.
Read more
Evaluating Neural Cartographic Relief Shading for Urban Environments: A Downtown Calgary Study Using High-Resolution DEM and DSM Data
Summary
This paper investigates the effectiveness of analytical and neural-based hillshading techniques in urban settings, specifically focusing on downtown Calgary using high-resolution digital elevation model (DEM) and digital surface model (DSM) data. The study compares traditional single-direction and multi-direction analytical hillshading methods with neural relief shading generated by Eduard, a machine-learning system originally designed for mountainous landscapes. The central question is whether parameter tuning in Eduard can yield visually compelling and cartographically useful results in urban environments, despite its training bias towards natural landforms. The research emphasizes the importance of terrain type, generalization, and detail parameters while maintaining a consistent large-scale shading style. The findings reveal both the strengths and limitations of Eduard in urban contexts, suggesting that while analytical methods remain reliable, neural approaches can produce unexpectedly strong results when appropriately tuned. The study concludes with recommendations for future model training focused on urban morphology, highlighting the need for tailored neural approaches in urban cartography.
Methodology
The study employs a comparative cartographic experiment, utilizing both analytical and neural-based hillshading methods. It evaluates the visual legibility and cartographic effectiveness of each method rather than relying solely on algorithmic accuracy metrics. The analysis includes varying terrain types, micro and macro generalization, and flat-area detail settings while keeping a consistent large-scale shading style.
Results
The results indicate that while traditional analytical hillshading methods are reliable, the neural-based Eduard system can produce visually compelling urban shadings when parameter tuning is applied. The study highlights specific scenarios where Eduard outperforms conventional methods and identifies failure modes due to its training bias towards alpine terrains.
Implications
The findings suggest that neural approaches can be adapted for urban cartography, potentially leading to improved visualization techniques for urban environments. This research advocates for future developments in neural model training that specifically address the complexities of urban morphology, which could enhance the quality of cartographic representations in densely built areas.
A Locally Tokenized Generative Model for Robust Time-Series Watermarking
Generative Models
Time Series
- Introduces L-VQVAE, a generative model that ensures local tokenization for time-series data.
- Develops LVQMark, a watermarking method that combines logit-bias injection with robust re-encoding.
- Addresses the issue of false-positive reliability in time-series watermarking by stabilizing re-encoding.
- Demonstrates improved detection power and reduced false positives under post-editing attacks.
Read more
A Locally Tokenized Generative Model for Robust Time-Series Watermarking
Summary
This paper addresses the challenges of watermarking in multivariate time series, particularly the reliability issues that arise under post-editing attacks. The authors identify that existing watermarking detectors suffer from instability due to globally coupled re-encoding, which can lead to bidirectional drift of the null distribution. To overcome this, they propose a novel generative model called L-VQVAE, which produces discrete tokens from a short temporal neighborhood, thus ensuring that perturbations only affect a limited subset of the recovered representation. The watermarking method, LVQMark, leverages this local tokenization to embed a red-green logit bias during autoregressive sampling, enhancing detection robustness. The paper demonstrates that this approach maintains high generation quality while improving detection power and controlling false-positive rates under various post-editing attacks across four benchmark datasets in finance, energy, and neuroimaging.
Methodology
The authors propose L-VQVAE, a generative model that utilizes a local tokenizer, a global decoder, and an autoregressive transformer, all sharing a single codebook. LVQMark is then developed as a watermarking method that embeds a logit bias during sampling and incorporates robust re-encoding to detect attacks effectively. The methodology emphasizes locality in token recovery to prevent global drift in the representation.
Results
Experiments conducted on four datasets reveal that LVQMark preserves high generation quality while achieving robust watermark detection. The method demonstrates stable false-positive behavior on attacked non-watermarked samples, significantly improving upon existing techniques that suffer from reliability issues under post-editing attacks.
Implications
The findings suggest that the proposed methods can enhance the reliability of watermarking in generative models, particularly for applications in synthetic data generation where provenance verification is critical. This has potential implications for fields such as finance, energy, and healthcare, where accurate data attribution is essential.
Orthogonal JEPA: Factorized Predictive States for Latent World Models
Computer Vision
Time Series
Robotics
- Orthogonal JEPA introduces a factorized approach to latent state prediction, enhancing the efficiency of predictive modeling in complex systems.
- The framework employs multiple learned components for target states, each with dedicated prediction pathways, improving gradient flow and reducing redundancy.
- Regularization techniques are integrated to maintain diversity in predictions and prevent encoder collapse.
- The method is evaluated across diverse applications, showcasing its versatility and effectiveness in various domains.
Read more
Orthogonal JEPA: Factorized Predictive States for Latent World Models
Summary
The paper introduces Orthogonal JEPA, a novel framework for latent world modeling that enhances the predictive capabilities of joint-embedding predictive architectures (JEPAs). Traditional JEPAs utilize a single target embedding and prediction pathway, which can lead to inefficiencies in complex systems where multiple signals interact. Orthogonal JEPA addresses this by employing orthogonal predictive factorization, which decomposes target states into multiple components, each predicted by a dedicated branch. This method allows for better allocation of predictive capacity, reducing redundancy and improving gradient flow for weaker signals. The framework incorporates several regularization techniques, including orthogonality objectives and variance regularization, to maintain diversity in the predicted components and prevent encoder collapse. The authors validate their approach through experiments across various domains, including vision, single-cell transcriptomics, health records, control tasks, and molecular dynamics, demonstrating improved representation quality and forecasting capabilities.
Methodology
The authors develop Orthogonal JEPA by creating a framework that utilizes orthogonal predictive factorization. This involves learning basis matrices that decompose target states into multiple components, each predicted through separate branches. The architecture includes context and target encoders, with regularization strategies to ensure effective learning and prevent overfitting. The framework is adaptable to different applications by allowing domain-specific context-target interfaces.
Results
Experiments demonstrate that Orthogonal JEPA significantly improves representation quality and forecasting accuracy across multiple domains, including visual tasks, biological data analysis, clinical event forecasting, control planning, and molecular dynamics. The results indicate that the factorized approach leads to better predictive performance and stability over long horizons compared to traditional JEPA models.
Implications
The findings suggest that Orthogonal JEPA can be applied to a wide range of predictive modeling tasks, particularly in complex systems where multiple interacting factors are present. This framework could enhance the performance of models in fields such as healthcare, robotics, and molecular biology, where accurate predictions of latent states are crucial.
Green BOA: Determining the environmental break-even point for ML-based data compression
Efficient ML
- The study defines the environmental break-even point for ML-based data compression in terms of carbon-equivalent costs.
- The break-even point is highly sensitive to the carbon intensity of the energy mix in different countries.
- ML-based compression algorithms can achieve better compression ratios but may have higher CO2 costs compared to traditional algorithms.
- The research highlights the importance of considering environmental impacts in the development and deployment of ML technologies.
Read more
Green BOA: Determining the environmental break-even point for ML-based data compression
Summary
This paper presents a study on the environmental break-even point for machine learning (ML)-based data compression algorithms, specifically focusing on a lossless compression algorithm developed at the University of Manchester. The authors investigate the carbon-equivalent costs associated with the infrastructure required for ML training and inference, comparing these costs to the carbon savings achieved through reduced disk storage requirements. The study highlights the significant environmental impact of big data storage, particularly in contexts like the Large Hadron Collider, where vast amounts of data are generated. The authors define the break-even point as the dataset size at which the carbon cost of using the ML compression algorithm equals the embodied carbon of the additional storage needed for uncompressed data. The analysis reveals that the break-even point is sensitive to the carbon intensity of the energy sources used in different countries. The results indicate that while ML-based compression can achieve better compression ratios, it may also incur higher CO2 costs compared to traditional algorithms, raising questions about the environmental justification for deploying ML-based methods at scale.
Methodology
The authors estimate the energy usage of an ML-based data compression algorithm on an Nvidia T4 GPU, considering various country-specific energy scenarios to convert energy consumption into carbon-equivalent costs. They analyze the carbon costs of manufacturing and operating disk storage devices over a five-year lifespan, comparing these costs to the carbon costs of training and running the ML compression algorithm.
Results
The analysis shows that the break-even point for the ML-based compression algorithm is influenced by the carbon intensity of the energy used during training and inference. The study finds that tape storage has a lower carbon footprint compared to HDDs, although it offers slower data access. The results suggest that while ML-based compression can provide better compression ratios, it may not always justify the additional environmental costs when deployed at scale.
Implications
This research underscores the need for a careful evaluation of the environmental impacts of ML technologies, particularly in data-intensive fields. It suggests that future work should focus on improving the throughput of ML-based compression algorithms while also considering their environmental costs. The findings could inform decisions on the adoption of ML-based methods in large-scale data storage and processing applications.
SAGE-XGBoost: Spatially Augmented Graph Embeddings–Machine Learning Framework for Natural Hazards Susceptibility Mapping under Data Scarcity
Graph Learning
- SAGE-XGBoost improves natural hazard susceptibility mapping under data scarcity.
- The framework combines data augmentation with graph embeddings for enhanced feature representation.
- It significantly outperforms conventional machine learning models in landslide and wildfire susceptibility mapping.
- Achieved AUC values of 0.97 for landslide and 0.95 for wildfire susceptibility.
Read more
SAGE-XGBoost: Spatially Augmented Graph Embeddings–Machine Learning Framework for Natural Hazards Susceptibility Mapping under Data Scarcity
Summary
This paper addresses the challenge of natural hazard susceptibility mapping, which is often limited by the availability of labeled data. The authors propose a novel framework called SAGE (Spatially Augmented Graph Embeddings) that enhances prediction capabilities under data-scarce conditions. SAGE integrates controlled noise-based data augmentation with neighborhood-based graph embeddings to derive local spatial statistics. A K-nearest neighbor graph is constructed, and the resulting features are reduced using principal component analysis and combined with environmental covariates and spatial coordinates. The SAGE-XGBoost model, which employs these features, was evaluated for landslide and wildfire susceptibility mapping. The results indicate that SAGE-XGBoost significantly outperforms traditional and spatially explicit machine learning models, achieving an absolute improvement of over 33 percentage points compared to Spatial XGBoost. The model attained AUC values of approximately 0.97 for landslide susceptibility and 0.95 for wildfire susceptibility. Feature importance analysis confirmed the significant contribution of graph embeddings to the model's predictive performance, enhancing spatial coherence and reducing local noise amplification. Overall, SAGE-XGBoost presents an efficient and transferable alternative to deep learning methods for environmental hazard assessment, particularly in scenarios with limited supervision.
Methodology
The methodology involves constructing a K-nearest neighbor graph to derive local spatial statistics, which are then reduced using principal component analysis. These features are integrated with environmental covariates and spatial coordinates to train the SAGE-XGBoost model, which utilizes XGBoost as the underlying machine learning algorithm.
Results
SAGE-XGBoost demonstrated superior performance in susceptibility mapping, achieving an absolute improvement of over 33 percentage points compared to Spatial XGBoost. The model reached AUC values of approximately 0.97 for landslide susceptibility and 0.95 for wildfire susceptibility, indicating high predictive accuracy.
Implications
The findings suggest that SAGE-XGBoost can serve as an effective tool for environmental hazard assessment, particularly in regions with limited data availability. This framework can be applied to various geospatial prediction tasks, enhancing decision-making in disaster management and risk reduction.
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Theory
Efficient ML
Time Series
- Introduces a statistically defined 'defer' action for managing uncertainty in continual learning.
- Develops a decision layer based on conditional Jensen–Shannon discrepancy for expert model management.
- Achieves zero false spawns and missed concepts in ideal conditions on synthetic streams.
- Demonstrates improved accuracy and efficiency in real-world datasets with fewer experts maintained.
Read more
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
Summary
This paper addresses the challenges faced by continual-learning systems in managing uncertainty regarding incoming data batches. Instead of treating uncertainty as a nuisance, the author proposes a statistically defined action of 'defer,' which allows systems to wait for more evidence before making decisions about whether to reuse an existing expert model or spawn a new one. The paper introduces a decision layer based on a two-axis task comparison using the conditional Jensen–Shannon discrepancy (CJSD) and its covariate counterpart. Key contributions include: (1) a formal decision semantics that defines reuse and spawn tests with an indifference zone, (2) a sequential evidence process that utilizes per-expert betting e-processes to guide decisions, and (3) a system mechanics that optimizes expert selection and minimizes decision lag. The proposed method demonstrates high accuracy and efficiency across various synthetic and real-world datasets, achieving zero false spawns and missed concepts in ideal conditions, while maintaining a minimal number of experts in more complex scenarios.
Methodology
The methodology involves a decision layer that utilizes a two-axis comparison of task discrepancies to determine whether to reuse, spawn, or defer expert models. The system employs sequential evidence processes with betting e-processes to assess the performance of experts based on incoming data batches. The decision-making is guided by thresholds defined within an indifference zone, ensuring statistically valid actions.
Results
The proposed system achieved zero false spawns and missed concepts on a synthetic stream with an optimal number of experts. In real-world datasets like INSECTS and Covertype, it effectively maintained fewer experts compared to traditional methods while improving accuracy. The system demonstrated robustness across various conditions, with a recommendation for an α-spending variant to enhance decision quality without compromising performance.
Implications
The findings suggest that the proposed decision-making framework can significantly improve the efficiency and accuracy of continual-learning systems in dynamic environments. This approach could be applied in various domains requiring adaptive learning, such as robotics, finance, and personalized recommendations, where managing model uncertainty is crucial.
Rationally Enriched Chebyshev Trunk Bases for DeepONet Surrogates of High Péclet Entrance Transport
Theory
Optimization
Efficient ML
- Introduction of a Rationally Enriched Chebyshev trunk for DeepONet models.
- Demonstrated improved accuracy in predicting solution profiles for high-Péclet transport problems.
- Significant reduction in profile-error metrics compared to vanilla and Chebyshev-trunk DeepONets.
- Effective suppression of artificial oscillations in near-wall predictions.
Read more
Rationally Enriched Chebyshev Trunk Bases for DeepONet Surrogates of High Péclet Entrance Transport
Summary
This paper presents a novel approach to modeling high-Péclet transport problems using a Rationally Enriched Chebyshev (REC) trunk within DeepONet surrogate models. The authors address the challenges posed by singularly perturbed problems characterized by thin boundary or wall layers, which are difficult to resolve using standard numerical methods. The REC trunk integrates Chebyshev polynomial elements with rational dictionary elements derived from the adaptive Antoulas–Anderson (AAA) algorithm. The performance of the REC-trunk DeepONet is evaluated against both a vanilla DeepONet and a Chebyshev-trunk DeepONet across three distinct problems involving varying diffusion-to-advection ratios. The results demonstrate that the REC-trunk DeepONet significantly reduces profile-error metrics, particularly in scenarios where the perturbation parameter is within a specific range, showcasing its effectiveness in predicting wall-normal temperature and concentration profiles while minimizing artificial oscillations. This advancement highlights the potential of combining rational and polynomial bases for improved accuracy in machine learning models applied to complex transport phenomena.
Methodology
The authors developed a DeepONet model with a Rationally Enriched Chebyshev trunk, combining Chebyshev polynomial bases with rational elements generated through the adaptive Antoulas–Anderson algorithm. The model was trained and evaluated on three benchmark problems characterized by singular perturbations, assessing its performance against standard DeepONet architectures.
Results
The REC-trunk DeepONet outperformed the vanilla DeepONet and showed comparable performance to the Chebyshev-trunk DeepONet, particularly excelling in scenarios with perturbation parameters between 1.00 × 10−4 and 1.78 × 10−4, achieving up to a 19.5% reduction in profile-error metrics. Additionally, it achieved reductions of up to 60.2% and 32.2% in wall-normal temperature and concentration profile errors, respectively, while effectively minimizing artificial oscillations.
Implications
The findings suggest that the REC trunk can enhance the predictive capabilities of machine learning models in complex transport scenarios, potentially benefiting applications in biomedical engineering, chemical engineering, and aerospace engineering where accurate modeling of boundary layers is critical.
Improved Confidence Estimates for Black-Box Large Language Models
NLP
Large Language Models
Interpretability
- Proposes a novel method for uncertainty quantification in LLMs using labeled datasets.
- Utilizes classifiers to predict response correctness based on existing uncertainty scores and features from similar queries.
- Demonstrates minimal computational overhead, making it practical for real-world applications.
- Empirical results show significant improvements over existing UQ methods.
Read more
Improved Confidence Estimates for Black-Box Large Language Models
Summary
This paper addresses the challenge of uncertainty quantification (UQ) in large language models (LLMs), which is crucial for their safe deployment in high-stakes environments. Existing UQ methods often operate in a zero-shot manner, generating uncertainty scores without labeled data, which can be insufficient for practical applications. The authors propose a novel approach that leverages a labeled dataset to enhance confidence estimates of LLM responses. By training classifiers that utilize both existing uncertainty scores and features derived from similar queries in a reference dataset, the method aims to predict the correctness of LLM responses more accurately. The framework is computationally efficient and flexible, allowing for the integration of various uncertainty scores into a single confidence estimate. Empirical results demonstrate that this approach consistently outperforms traditional UQ methods across multiple datasets and models, highlighting the importance of labeled data in improving UQ for LLMs.
Methodology
The authors partition a labeled dataset into training and reference sets. They augment training queries with uncertainty scores as features and extract neighborhood-based statistics from the reference set. A classifier is then trained to predict response correctness based on these augmented features, moving beyond the zero-shot paradigm of traditional UQ methods.
Results
The proposed classifiers consistently outperform existing UQ methods across various datasets and models, providing calibrated confidence estimates. The results indicate that leveraging labeled data significantly enhances the reliability of uncertainty estimates, with the optimal classifier model varying by task.
Implications
This work has significant implications for the deployment of LLMs in critical applications such as healthcare and scientific research, where reliable uncertainty quantification is essential. The proposed method can improve decision-making processes by providing more accurate confidence estimates.