AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
64
Papers today
8h
Update frequency
7
Days of history
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
Reinforcement Learning
Large Language Models
Efficient ML
- Introduces LongStraw, an efficient execution stack for long-context RL training.
- Achieves training with context lengths exceeding 2 million tokens under fixed GPU budgets.
- Demonstrates significant memory savings through innovative state management techniques.
- Validates the approach on two distinct model architectures: Qwen and GLM.
Read more
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
Summary
The paper addresses the growing disparity between the context lengths used during reinforcement learning (RL) post-training and those supported during inference, particularly as inference systems approach million-token contexts. The authors introduce LongStraw, an innovative architecture-aware execution stack designed to facilitate long-context RL post-training while adhering to fixed GPU budgets. LongStraw employs Group Relative Policy Optimization (GRPO) and optimizes memory usage by evaluating shared prompts once, retaining only essential model-specific states, and replaying response branches sequentially. This approach significantly reduces GPU memory requirements, allowing for training with context lengths exceeding 2 million tokens. The authors implement LongStraw across two model families, demonstrating its capability to handle up to 4.46 million positions on a multi-GPU setup. The findings suggest that managing state lifetime and ownership is crucial for extending practical context limits in RL training, thereby lowering hardware barriers and enabling broader research opportunities in long-context training.
Methodology
The authors developed LongStraw, which utilizes an architecture-aware execution strategy to optimize memory usage during RL post-training. It evaluates shared prompts once, retains only necessary model-specific states, and replays response branches sequentially, thus reducing the live training graph and memory consumption.
Results
LongStraw successfully completed grouped scoring and response backward for the Qwen model at 2.1 million positions with minimal memory overhead. A stress test extended the execution capacity to 4.46 million positions, validating the end-to-end execution path for the GLM model across all layers.
Implications
The LongStraw framework has the potential to democratize access to long-context RL training by reducing reliance on large GPU clusters, enabling smaller teams and researchers to explore advanced RL applications without significant hardware investments.
Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction
Graph Learning
Efficient ML
Theory
- Introduction of a topology-aligned inductive bias for molecular property prediction.
- Development of two architectures: Iso-QGNN (quantum) and Iso-CGNN (classical) with identical parameter counts.
- Competitive performance achieved with only 64 trainable parameters, demonstrating high data efficiency.
- Both models perform comparably, indicating the inductive bias's significance over the quantum substrate.
Read more
Implementations of Quantum and Classical Topology-Aligned Architectures for Molecular Property Prediction
Summary
This paper addresses the challenge of parameter-efficient learning in quantum chemistry, particularly in low-data and resource-constrained environments. The authors propose a topology-aligned inductive bias where the model architecture reflects the molecular bond graph, with atoms corresponding to computational units and bonds determining interactions through shared learnable parameters. They instantiate this concept in two architectures: a variational quantum circuit (Iso-QGNN) and a parameter-matched classical message-passing model (Iso-CGNN). Both models are evaluated on binary classification tasks related to molecular properties using the QM9 benchmark dataset. The results demonstrate that both architectures achieve competitive performance with only 64 trainable parameters, reaching test AUCs of approximately 0.88 for the quantum model and 0.91 for the classical model on the HOMO-LUMO gap task, and around 0.78 for the dipole moment task. The models achieve 90% of their asymptotic performance with about 250 training molecules, indicating the effectiveness of the topology-aligned inductive bias in enhancing parameter efficiency. The findings suggest that the inductive bias is a crucial factor in achieving performance at the QM9 scale, with implications for benchmarking quantum machine learning against classical methods.
Methodology
The authors utilize the QM9 benchmark dataset, which contains approximately 134,000 small organic molecules. They focus on two binary classification tasks: predicting the HOMO-LUMO energy gap and the electric dipole moment. The Iso-QGNN architecture employs a variational quantum circuit that reflects molecular topology, while the Iso-CGNN uses a classical message-passing network with shared parameters for bonded nodes. Both models are trained and evaluated under matched conditions to isolate the effects of the inductive bias from the quantum substrate.
Results
The Iso-QGNN and Iso-CGNN models achieved test AUCs of approximately 0.88 and 0.91, respectively, on the HOMO-LUMO gap task, and around 0.78 for the dipole moment task. Both models reached 90% of their asymptotic performance with about 250 training molecules, demonstrating their efficiency in parameter usage and training data requirements.
Implications
The findings suggest that the topology-aligned inductive bias is a key factor in achieving efficient learning in molecular property prediction. This has broader implications for the development of quantum machine learning models, particularly in how they are benchmarked against classical counterparts. It also highlights potential areas where quantum implementations may provide advantages in future applications.
Asymmetric Peak-Aware Loss for Peak-Critical Time Series Forecasting
Time Series
- Introduction of Asymmetric Peak-Aware Loss (APAL) to improve peak-critical forecasting.
- Development of a peak-critical evaluation protocol that includes tail error and peak metrics.
- APAL shows improved performance in forecasting rare demand spikes compared to traditional symmetric loss functions.
- The methodology is model-agnostic and can be applied across various forecasting backbones.
Read more
Asymmetric Peak-Aware Loss for Peak-Critical Time Series Forecasting
Summary
This paper addresses the challenges in time-series forecasting, particularly in applications where under-prediction poses a higher risk than over-prediction, such as crowd demand forecasting. The authors introduce Asymmetric Peak-Aware Loss (APAL), a model-agnostic objective that penalizes under-predictions more heavily and emphasizes peak regions in the forecast. This approach aims to improve the accuracy of rare demand spikes, which are critical for operational decision-making. The paper also proposes a peak-critical evaluation protocol that includes metrics for tail error and peak detection, complementing traditional metrics like MAE and MSE. The effectiveness of APAL is evaluated on pedestrian demand forecasting using datasets from the City of Melbourne and a beach visitor count dataset, as well as additional benchmarks from various domains. The results demonstrate that APAL enhances tail accuracy and peak prediction quality while allowing for a controlled trade-off with aggregate error, making it a practical solution for peak-critical forecasting scenarios.
Methodology
The authors propose APAL, which combines asymmetric cost penalties for under- and over-predictions and emphasizes peak regions in the forecast. The evaluation protocol includes standard metrics (MAE, MSE) alongside tail error metrics (Top-10%, Top-1%) and peak-event detection metrics (precision, recall, F1 score). The methodology is tested on pedestrian demand forecasting datasets and additional benchmarks from other domains.
Results
The application of APAL across five forecasting backbones and ten datasets resulted in improved tail accuracy and peak-event quality. The results indicate a clear and tunable trade-off with aggregate error, demonstrating the effectiveness of APAL in scenarios where peak predictions are critical.
Implications
The findings suggest that APAL can significantly enhance forecasting performance in applications where accurate prediction of peak events is crucial, such as urban mobility, crowd management, and resource allocation. The proposed evaluation metrics can help practitioners better assess model performance in peak-critical contexts.
Microstructure-Conditioned Surrogate Models for Graded Multiscale Optimization of Mycelium Composites
Optimization
- Introduction of HyPRNN, a microstructure-conditioned surrogate model for multiscale optimization.
- Demonstrated a 42% reduction in peak stress for optimized graded structures compared to random microstructures.
- Conditioning on manufacturing variables enhances the model's applicability to complex geometries.
- The approach enables efficient multiscale simulations with limited data, addressing challenges in traditional methods.
Read more
Microstructure-Conditioned Surrogate Models for Graded Multiscale Optimization of Mycelium Composites
Summary
This paper presents a novel approach to optimizing mycelium-woodchip composites through the development of microstructure-conditioned surrogate models. The authors introduce a hybrid physics-data surrogate model, termed HyPRNN, which utilizes a hypernetwork to condition a Physically Recurrent Neural Network (PRNN) on microstructural variables. This method allows for accurate predictions of multiscale mechanical behavior, even with limited training data. The study demonstrates the effectiveness of the HyPRNN in performing multiscale simulations for functionally graded structures, achieving a significant reduction in peak stress by 42% compared to structures with random microstructures. Additionally, the authors explore conditioning the model on manufacturing variables, providing a practical pathway to engineer microscale properties for desired macroscale performance. The findings underscore the potential of microarchitectured materials and the role of conditioned surrogate models in accelerating the design and development of sustainable materials.
Methodology
The authors developed a hybrid physics-data surrogate model called HyPRNN, which combines a Physically Recurrent Neural Network (PRNN) with a hypernetwork to condition the model on microstructural variables. This approach allows the model to account for geometric variations and perform multiscale simulations efficiently. The methodology includes iterative optimization of graded structures to enhance mechanical performance.
Results
The HyPRNN was validated against a full FE2 simulation, showing its capability to accurately predict multiscale mechanical behavior. The optimization of a graded multiscale disk resulted in a 42% reduction in peak stress compared to a structure with a random microstructure. The model's conditioning on manufacturing variables further improved its effectiveness in capturing complex microstructural influences.
Implications
The findings suggest that microstructure-conditioned surrogate models can significantly enhance the design and optimization of sustainable materials, particularly in applications requiring functionally graded structures. This approach may lead to more efficient manufacturing processes and improved material performance in various engineering fields.
MESHA: Mechanism-Enforced Sequential Halving for Strategic Linear Bandits
Theory
- MESHA integrates a uniform sampling rule with an epoch-wise Grim Trigger Condition to combat strategic misreporting by arms.
- The algorithm proves that arms will comply with the GTC under Nash Equilibrium to maximize their selection probability.
- MESHA shows a bounded failure probability within a fixed budget, outperforming existing state-of-the-art algorithms in strategic settings.
- Numerical experiments confirm MESHA's effectiveness compared to baseline methods that rely on optimal design sampling rules.
Read more
MESHA: Mechanism-Enforced Sequential Halving for Strategic Linear Bandits
Summary
This paper presents MESHA, an innovative algorithm designed for Best Arm Identification (BAI) in strategic linear bandits, where arms can misreport their feature vectors to enhance their chances of being selected as the best arm. The authors introduce a mechanism that combines a naive uniform sampling rule with an epoch-wise Grim Trigger Condition (GTC) to mitigate the impact of strategic behavior. The GTC effectively eliminates arms that significantly deviate from their true features, ensuring that arms are incentivized to report accurately under a Nash Equilibrium. The paper proves that MESHA maintains a bounded failure probability within a fixed budget T, contrasting with existing state-of-the-art linear BAI algorithms that rely on optimal design (OD)-based sampling rules, which are susceptible to manipulation by self-interested arms. The authors validate MESHA's performance through extensive numerical experiments, demonstrating its superiority over baseline methods that utilize OD-based sampling and feature-agnostic approaches. This work addresses a critical gap in the literature regarding BAI in strategic environments, providing a robust solution that balances exploration and exploitation effectively.
Methodology
The authors develop the MESHA algorithm, which employs a naive uniform sampling strategy combined with an epoch-wise Grim Trigger Condition (GTC) to address the strategic behavior of arms in linear bandits. The GTC is designed to eliminate arms that report features significantly deviating from their true characteristics, thus enforcing truthful reporting under Nash Equilibrium. The theoretical analysis includes deriving performance guarantees and failure probabilities for MESHA within a fixed budget.
Results
The results indicate that MESHA significantly outperforms existing linear BAI algorithms that depend on optimal design sampling rules, which are vulnerable to manipulation by self-interested arms. The numerical experiments demonstrate MESHA's superior ability to accurately identify the best arm within the constraints of a fixed budget.
Implications
The findings suggest that MESHA can be effectively applied in scenarios where strategic behavior is prevalent, such as hiring platforms, recommendation systems, and other decision-making processes involving self-interested agents. This work opens avenues for further research into mechanism design in bandit problems.
Beyond Backbone Backpropagation: A Decoupled Strategy for Efficient Transfer Learning
Efficient ML
Computer Vision
- Introduces a decoupled training strategy that eliminates backbone backpropagation.
- Proposes normalization tuning for efficient domain adaptation.
- Develops margin-based weighted training for improved classifier performance.
- Achieves competitive accuracy on medical benchmarks with reduced training time.
Read more
Beyond Backbone Backpropagation: A Decoupled Strategy for Efficient Transfer Learning
Summary
This paper presents a novel decoupled training framework aimed at enhancing the efficiency of transfer learning in deep learning models, particularly in resource-constrained environments such as medical imaging. The authors argue that traditional backbone backpropagation incurs significant computational overhead, which can be mitigated by decoupling feature extraction from classifier optimization. Their approach involves minimal adaptation of normalization layers, specifically targeting Batch Normalization in CNNs and Layer Normalization in Transformers, to bridge the domain gap without full backbone updates. Additionally, they introduce a margin-based weighted training strategy for the classifier head, which prioritizes ambiguous samples to improve classification performance. The proposed method was evaluated across various CNN and Transformer architectures on three medical datasets, demonstrating competitive accuracy with significantly reduced training times. This efficiency not only lowers the computational burden but also contributes to a more environmentally sustainable approach to deep learning, making it feasible for clinical settings that lack high-performance computing resources.
Methodology
The methodology involves a decoupled framework where feature extraction is separated from classifier optimization. Normalization layers are adapted with minimal updates, and a margin-based weighted loss is applied to the classifier head to enhance decision boundaries. The approach is tested on various CNN and Transformer architectures across multiple medical datasets.
Results
The proposed approach significantly reduces training time while maintaining competitive accuracy compared to traditional methods. It allows for training on standard CPU infrastructure, making it accessible for facilities without high-performance computing capabilities. The method also leads to a substantial reduction in CO2 emissions associated with model training.
Implications
This work has significant implications for deploying deep learning models in resource-constrained environments, particularly in healthcare. It offers a practical solution for efficient transfer learning, enabling broader access to advanced machine learning techniques in clinical settings and potentially reducing the digital divide in technology access.
Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems
Theory
Large Language Models
Optimization
- Introduces a unified risk-based framework for event-triggered LLM invocation.
- Proves six theoretical results related to trigger policies and risk functions.
- Empirically validates the framework using real sensor data and compares it against multiple baselines.
- Demonstrates high diagnostic quality and effective cost sensitivity analysis.
Read more
Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems
Summary
This paper addresses the challenge of determining when to invoke Large Language Models (LLMs) in streaming inference systems, where lightweight models are paired with LLMs for enhanced semantic understanding. The author formulates this problem as a risk-based sequential stopping problem, introducing a trigger policy that activates when a risk functional exceeds a predefined threshold. The paper presents six theoretical results, including guarantees on minimum inter-event times to prevent trigger chattering, optimality of threshold policies, and convergence results for adaptive thresholds. The framework encompasses various classical trigger mechanisms, such as event-triggered and optimal stopping rules. Empirical validation is conducted using turbofan degradation data, demonstrating the effectiveness of the proposed approach against several baselines. The findings reveal sublinear regret in decision-making, high diagnostic quality for LLM outputs, and the superiority of anomaly-score-driven risk functions. Overall, the work provides a principled methodology for balancing the costs of LLM invocation with the risks of missing critical events in real-time systems.
Methodology
The methodology involves framing the decision of when to invoke an LLM as a sequential decision-making problem, utilizing a risk functional that aggregates anomaly scores and uncertainty estimates. The author proves theoretical results regarding optimal stopping and adaptive thresholds, and conducts empirical evaluations on real-world data to validate the assumptions and performance of the proposed framework.
Results
The results confirm that the proposed trigger policies achieve sublinear regret, with 92.9% of LLM diagnoses meeting a grounding score threshold. The analysis shows that the choice of risk function significantly impacts the invocation-miss-rate tradeoff, with anomaly-score-driven functions outperforming others by an order of magnitude on the Pareto AUC.
Implications
The findings have significant implications for the design of real-time streaming systems in various domains, including industrial sensor networks and autonomous vehicles, where timely and accurate decision-making is critical. The framework can enhance the efficiency of LLM usage, reducing costs while maintaining high diagnostic quality.
A Temporal Machine Learning-Based Time-to-Event Model for Predicting ALS Progression and Healthcare Utilization
Time Series
- Development of a digital twin-inspired framework for predicting ALS progression.
- Integration of longitudinal ALSFRS-R data with survival modeling for individualized predictions.
- Identification of lower limb function as a key predictor for wheelchair access.
- Implementation of a temporal machine learning model to capture disease progression.
Read more
A Temporal Machine Learning-Based Time-to-Event Model for Predicting ALS Progression and Healthcare Utilization
Summary
This paper addresses the challenge of predicting clinically meaningful milestones in amyotrophic lateral sclerosis (ALS), a progressive neurodegenerative disease characterized by significant variability in disease progression. The authors propose a novel time-to-event model inspired by digital twin technology, which integrates longitudinal data from the ALS Functional Rating Scale–Revised (ALSFRS-R) with survival modeling to facilitate individualized predictions of functional decline and assistive device utilization. The study constructs a harmonized longitudinal dataset that includes diagnosis records, ALSFRS-R assessments, activities of daily living, and demographic information, ensuring data quality and temporal alignment. Using correlation-based clustering, the authors identify coherent functional domains across various systems affected by ALS. Generalized additive mixed models are employed to characterize nonlinear, domain-specific functional decline. A temporal machine learning model is developed to predict longitudinal functional decline and capture stage-dependent disease progression. The results indicate that lower limb function, particularly walking and stair climbing, are strong predictors of earlier wheelchair access. The proposed digital twin-inspired time-to-event model generates individualized survival curves and dynamically predicts wheelchair-free survival, offering a scalable and interpretable approach for linking ALS progression with personalized decision support, with potential applications in proactive care planning and clinical trial stratification.
Methodology
The authors constructed a harmonized longitudinal dataset and applied correlation-based clustering to identify functional domains. They utilized generalized additive mixed models to analyze nonlinear functional decline and developed a temporal machine learning model alongside Cox proportional hazards modeling to predict disease progression and assistive device utilization.
Results
The study found that lower limb functions, especially walking and stair climbing, were significant predictors of earlier wheelchair access. The temporal machine learning model successfully generated individualized survival curves, allowing for dynamic predictions of wheelchair-free survival.
Implications
The proposed framework provides a clinically actionable tool for predicting ALS progression, which can enhance personalized decision-making, improve proactive care planning, and inform clinical trial designs in precision medicine.
EMAGN: Efficient Multi-Attention Graph Network via Learned Clustering for Scalable Traffic Forecasting
Graph Learning
Time Series
Efficient ML
- Introduces EMAGN, a scalable traffic forecasting model with linear complexity self-attention.
- Achieves competitive accuracy while significantly reducing training and inference times.
- Demonstrates the ability to operate under higher configurations without memory issues.
- Outperforms existing linear attention models in both accuracy and efficiency.
Read more
EMAGN: Efficient Multi-Attention Graph Network via Learned Clustering for Scalable Traffic Forecasting
Summary
Traffic forecasting is a complex task due to the nonlinear spatial and temporal dependencies present in traffic data. Traditional self-attention mechanisms, while effective, suffer from scalability issues due to their quadratic computational and memory complexity. This paper introduces the Efficient Multi-Attention Graph Network (EMAGN), which linearizes the spatial attention mechanism by employing learned clustering matrices to group key and value vectors into super-clusters. This approach reduces the computational complexity from O(N²d) to O(NMd), where M is significantly smaller than N, thus maintaining the flexibility of attention for dynamic dependency modeling. The authors demonstrate that EMAGN achieves a mean absolute error (MAE) within 2.7–3.2% of the full-attention GMAN while significantly reducing training time by 32%, inference time by 38%, and GPU memory usage by 58%. Notably, EMAGN can operate on a standard 11 GB GPU even at higher configurations, unlike full-attention GMAN, which runs out of memory. Additionally, EMAGN outperforms existing linear attention mechanisms like Linformer and Performer in both accuracy and efficiency, showcasing its effectiveness in traffic forecasting tasks.
Methodology
EMAGN employs a linear-complexity self-attention mechanism derived from fast high-dimensional Gaussian filtering. It uses learned clustering matrices to adaptively group key and value vectors into super-clusters, allowing for efficient computation of dependencies while maintaining performance.
Results
EMAGN achieves a mean absolute error (MAE) within 2.7–3.2% of the full-attention GMAN, while reducing training time by 32%, inference time by 38%, and GPU memory usage by 58%. It can handle configurations that exceed the memory limits of traditional models, demonstrating enhanced scalability.
Implications
The findings suggest that EMAGN can be effectively deployed in real-time traffic forecasting applications, particularly in resource-constrained environments. Its ability to maintain performance while reducing computational demands opens avenues for broader applications in Intelligent Transportation Systems (ITS).
Leveraging unlabelled data for generalizable neural population decoding
Time Series
Interpretability
Multimodal
- Introduction of MOJO, a joint SSL-SL framework for spike-tokenizing models.
- Demonstrated superior performance over traditional SL-only models, particularly in few-shot learning scenarios.
- Enhanced interpretability of neuronal representations and improved performance on classification tasks.
- Generalization of the framework to human ECoG data, achieving competitive results.
Read more
Leveraging unlabelled data for generalizable neural population decoding
Summary
This paper introduces MOJO (Masked autOencoder-based JOint training), a novel framework that enhances neural population decoding by integrating self-supervised learning (SSL) with supervised learning (SL). Traditional spike-based models have been limited by their reliance on labeled data, which restricts their training to datasets with paired behavioral labels. MOJO addresses this limitation by allowing models to leverage unlabelled data through masked autoencoding, facilitating the extraction of meaningful representations from neural data. The authors evaluated MOJO on diverse datasets, including monkey motor cortex recordings and mouse visual decision-making tasks, demonstrating that it significantly outperforms models trained solely with SL, especially in scenarios with limited labeled data. The framework also enhances interpretability of neuronal representations and generalizes to human electrocorticography (ECoG) data during speech tasks, achieving performance comparable to specialized neuro-foundation models. Overall, MOJO represents a significant advancement in the flexibility and scalability of neural decoding models, enabling more effective use of unlabelled data across various tasks and species.
Methodology
The MOJO framework combines self-supervised learning via masked autoencoding with supervised learning objectives. It tokenizes neural data at the spike level and optimizes both learning objectives simultaneously, allowing for effective representation learning from both labeled and unlabelled data. The framework was tested on various neural datasets using both Transformer and State-Space Model architectures.
Results
MOJO consistently outperformed purely supervised models across multiple datasets, including those involving monkey reaching tasks and mouse visual decision-making. The framework showed particularly strong performance in few-shot learning scenarios, where limited labeled data was available. Additionally, it provided more interpretable neuronal representations and successfully generalized to human ECoG data, achieving results comparable to specialized neuro-foundation models.
Implications
The findings suggest that MOJO can significantly enhance the training of neural decoders in neurotechnologies, such as brain-computer interfaces, by leveraging unlabelled data. This approach could lead to more flexible and scalable models that can adapt to various neural data modalities and tasks, ultimately improving the efficacy of neurotechnological applications.
EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models
Generative Models
Optimization
Large Language Models
- EXPLORE is the first framework to integrate test-time scaling with language model decoding for analog topology generation.
- The framework employs simulator-guided Monte Carlo Tree Search (MCTS) to enhance the efficiency of topology generation.
- By filtering structural tokens, EXPLORE reduces simulation trials, making it feasible to scale to higher-complexity circuits.
- The success rate for generating valid topologies improved from 12% to 65% at a tight tolerance of 0.01.
Read more
EXPLORE: Exploration with Guided Search for Analog Topology Generation using Language Models
Summary
The paper presents EXPLORE, a novel framework designed to automate the generation of analog circuit topologies, addressing the limitations of existing one-shot generation methods that struggle with complex circuits due to their vast search spaces and limited training datasets. By integrating simulator-guided Monte Carlo Tree Search (MCTS) with transformer-based decoding, EXPLORE enhances the test-time scaling capabilities for analog topology generation. The framework utilizes language model priors while strategically bypassing high-confidence structural tokens to optimize the use of simulation resources. The authors constructed a large-scale dataset of higher-complexity analog topologies to facilitate this process. The results demonstrate a significant improvement in success rates and mean squared error (MSE) compared to previous methods, establishing EXPLORE as a pioneering approach in the field of automated analog circuit design.
Methodology
EXPLORE combines a transformer-based language model with a simulator-guided Monte Carlo Tree Search (MCTS) to explore the design space of analog circuit topologies. The framework leverages language model priors to guide the search process and employs structured-token filtering to minimize the number of simulations required. This approach allows for efficient exploration of complex circuit designs while utilizing a large-scale dataset of analog topologies.
Results
The implementation of EXPLORE resulted in a success rate of 65% for generating valid analog topologies at a tight tolerance of 0.01, significantly higher than the 12% success rate of one-shot generation methods and 33% for sampling-and-filter approaches. Additionally, EXPLORE achieved over 20 times lower MSE compared to the sampling-and-filter baseline under the same search budget, demonstrating faster convergence and improved efficiency.
Implications
The advancements presented in EXPLORE have the potential to revolutionize the field of analog circuit design by automating the topology generation process, thereby reducing the manual effort required and accelerating development cycles. This framework could lead to more efficient design workflows and enable the creation of customized circuits that meet diverse application demands.
Sharp Stability Threshold and Certification for Designing Stable Residual Architectures
Theory
Optimization
- Establishes a stability threshold for residual architectures with the condition q ≤ 1.
- Demonstrates that exceeding this threshold can lead to divergence in training.
- Introduces an arithmetic framework for input-magnitude exponents to guide architectural design.
- Confirms through experiments that architectures with q ≤ 1 train stably, regardless of normalization layers.
Read more
Sharp Stability Threshold and Certification for Designing Stable Residual Architectures
Summary
This paper introduces the sublinear-growth principle for deep residual architectures, establishing a sharp stability threshold for the input-magnitude exponent of residual block velocity fields. The authors demonstrate that the condition q ≤ 1 is both necessary and sufficient for stable training, where q represents the growth rate of the velocity field. They derive this threshold through classical ODE theory and optimal-control analysis, showing that at q > 1, velocity fields can lead to divergence during training. The paper also discusses the role of normalization layers in stabilizing training and proposes an arithmetic framework for input-magnitude exponents, allowing for efficient certification of architectural designs. Experimental results on Mamba and PatchTST architectures confirm that models adhering to the q ≤ 1 criterion train stably, emphasizing the importance of the input-magnitude exponent over the presence of normalization layers.
Methodology
The authors utilize classical ODE theory to analyze the stability of residual architectures and apply optimal-control analysis through the Hamilton–Jacobi–Bellman equation to derive the stability threshold. They develop an arithmetic framework for input-magnitude exponents to facilitate the design and certification of stable architectures.
Results
The main results indicate that the condition q ≤ 1 is critical for ensuring stable training in deep residual architectures. The authors successfully demonstrate that architectures designed with this criterion exhibit stable training dynamics, as confirmed by experiments on Mamba and PatchTST models.
Implications
The findings suggest that architectural design in deep learning can be systematically guided by the proposed stability criterion, potentially leading to more robust and efficient neural network architectures. This could influence future research in deep learning design principles and practices.
Mono-Z Dark Matter Search with Neural Spline Flows Using CMS Run 2015D Open Data
Generative Models
Theory
- First application of Neural Spline Flow likelihood-ratio scoring in mono-Z DM search using CMS open data.
- Defined signal and control regions based on kinematic properties of events.
- Set observed upper limits on signal strength for scalar, vector, and axial-vector mediators.
- Demonstrated the potential of NSFs to model complex event densities in high-energy physics.
Read more
Mono-Z Dark Matter Search with Neural Spline Flows Using CMS Run 2015D Open Data
Summary
This paper presents a novel search for dark matter (DM) produced in association with a leptonically decaying Z boson, utilizing CMS Run 2015D open data with an integrated luminosity of 2.32 fb−1. The authors focus on the mono-Z → ℓ+ℓ− final state, specifically in the channels Z → µ+µ− and Z → e+e−, requiring opposite-sign same-flavor dilepton pairs with invariant mass between 60 and 120 GeV. A signal region is defined based on missing transverse energy (Emiss_T) and angular separation criteria. The authors employ Neural Spline Flows (NSFs) to model the background and DM signal densities, training separate flows for Standard Model (SM) and DM hypotheses. The likelihood-ratio score is computed to assess the presence of DM against the SM background. The analysis yields upper limits on the signal strength for various mediator types, indicating no significant evidence for DM but highlighting the effectiveness of NSFs in collider physics applications.
Methodology
The authors utilized Neural Spline Flows to model the background density from control-region Drell–Yan events and the signal density from Monte Carlo samples with different mediator types. They constructed a 37-dimensional feature vector from kinematic observables and applied a likelihood-ratio score to evaluate the consistency of observed events with SM predictions versus DM hypotheses.
Results
The analysis resulted in observed upper limits on the signal-strength parameter for scalar, vector, and axial-vector mediators, with values of µ < 0.0177, µ < 0.0362, and µ < 0.0498 respectively. The observed limits exceeded expected limits by factors of 7-12, attributed to background modeling discrepancies rather than evidence for DM.
Implications
The findings suggest that while no significant DM signal was detected, the methodology employed could enhance future searches for DM in collider experiments. The use of NSFs may provide a more robust framework for modeling complex event distributions in high-energy physics.
Privacy Leakage in Federated Learning in Radiology Reports: A Comparative Evaluation of Tokenizer-Driven Privacy Risks
Federated Learning
NLP
Large Language Models
- Federated learning can lead to significant privacy leakage in radiology reports despite not sharing raw data.
- Different tokenizer designs influence the extent of privacy leakage, with domain-specific tokenizers like RadBERT performing better in reconstructing clinical terms.
- Exact sentence reconstruction accuracy ranged from 31% to 44% across different tokenizers and batch sizes.
- Increased batch sizes generally resulted in lower reconstruction fidelity.
Read more
Privacy Leakage in Federated Learning in Radiology Reports: A Comparative Evaluation of Tokenizer-Driven Privacy Risks
Summary
This paper investigates the privacy risks associated with federated learning (FL) in the context of radiology reports, focusing on how different tokenizer designs can influence the potential for privacy leakage. The authors quantify the extent to which sensitive information can be reconstructed from model updates in an FL setting, using three different transformer tokenization strategies: GPT-2, RadBERT, and LLaMA-2. The study employs a controlled evaluation with six FL clients training a GPT-2-style transformer on two public radiology datasets, comprising over 368,000 diagnostic reports and discharge summaries. The results reveal that significant portions of radiology report text can be reconstructed from FL gradients, with reconstruction fidelity varying across tokenizers. Notably, RadBERT demonstrated higher reconstruction fidelity and better recovery of clinical terms compared to the other tokenizers. The findings underscore the importance of tokenizer design in privacy evaluations for clinical language models and suggest that additional safeguards, such as secure aggregation and differential privacy, are necessary to comply with data protection regulations like HIPAA and GDPR.
Methodology
The study involved training a GPT-2-style transformer model across six federated learning clients using two public radiology corpora. Three different tokenizers (GPT-2, RadBERT, LLaMA-2) were employed, and the models were evaluated under an active malicious-server threat model. Reconstruction fidelity was assessed using metrics such as exact sentence accuracy, S-BLEU, G-BLEU, and ROUGE-L over multiple runs.
Results
The study found that exact sentence reconstruction accuracy varied between 31% and 44% across different tokenizers and datasets. For instance, at a batch size of 64 on the Discharge dataset, the accuracies were 42.1% for GPT-2, 42.3% for RadBERT, and 39.4% for LLaMA-2, with accuracy decreasing as batch size increased. RadBERT consistently yielded higher reconstruction fidelity and better recovery of clinical terms compared to the other tokenizers.
Implications
The findings highlight the critical need for careful consideration of tokenizer design in federated learning applications, particularly in sensitive domains like healthcare. The results suggest that without additional privacy safeguards, federated learning systems may not adequately protect patient confidentiality, necessitating the integration of advanced privacy-preserving techniques to comply with regulatory standards.
NeuroGRIP: Retrieval-Augmented Graph Refinement for Knowledge-Grounded EEG Seizure Diagnosis
Graph Learning
Time Series
Interpretability
- NeuroGRIP integrates external medical knowledge to refine EEG seizure diagnosis graphs.
- The framework constructs a domain-specific knowledge base from clinical guidelines.
- It employs large language models for extracting structured biomedical knowledge.
- NeuroGRIP enhances seizure detection accuracy and interpretability of predictions.
Read more
NeuroGRIP: Retrieval-Augmented Graph Refinement for Knowledge-Grounded EEG Seizure Diagnosis
Summary
NeuroGRIP introduces a novel framework for improving EEG seizure diagnosis by integrating external medical knowledge into spatial-temporal graph neural networks (STGNNs). The framework addresses the challenges of noisy and clinically implausible graph structures that arise from purely data-driven approaches. By constructing a large-scale knowledge base from authoritative clinical guidelines, NeuroGRIP utilizes large language models (LLMs) to extract structured biomedical entities and relations, forming a textual knowledge graph. The framework employs a Semantic Alignment Query module to project STGNN-generated EEG node embeddings into the semantic space of the knowledge graph, enabling alignment-aware query construction. This process retrieves relevant clinical relations through FAISS-based similarity search, allowing for the assignment of confidence scores to predicted edges based on retrieved evidence. The resulting refined graph enhances both the accuracy of seizure detection and the interpretability of predictions by grounding them in clinically validated knowledge. Extensive experiments on TUSZ and CHB-MIT datasets demonstrate the effectiveness of NeuroGRIP in improving diagnostic performance and providing explainable clinical insights.
Methodology
NeuroGRIP constructs a knowledge graph from clinical guidelines and uses large language models to extract relevant biomedical entities and relations. It employs a Semantic Alignment Query module to align EEG node embeddings with the knowledge graph, retrieves relevant clinical relations using FAISS for similarity search, and assigns confidence scores to edges based on the retrieved evidence, ultimately refining the graph structure.
Results
Experiments on TUSZ and CHB-MIT datasets show that NeuroGRIP significantly improves seizure detection accuracy compared to traditional STGNN approaches, while also enhancing the interpretability of the predictions by grounding them in clinically validated knowledge.
Implications
The proposed framework paves the way for knowledge-enhanced and explainable clinical diagnosis in EEG-based seizure detection, potentially improving clinical trust and decision-making in epilepsy management.
Federated Explainable Artificial Intelligence: Roles, Architectures, Evaluation, and Open Challenges
Federated Learning
Interpretability
- FedXAI combines privacy-preserving Federated Learning with Explainable AI to enhance model transparency.
- A multi-axis taxonomy is introduced to categorize FedXAI methods based on various dimensions.
- Current evaluation practices for FedXAI are fragmented, lacking standardized benchmarks and metrics.
- Open challenges include explainability under non-IID data and security threats related to explanations.
Read more
Federated Explainable Artificial Intelligence: Roles, Architectures, Evaluation, and Open Challenges
Summary
This paper presents a comprehensive survey on Federated Explainable Artificial Intelligence (FedXAI), which integrates the principles of Federated Learning (FL) and Explainable Artificial Intelligence (XAI). The authors highlight the necessity of combining privacy-preserving model training with explainability, particularly in high-stakes domains where transparency is crucial. The paper introduces a multi-axis taxonomy for categorizing FedXAI methods based on various factors such as the role of explainability, model types, and integration levels. It reviews existing approaches, ranging from model-agnostic explanations to interpretable federated models, and discusses the challenges of evaluating these systems, including the lack of standardized benchmarks for explanation quality and privacy concerns. The authors identify several open challenges in the field, such as addressing non-IID data, security threats related to explanations, and the need for communication-efficient XAI. The survey serves as a foundational reference for developing trustworthy and transparent federated AI systems.
Methodology
The authors conducted a systematic review of the literature on FedXAI, categorizing existing methods and evaluating their roles, architectures, and challenges. They introduced a taxonomy to organize the findings and highlighted gaps in current evaluation practices.
Results
The survey reveals a growing body of work at the intersection of FL and XAI, emphasizing the need for integrated approaches that address both privacy and explainability. It identifies significant challenges in evaluating explanation quality and the risks associated with explanations in federated settings.
Implications
The findings suggest that integrating explainability into federated systems is essential for applications in healthcare, finance, cybersecurity, and energy systems, where decision-making must be transparent and trustworthy. The survey provides a framework for future research and development of FedXAI systems.
The Hyperspherical Geometry of CLIP Latent Space: A Semantic Mixture Model
Multimodal
- Introduces MovMF-CLIP, a model that captures the hyperspherical geometry of CLIP latent space using Mixtures of von Mises–Fisher distributions.
- Demonstrates that Gaussian-based models inadequately represent the multimodal and directional nature of CLIP embeddings.
- Achieves significant improvements in long-tailed and OOD detection, reducing false positive rates substantially.
- Provides intrinsic interpretability by allowing embeddings to be decomposed into sparse combinations of semantic prototypes.
Read more
The Hyperspherical Geometry of CLIP Latent Space: A Semantic Mixture Model
Summary
This paper addresses the limitations of existing probabilistic models for the CLIP (Contrastive Language–Image Pretraining) latent space, which are typically based on Gaussian assumptions. The authors propose a new model, MovMF-CLIP, which utilizes Mixtures of von Mises–Fisher (MovMF) distributions to better capture the intrinsic hyperspherical geometry of the CLIP representations. The proposed method employs the Expectation-Maximization (EM) algorithm for efficient learning, allowing each mixture component to correspond to a coherent semantic concept. This approach yields a closed-form likelihood that aligns with the hyperspherical structure, enhancing the interpretability and accuracy of density estimation. The authors demonstrate that MovMF-CLIP significantly improves performance in long-tailed and out-of-distribution (OOD) detection tasks, as well as providing a natural semantic decomposition of embeddings. The findings suggest that the CLIP latent space is more accurately represented as a hyperspherical semantic mixture rather than an isotropic Gaussian, offering a robust framework for understanding multimodal representations.
Methodology
The authors perform covariance-based geometric calibration via whitening to address global anisotropy in the CLIP latent space. They then model the normalized representations using Mixtures of von Mises–Fisher distributions, estimated through the Expectation-Maximization algorithm. This method allows for a principled density estimation that respects the hyperspherical geometry of the embeddings.
Results
MovMF-CLIP significantly reduces the false positive rate for OOD detection from 67.76% to 48.00% on MS-COCO, and from 75.05% to 33.48% for tail concepts. It achieves the highest Semantic Relevance score of 0.673 while providing a 13-fold speedup in inference time compared to the next best method. Overall, the model consistently enhances robustness, interpretability, and stability across various tasks.
Implications
The findings suggest that a more accurate modeling of the CLIP latent space can lead to better performance in multimodal tasks, particularly in scenarios involving rare or out-of-distribution concepts. The framework could be applied to improve robustness and interpretability in other machine learning models that operate in high-dimensional spaces.
MetaPerch: Learning from metadata for bioacoustics foundation models
Audio & Speech
Multimodal
- MetaPerch utilizes metadata as auxiliary targets to improve species identification in bioacoustics.
- The model employs a multi-task learning approach to jointly train on species identification and metadata prediction.
- Extensive experiments reveal that metadata significantly enhances model performance across various datasets.
- The study provides insights into the importance of different metadata modalities and their effects on training.
Read more
MetaPerch: Learning from metadata for bioacoustics foundation models
Summary
The paper introduces MetaPerch, a novel bioacoustic foundation model that leverages metadata from citizen science platforms to enhance species identification performance. Traditional bioacoustic models primarily rely on vocalization data, but this work emphasizes the untapped potential of recording metadata, such as location and time, as auxiliary supervision signals. By employing a multi-task learning approach, MetaPerch is trained to predict various metadata alongside species identification, allowing the model to learn richer representations that generalize better across different ecological and acoustic domains. The authors conduct extensive empirical studies across 17 bioacoustic datasets to evaluate the impact of nine diverse metadata sources on model performance, demonstrating that incorporating metadata significantly improves species identification accuracy. The findings highlight the importance of metadata in bridging domain gaps and addressing challenges in real-world passive acoustic monitoring settings.
Methodology
The authors developed MetaPerch using a multi-task learning framework that jointly trains the model on species identification and multiple metadata prediction tasks. They conducted empirical studies to assess the effects of different metadata sources on model performance across 17 bioacoustic datasets, employing various loss formulations and design choices to optimize training.
Results
The results indicate that incorporating metadata as auxiliary supervision leads to improved species identification performance. The empirical analysis shows that different metadata modalities contribute variably to the model's robustness and generalization capabilities, effectively addressing challenges related to domain shifts in bioacoustic monitoring.
Implications
The findings suggest that leveraging metadata can significantly enhance the effectiveness of bioacoustic models, making them more applicable in real-world monitoring scenarios. This approach could facilitate better biodiversity assessments and conservation efforts by improving the accuracy of species identification in diverse ecological contexts.
Analytical study of the optimal combination of binary classifiers based on classifiers-induced partitioning of the training set
Theory
Optimization
Efficient ML
- Establishes conditions for the existence and uniqueness of the minimum of convexified empirical risk.
- Derives analytical formulas for optimal weights, avoiding iterative optimization methods.
- Introduces the concept of Ï•-frontiers to assess classifier stability and data quality.
- Analyzes the implications of classifier-induced partitioning on the training set structure.
Read more
Analytical study of the optimal combination of binary classifiers based on classifiers-induced partitioning of the training set
Summary
This paper investigates the optimal linear combination of binary classifiers by structuring the dataset through truth tables. The authors propose a method that partitions data into equivalence classes, which facilitates a rigorous analysis of the convexified empirical risk using a multidimensional generalization of classification calibrated functions. They establish sufficient conditions for the existence and uniqueness of the global minimum of the convexified empirical risk, particularly when dealing with a large number of classifiers. The analysis is detailed for three classifiers, identifying configurations that lead to unique or non-unique solutions. The authors derive explicit analytical formulas for optimal weights using Exponential (Boost) and Logistic (Logit) loss functions, circumventing the need for iterative optimization. Additionally, they introduce the concept of Ï•-frontiers to evaluate classifier stability and data quality. The study highlights the challenges of achieving a unique optimal classifier and the complexities involved in numerical optimization, particularly in cases where the minimum may not exist.
Methodology
The authors utilize a logical structuring of the dataset via truth tables to partition data into equivalence classes. They generalize the convexified empirical risk to a multidimensional context and derive conditions for optimality. The analysis includes deriving explicit formulas for optimal weights based on specific loss functions and studying the stability of classifiers through Ï•-frontiers.
Results
The paper presents sufficient conditions for the existence and uniqueness of the optimal classifier's weights. It also provides explicit formulas for calculating these weights and demonstrates the potential for non-uniqueness or non-existence of optimal classifiers in certain scenarios. The introduction of Ï•-frontiers allows for a new way to evaluate classifier stability.
Implications
The findings have significant implications for ensemble learning and classifier design, particularly in improving the robustness and accuracy of classifiers by leveraging the diversity of weak classifiers. The analytical approach may enhance the understanding of classifier interactions and the impact of data quality on classification outcomes.
What's in a Smoothness Constant? Tighter Rates for Local SGD with Bounded Second-order Heterogeneity
Optimization
Federated Learning
Theory
- Establishes first convergence guarantees for Local SGD under bounded second-order heterogeneity in general convex settings.
- Introduces a trajectory-dependent control of heterogeneity, improving upon traditional uniform bounds.
- Demonstrates that Local SGD can outperform Mini-batch SGD without restrictive assumptions on first-order heterogeneity.
- Provides nearly tight upper and lower bounds for Local SGD performance.
Read more
What's in a Smoothness Constant? Tighter Rates for Local SGD with Bounded Second-order Heterogeneity
Summary
This paper addresses the convergence guarantees of Local Stochastic Gradient Descent (SGD), also known as Federated Averaging, under the assumption of bounded second-order heterogeneity. While Local SGD has shown practical advantages over Mini-batch SGD, theoretical understanding of its performance in heterogeneous data settings has been limited. The authors build on previous work that established the benefits of second-order heterogeneity for strongly convex objectives and extend these findings to general convex settings. They provide improved convergence guarantees and establish nearly tight upper and lower bounds for Local SGD. The key innovation is a trajectory-dependent control of heterogeneity, which allows for a more nuanced understanding of how local updates can outperform averaging methods without relying on restrictive assumptions about first-order heterogeneity. Additionally, the paper presents a lower bound for serial SGD with replacement, illustrating the impact of high-curvature clients on performance. Overall, the results contribute to a more refined convergence theory for Local SGD, highlighting its advantages in realistic distributed optimization scenarios.
Methodology
The authors utilize a theoretical approach to analyze Local SGD's convergence properties by establishing new bounds based on second-order heterogeneity. They employ a self-bounding recursion to couple trajectory-wise heterogeneity with iterate error, allowing for a more accurate assessment of Local SGD's performance in various settings.
Results
The paper presents improved convergence guarantees for Local SGD in general convex settings, showing that it can achieve better performance than Mini-batch SGD under bounded second-order heterogeneity. The authors also derive nearly tight upper and lower bounds for Local SGD, confirming the effectiveness of their theoretical framework.
Implications
The findings suggest that Local SGD is a robust optimization method for distributed learning scenarios, particularly in environments with heterogeneous data distributions. This work could enhance the design of federated learning systems, making them more efficient and effective in practice.
TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents
Reinforcement Learning
Large Language Models
- TRACE provides a dense credit-assignment method for long-horizon reinforcement learning tasks.
- The approach utilizes a frozen reference model to evaluate intermediate actions without needing additional critics.
- Significant performance improvements were observed on closed-web and open-web benchmarks.
- TRACE enables faster convergence and earlier improvements in learning curves compared to traditional methods.
Read more
TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents
Summary
The paper introduces TRACE (Turn-level Reward Assignment via Credit Estimation), a novel approach to credit assignment in reinforcement learning for multi-turn agents. Traditional outcome rewards are insufficient for long-horizon tasks, as they provide sparse and high-variance feedback that fails to differentiate between useful and irrelevant actions. TRACE addresses this by assigning dense rewards at tool-call boundaries, utilizing a frozen reference model to evaluate the predictability of the final answer based on intermediate actions. By transforming log-probabilities into log-ratio state values, TRACE computes per-action rewards through Temporal-Difference (TD) changes, allowing for effective credit assignment without the need for additional critics or process-label training. The method significantly enhances the tool-use capabilities of agents in long-horizon complex search tasks, demonstrating improved performance on both closed-web and open-web benchmarks without requiring cold-start supervised fine-tuning or live-web data. The results indicate that TRACE not only accelerates learning but also leads to better exploration and interaction with the environment, ultimately refining tool use for complex tasks.
Methodology
TRACE represents rollouts as state transitions at tool-call boundaries and uses a frozen reference model to derive log-ratio state values. It assigns per-action rewards based on Temporal-Difference changes, allowing for dense credit assignment without additional supervision or training stages.
Results
On the closed-web BrowseComp-Plus benchmark, TRACE improved the performance of Qwen3-4B from 7.2 to 35.6 and Qwen3-30B-A3B from 8.4 to 42.6. The learned search behavior also transferred to open-web benchmarks, with notable scores achieved on various tasks, indicating enhanced tool-use capabilities.
Implications
TRACE has the potential to improve the performance of long-horizon agents in various applications, including web navigation, software engineering, and complex task execution, by enabling more effective exploration and interaction with environments.
VAIOM: Continuous-Input, Discrete-Output Decoder-Only Financial Sequence Modeling
Time Series
- VAIOM separates input representation from output likelihood, allowing for continuous inputs and categorical outputs.
- The model outperforms traditional baselines in predicting next-period normalized returns in financial sequences.
- Continuous-input representation improves performance over discrete-token branches while maintaining categorical objectives.
- Full-sequence autoregressive supervision enhances model training compared to last-position training.
Read more
VAIOM: Continuous-Input, Discrete-Output Decoder-Only Financial Sequence Modeling
Summary
The paper introduces VAIOM (Vector-Input Autoregressive Inference for Ordinal-Return Modeling), a novel decoder-only Transformer designed for probabilistic next-return modeling in financial sequences, specifically targeting one-hour foreign-exchange (FX) bars. Unlike traditional models that rely on discrete symbolic inputs, VAIOM utilizes continuous multivariate financial-event vectors to maintain the numerical structure of the data. The model predicts a categorical distribution over the next volatility-normalized return bucket, allowing for effective cross-entropy training and likelihood evaluation. The Hybrid Continuous Input (HybridContIn) model integrates continuous event features with categorical asset metadata and employs a Mixture-of-Market-States (MoMS) return head. The study evaluates the model's performance against various baselines, including Frequency, Markov, and LightGBM, demonstrating that VAIOM consistently outperforms these models in terms of return likelihood. Key findings reveal that continuous input representation enhances performance, full-sequence autoregressive supervision is superior to last-position training, and auxiliary objectives improve return likelihood. The results establish a compact framework for financial sequence modeling that effectively captures the complexities of financial data.
Methodology
VAIOM employs a decoder-only Transformer architecture that processes continuous multivariate financial-event vectors as input. It predicts a categorical distribution over the next volatility-normalized return bucket, utilizing cross-entropy training. The model incorporates various auxiliary objectives and a Mixture-of-Market-States return head to enhance performance. Model selection and evaluation are conducted using a validation set, with final performance assessed on held-out test data.
Results
The selected VAIOM model consistently outperformed train-fitted baselines, achieving paired gains of approximately 0.029–0.043 bits per event over LightGBM across multiple test periods. Validation experiments confirmed that continuous-input representation and full-sequence autoregressive supervision significantly improved return likelihood.
Implications
The findings suggest that VAIOM can effectively model complex financial sequences, providing a robust tool for probabilistic forecasting in financial markets. This approach may enhance decision-making processes in trading and risk management by offering more accurate predictions of market behavior.
Muse: Representation Geometry of Muon Beyond Normalized Momentum
Optimization
Large Language Models
Theory
- Muse optimizers introduce a representation-indexed view of Muon-style optimization.
- Different matrix representations significantly affect the optimizer's geometry and performance.
- Balanced non-native representations can match the performance of native representations in training.
- Shorter dimensions in matrix representations weaken scaling and singular-channel support.
Read more
Muse: Representation Geometry of Muon Beyond Normalized Momentum
Summary
This paper introduces Muse, a family of Muon-style optimizers that explores the impact of different matrix representations on optimization geometry. Muon optimizers utilize a polar map for matrix momentum updates, which depend on the representation of parameter blocks prior to orthogonalization. Muse examines how various Frobenius-isometric representations (native, nearest-square, skinny, and vector) influence the optimizer's geometry and performance. The study reveals that the shorter dimension of the matrix representation affects the number of singular channels, pullback scaling, and convergence bounds in stochastic nonconvex settings. Through experiments with LLaMA2-130M and LLaMA2-600M models, the authors demonstrate that balanced non-native representations can achieve performance comparable to native representations, while reducing the shorter dimension leads to diminished scaling and singular-channel support, resulting in behavior akin to normalized momentum. The findings highlight the significance of representation choice in Muon optimizers and provide insights into the underlying mechanisms of performance variation.
Methodology
The authors analyze the geometry induced by various Frobenius-isometric representations of matrix blocks and establish a polar steepest-descent direction under spectral/nuclear-norm geometry. They conduct pretraining experiments on LLaMA2 models to evaluate the performance of different representations, using fixed-momentum diagnostics to assess the impact of representation on optimization outcomes.
Results
The results indicate that Muse optimizers with balanced non-native representations can achieve validation losses comparable to those of native representations. The study also finds that reducing the shorter dimension of matrix representations leads to a decline in performance, aligning behavior closer to normalized momentum. The analysis establishes bounds on rank and norm that differentiate matrix-valued polar updates from vector endpoints.
Implications
The findings suggest that careful selection of matrix representations in Muon-style optimizers can enhance performance in large language model training, potentially leading to more efficient training strategies and improved convergence in complex optimization landscapes.
Accuracy-Preserving Stability Regularization for Large-Scale Retail Demand Forecasting
Time Series
- Introduces a training-time penalty for consecutive forecasts to improve forecast stability.
- Demonstrates that stability regularization can enhance forecast stability without significantly degrading point accuracy.
- Evaluates the proposed method using a temporal-structured pipeline with various operational features.
- Shows improvements in Forecast Stability Score over XGBoost while keeping RMSE changes minimal.
Read more
Accuracy-Preserving Stability Regularization for Large-Scale Retail Demand Forecasting
Summary
This paper addresses the challenge of retail demand forecasting, particularly focusing on the stability of forecast paths in addition to point accuracy. Traditional forecasting methods often prioritize minimizing point errors without considering the abrupt movements between consecutive forecasts, which can lead to operational inefficiencies. The authors propose a novel training-time penalty on consecutive forecasts to enhance forecast stability while maintaining point accuracy. They implement this approach within a temporal-structured pipeline that integrates recent demand embeddings with various operational features such as calendar, price, and store characteristics. The study evaluates the proposed stability-aware hybrid model against XGBoost on the M5 demand series dataset at different scales. The results indicate significant improvements in forecast stability scores, with minimal changes in root mean square error (RMSE), demonstrating that training-time regularization can effectively balance the trade-off between accuracy and stability. This work extends the evaluation of forecasting methods beyond mere point-error minimization to include operational usability, highlighting the importance of stability in retail demand forecasting.
Methodology
The authors adapt a total-variation penalty for consecutive forecasts into the XGBoost objective function, contrasting it with post-hoc smoothing techniques. They utilize a temporal-structured pipeline that combines a compact demand-history embedding with various operational features to assess the impact of stability regularization on forecast performance.
Results
The stability-aware hybrid model improves the Forecast Stability Score over XGBoost by 6.91%, 6.66%, and 7.68% at scales of 1000, 3000, and 4000 series, respectively, while RMSE changes remain within 0.72% across different random seeds. In contrast, post-hoc smoothing reduces raw movement but incurs a larger RMSE cost.
Implications
The findings suggest that incorporating stability into the training process of forecasting models can lead to more reliable and usable forecasts in operational settings, particularly in retail environments where demand patterns can be erratic. This approach could inform future developments in forecasting methodologies and operational decision-making processes.
Accelerating A/B-Tests with Counterfactual Estimation: Reducing Variance through Policy Overlap
Theory
Efficient ML
- Introduces a novel A/B-testing protocol leveraging policy overlap to reduce variance.
- Establishes a variance dominance theorem proving that the proposed estimators outperform the standard Difference-in-Means estimator.
- Identifies an optimal traffic allocation strategy based on policy divergence.
- Proposes the Δ-MRDR estimator for minimizing ATE estimation variance.
Read more
Accelerating A/B-Tests with Counterfactual Estimation: Reducing Variance through Policy Overlap
Summary
This paper addresses the inefficiencies in standard A/B-testing protocols, particularly the variance issues that impede the statistical power necessary for reliable treatment effect assessment. The author identifies that when treatment and control policies agree on an action, the resulting outcomes contribute noise without providing useful information about the treatment effect, leading to inflated confidence intervals. To mitigate this, the paper proposes a novel experimental protocol that utilizes policy overlap to enhance the efficiency of A/B-tests. By framing the randomized treatment assignment as a meta-policy and employing Δ-Off-Policy Estimation (Δ-OPE) methods, the author demonstrates how to obtain unbiased estimates of average treatment effects (ATE) while reducing variance. The theoretical framework established shows that the proposed approach can recover standard A/B-testing practices while offering improved variance scaling based on policy divergence. Empirical results validate the theoretical findings, suggesting significant potential for improving the evaluation of recommender systems, information retrieval, and large language model interfaces.
Methodology
The paper employs a theoretical framework that applies Δ-Off-Policy Estimation (Δ-OPE) to A/B-testing data by interpreting treatment assignment as a meta-policy. It includes the development of new estimators (Δ-MRDR and Δ-DCG) and a variance dominance theorem to analyze the performance of these estimators compared to traditional methods.
Results
The proposed estimators demonstrate a strict variance reduction compared to the standard Difference-in-Means estimator, particularly in scenarios where treatment and control policies have significant overlap. The empirical validation shows that the new methods can achieve substantial variance reduction with minimal engineering overhead.
Implications
The findings have significant implications for the design and analysis of online experiments, particularly in fields such as recommender systems and information retrieval, where variance reduction can lead to more efficient and reliable evaluations of treatment effects.
CoDiffGRN: Rethinking Gene Regulatory Network Inference via the BEELINE-KGC Benchmark and Co-evolutionary Discrete Diffusion
Graph Learning
- Introduces BEELINE-KGC, a new benchmark for GRN inference that emphasizes inductive generalization and top-K ranking.
- Presents CoDiffGRN, a co-evolutionary discrete diffusion model that captures the conditional nature of gene regulations.
- Implements a cell-cluster-based discretization strategy to enhance the robustness of predictions for unseen genes.
- Demonstrates significant improvements in regulatory discovery compared to existing state-of-the-art methods.
Read more
CoDiffGRN: Rethinking Gene Regulatory Network Inference via the BEELINE-KGC Benchmark and Co-evolutionary Discrete Diffusion
Summary
The paper addresses the challenges in inferring gene regulatory networks (GRNs) from single-cell transcriptomic data, highlighting the misalignment between existing methods and real-world biological needs. Current benchmarks primarily use transductive splits and global classification metrics, which do not effectively support the discovery of high-confidence regulatory interactions, especially for unseen genes. To overcome these limitations, the authors propose a new benchmark, BEELINE-KGC, which incorporates an inductive gene-holdout split and knowledge graph completion metrics for evaluating top-ranked predictions. They also introduce CoDiffGRN, a novel co-evolutionary discrete diffusion framework that models gene expression states and regulatory interactions to enhance inductive generalization and improve regulatory discovery. The methodology includes a cell-cluster-based discretization strategy and TF-ALL Subgraph Sampling (TASS) for scalable training. Extensive experiments demonstrate that CoDiffGRN achieves state-of-the-art performance in novel regulatory discovery, significantly outperforming existing methods and validating the effectiveness of the proposed design through ablation studies.
Methodology
The authors reformulate GRN inference as an inductive, ranking-centric graph completion problem. They develop the BEELINE-KGC benchmark with an inductive gene-holdout split and knowledge graph completion metrics. CoDiffGRN employs a co-evolutionary discrete diffusion model that integrates gene expression states and regulatory interactions, utilizing a gene-state-dependent transition matrix to condition edge transitions on the states of their endpoint genes. The TF-ALL Subgraph Sampling (TASS) method is introduced for scalable training.
Results
CoDiffGRN establishes new state-of-the-art performance in novel regulatory discovery, significantly outperforming existing methods. The experiments conducted on the BEELINE-KGC benchmark reveal severe generalization bottlenecks in current state-of-the-art models, while CoDiffGRN effectively mitigates these issues.
Implications
The findings suggest that the proposed methods can significantly enhance the discovery of regulatory interactions in biological research, particularly for novel genes and rare cell types. This could lead to advancements in understanding gene regulation mechanisms and their implications in various biological contexts.
HyperShadow: A Benchmark for Detecting 3D Projections of Higher-Dimensional Spatial Objects
Computer Vision
Theory
- HyperShadow is the first benchmark for detecting 3D projections of higher-dimensional spatial objects.
- Traditional intrinsic-dimension estimation methods fail to accurately identify shadows, achieving only 71-73% accuracy.
- A compact learned point network achieves 96.2% accuracy in detecting projections.
- The rigidity witness statistic effectively separates classes with an AUROC of 0.982.
Read more
HyperShadow: A Benchmark for Detecting 3D Projections of Higher-Dimensional Spatial Objects
Summary
The paper introduces HyperShadow, a novel benchmark designed to detect 3D projections of higher-dimensional spatial objects, specifically in dimensions 4 to 6. Unlike traditional datasets labeled as '4D', which typically combine three spatial dimensions with time, HyperShadow focuses solely on spatial dimensions. The benchmark consists of 10,800 static point clouds and 1,800 temporal sequences, allowing for the classification of native 3D shapes versus their projections. The study highlights the limitations of intrinsic-dimension estimation methods, which achieve only 71-73% accuracy on this task, contrasting with a proposed 190k-parameter point network that reaches 96.2% accuracy. Additionally, the paper introduces a zero-parameter rigidity witness statistic that effectively distinguishes between rigid and non-rigid motions in temporal data, achieving an AUROC of 0.982. The dataset, models, and code are publicly available, providing a controlled environment for exploring the detectability of higher-dimensional projections without making claims about physical reality.
Methodology
The methodology involves generating a dataset of point clouds representing native 3D shapes and their projections from higher dimensions. The study employs a compact learned point network for classification and introduces a rigidity witness statistic for temporal data analysis. The benchmark includes various corruption tiers to assess model robustness and generalization.
Results
The proposed point network achieved 96.2% accuracy in distinguishing between native 3D shapes and their higher-dimensional projections, significantly outperforming traditional methods. The rigidity witness statistic demonstrated an AUROC of 0.982, effectively identifying rigid versus non-rigid motions in temporal sequences.
Implications
HyperShadow provides a foundational tool for researchers to explore the characteristics of higher-dimensional projections and their detectability, potentially influencing future studies in computer vision and machine learning. It opens avenues for understanding geometric regularities and their implications in various applications, including robotics and computer graphics.
Scalable Training of Continuous-Time Spiking Neural Networks with Differentiable Spike-Time Discretization
Efficient ML
Theory
Time Series
- Introduces a memory-efficient framework for training continuous-time SNNs using differentiable spike-time discretization (DSTD).
- Reduces memory consumption from O(NoutNin) to O(NoutM) for time-to-first-spike coding.
- Implements synfire-chain-inspired temporal regularization to improve layer-wise firing organization and reduce dead-neuron issues.
- Achieves up to 100-fold reduction in peak memory and 20-fold decrease in training time compared to traditional methods.
Read more
Scalable Training of Continuous-Time Spiking Neural Networks with Differentiable Spike-Time Discretization
Summary
This paper addresses the challenges of training deep continuous-time spiking neural networks (SNNs), which are limited by the memory requirements for exact spike-time computation. The authors propose a memory-efficient training framework utilizing differentiable spike-time discretization (DSTD) for leaky integrate-and-fire (LIF) neurons. DSTD maps irregular presynaptic spikes to weighted events at fixed time points, significantly reducing the memory needed for candidate spike times. The authors also introduce a temporal regularization inspired by synfire chains to organize layer-wise firing windows, which mitigates dead-neuron failures and enhances processing efficiency. The proposed methods lead to a drastic reduction in peak memory consumption and training time, enabling the training of deep SNNs on standard datasets like CIFAR-10 and Fashion-MNIST. The results demonstrate that the DSTD framework can effectively approximate continuous-time dynamics while maintaining organized spike propagation across layers, thus facilitating scalable training of deep SNN architectures.
Methodology
The authors extend the differentiable spike-time discretization (DSTD) to leaky integrate-and-fire (LIF) neurons, mapping irregular spike trains to fixed time points. They also incorporate temporal regularization to encourage organized firing patterns across layers, which helps in maintaining gradient flow and reduces dead-neuron occurrences.
Results
The proposed DSTD method significantly lowers peak memory consumption by nearly two orders of magnitude and reduces training time by up to 20-fold. The framework successfully trains deep SNNs, achieving effective spike-time propagation and organized firing across layers, demonstrated through experiments on CIFAR-10 and Fashion-MNIST.
Implications
This work has potential applications in neuromorphic computing and computational neuroscience, enabling the development of more efficient and scalable spiking neural network architectures that can be implemented in real-time systems.
Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems
Reinforcement Learning
Robotics
Optimization
- Introduction of the SVR-MARL framework for multi-AUV systems.
- Focus on practical communication constraints rather than idealized models.
- Demonstration of improved task efficiency in covert operations.
- Emphasis on the utility of sensed information for cooperative tasks.
Read more
Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems
Summary
This paper addresses the challenges faced by autonomous underwater vehicles (AUVs) during covert cooperative missions, where reliance on active sonar and frequent communications can lead to exposure risks. The authors propose a novel framework called Sensed Information Value Realization Multi-Agent Reinforcement Learning (SVR-MARL), which aims to enhance cooperative task efficiency while minimizing communication and exposure risks. The framework incorporates realistic communication constraints and focuses on the utility of information for cooperative tasks rather than treating communication as an idealized flow. Through a case study on covert multi-AUV cooperative localization and tracking, the authors demonstrate that their approach significantly improves task efficiency by allowing AUVs to share critical information while adhering to covert operational requirements. The study highlights the importance of balancing information sharing and covert operations in multi-agent systems operating in complex underwater environments.
Methodology
The authors developed the SVR-MARL framework, which integrates multi-agent reinforcement learning with a focus on the value of sensed information for cooperative tasks. The framework allows AUVs to learn distributed policies under realistic communication constraints, facilitating effective information sharing while maintaining covert operations. A case study involving cooperative localization and tracking was conducted to evaluate the framework's performance.
Results
The results from the case study indicate that the SVR-MARL framework significantly enhances the efficiency of collaborative tasks among AUVs while reducing unnecessary communication and exposure risks. The framework effectively balances the need for information sharing with the constraints of covert operations, demonstrating its practical applicability in real-world scenarios.
Implications
The findings suggest that the SVR-MARL framework can be applied to various underwater missions, such as environmental monitoring, security operations, and resource exploration, where covert operations are critical. The approach may also inform future developments in multi-agent systems and communication strategies in complex environments.
Decoding Market Emotion from Blockchain Activity: A Data-Driven Sentiment Classifier
NLP
Time Series
Interpretability
- Integration of blockchain data with social media sentiment to explain market behavior.
- Focus on sentiment analysis rather than price prediction.
- Gradient Boosting (XGBoost) achieved an F1-score of around 0.84 for sentiment classification.
- SHAP values were used for model interpretability, enhancing transparency.
Read more
Decoding Market Emotion from Blockchain Activity: A Data-Driven Sentiment Classifier
Summary
This study investigates the sentiment surrounding Bitcoin by integrating on-chain data, financial metrics, and social media sentiment, particularly from Twitter. Unlike traditional models that focus on price prediction, this research aims to explain market sentiment through a comprehensive analysis of blockchain transactions and social media posts. The authors developed a unified dataset that combines these diverse data sources, allowing for a more nuanced understanding of market behavior. Various machine learning models were evaluated, with Gradient Boosting (XGBoost) demonstrating the highest reliability in sentiment classification, achieving an average F1-score of approximately 0.84. The use of SHAP (SHapley Additive exPlanations) provided insights into the contribution of on-chain features, enhancing model interpretability. The findings suggest that the integration of blockchain activity and social media sentiment can yield valuable insights for cryptocurrency analysis, offering a new perspective for investors and analysts.
Methodology
The authors combined on-chain data, historical Bitcoin price data, and daily Twitter sentiment classifications to create a unified dataset. They employed multiple machine learning models, particularly focusing on Gradient Boosting (XGBoost), and utilized cross-validation for model evaluation. SHAP was applied to interpret the model's predictions and quantify the influence of on-chain features.
Results
The study found that the integration of blockchain activity and social media sentiment provides meaningful predictive signals regarding market sentiment. The XGBoost model achieved an average F1-score of about 0.84, indicating strong performance in sentiment classification. The use of SHAP values improved the interpretability of the model, allowing for a better understanding of how on-chain features contribute to sentiment predictions.
Implications
This research offers a novel approach to cryptocurrency analysis by providing insights into market sentiment through the lens of blockchain activity and social media. It can aid investors in making informed decisions and enhance the understanding of market dynamics. The findings also suggest potential for future improvements using deep learning techniques.
Explainable Geospatial AI for Satellite Ground Station Siting Using LiDAR-Derived Terrain Intelligence
Interpretability
- Introduces a machine learning framework for predicting Representative Clutter Height (RCH) using LiDAR and open geospatial data.
- Achieves a mean absolute error of 1.79 m, significantly improving upon traditional fixed clutter height methods.
- Utilizes LightGBM for its accuracy and efficiency, along with SHAP for feature attribution analysis.
- Demonstrates the potential for improved site selection and reduced uncertainty in satellite ground station siting.
Read more
Explainable Geospatial AI for Satellite Ground Station Siting Using LiDAR-Derived Terrain Intelligence
Summary
This paper addresses the challenge of accurately predicting Representative Clutter Height (RCH), a crucial parameter for radio propagation and interference analysis in satellite ground station siting. Traditional methods rely on fixed clutter heights based on land use categories, which often overlook significant variations within these classes, leading to conservative site exclusions and poor ranking. The authors propose an interpretable machine learning framework that utilizes LiDAR-derived terrain intelligence and various open geospatial data sources to predict RCH. The model employs LightGBM, achieving a mean absolute error (MAE) of 1.79 m and an R² of 0.765, significantly outperforming the ITU baseline by over 60%. The study emphasizes the importance of feature attribution analysis using SHAP, identifying key predictors such as tree canopy cover and land-cover semantics. The framework is designed to be globally deployable and interpretable, enhancing the accuracy of clutter modeling without sacrificing usability. The findings suggest that improved RCH predictions can lead to better site selection and spectrum coordination for satellite operations.
Methodology
The authors developed a supervised learning model using LightGBM, trained on LiDAR-derived labels and various open geospatial features, including land cover, terrain, and demographic data. They defined RCH using the 75th percentile clutter height statistic and employed SHAP for feature importance analysis.
Results
The model achieved a mean absolute error of 1.79 m and an R² of 0.765 on held-out U.S. data, outperforming fixed ITU-R clutter defaults by over 60%. The analysis also highlighted the importance of tree canopy cover and land-cover semantics as key predictors.
Implications
The proposed framework can enhance the accuracy of satellite ground station siting and spectrum coordination by providing more reliable RCH estimates. This can lead to better-informed decisions in the planning and deployment of satellite communication infrastructure.
Supervised Fine-Tuning vs. In-Context Learning: An Equilibrium Analysis of LLM Personalization under Congestion
Large Language Models
NLP
Theory
- SFT and ICL are optimal in different regimes influenced by pretraining coverage and data quality.
- Congestion from multiple users can flip the performance ranking of SFT and ICL.
- Equilibrium resource consumption is non-monotonic, with various factors affecting congestion levels.
- Offering both SFT and ICL maximizes platform profits, aligning with observed trends in AI service offerings.
Read more
Supervised Fine-Tuning vs. In-Context Learning: An Equilibrium Analysis of LLM Personalization under Congestion
Summary
This paper explores the trade-offs between Supervised Fine-Tuning (SFT) and In-Context Learning (ICL) for personalizing Large Language Models (LLMs) in the context of resource congestion. The authors develop a continuum-user model that captures the economic and statistical dynamics of LLM personalization, addressing how users should choose between SFT and ICL based on the availability of computational resources. The analysis reveals that SFT and ICL perform optimally in different scenarios, influenced by factors such as pretraining coverage and data signal-to-noise ratios. Congestion from multiple users can alter the effectiveness of these methods, leading to non-monotonic resource consumption patterns. The study also provides insights into platform strategies for offering personalization options, demonstrating that providing both SFT and ICL can maximize platform profits despite increased computational demands. Experimental validation using GPT-2 on linear regression tasks supports the theoretical findings, indicating that the choice of personalization method significantly impacts performance and user experience.
Methodology
The authors develop a continuum-user model that employs linear approximations of personalization algorithms to derive error formulas for SFT and ICL. They analyze user behavior under congestion and study platform pricing strategies, using theoretical and experimental approaches to validate their findings.
Results
The analysis shows that SFT dominates when pretraining coverage exceeds a critical threshold, while ICL performs better in certain conditions. The study identifies non-monotonic relationships in resource consumption based on pretraining precision and task difficulty. The findings also demonstrate that offering both personalization methods does not harm platform profits, which is supported by empirical data from major AI platforms.
Implications
The results have significant implications for AI service providers in designing personalization strategies and pricing models. Understanding the dynamics of user choices under congestion can help platforms optimize resource allocation and improve user satisfaction. The findings also contribute to the broader AI-economics literature by integrating behavioral insights into the personalization of LLMs.
Counterfactual Optimal Action Trees (COAT): Interpretable Prescriptive Policies from Observational Data
Optimization
Interpretability
- COAT combines counterfactual outcome estimation with mixed-integer optimization for interpretable decision-making.
- The framework was validated in a live pilot with a major airline, achieving significant revenue increases.
- COAT addresses the challenge of deploying AI in regulated environments by ensuring decisions are interpretable and compliant.
- The study illustrates the potential of operations research to enhance AI-driven decision systems.
Read more
Counterfactual Optimal Action Trees (COAT): Interpretable Prescriptive Policies from Observational Data
Summary
The paper introduces the Counterfactual Optimal Action Tree (COAT) framework, designed to learn interpretable prescriptive policies from observational data. COAT integrates counterfactual outcome estimation with large-scale mixed-integer optimization, utilizing column generation to create feasible and transparent decisions that adhere to business and regulatory constraints. The framework is applied to airline ancillary pricing, where it demonstrated a 6.9% increase in upsell revenue per booking during a 17-week field pilot with a major airline, projecting an annual revenue increase of $50–$150 million. The study emphasizes the importance of bridging predictive AI with operational decision-making, highlighting COAT's role in producing interpretable policies that can be deployed responsibly in high-stakes environments. The results underscore the potential of operations research to enhance AI applications in real-world settings, particularly in sectors requiring compliance and transparency.
Methodology
COAT operates in two stages: first, it estimates counterfactual outcomes for various actions, and second, it solves a constrained optimization problem to derive an interpretable policy represented as an action tree. The optimization is achieved through a path-based formulation using column generation, allowing for scalability in large policy spaces.
Results
In a 17-week pilot, COAT increased upsell revenue per booking by 6.9%, translating to an estimated $50–$150 million in additional annual revenue for the airline. The evaluation employed the synthetic control method to provide credible causal evidence in a non-experimental setting.
Implications
The COAT framework has significant implications for industries requiring interpretable and compliant decision-making systems, particularly in high-stakes environments like airlines. Its successful implementation can lead to broader adoption of AI-driven decision initiatives, enhancing operational efficiency and revenue generation while adhering to regulatory standards.
ExTernD: Expanded-Rank Ternary Decomposition Ternary LLM PTQ with Accuracy Approaching Any Quantization Level
Large Language Models
Efficient ML
Optimization
- ExTernD allows for flexible inner rank expansion, enabling precise control over quantization accuracy.
- The method can achieve accuracy levels approaching bf16, surpassing limitations of fixed-plane ternary quantization methods.
- A batched block-ALS GPU algorithm significantly improves computational efficiency while maintaining accuracy.
- Empirical results show ExTernD's effectiveness on various LLMs, achieving competitive performance metrics.
Read more
ExTernD: Expanded-Rank Ternary Decomposition Ternary LLM PTQ with Accuracy Approaching Any Quantization Level
Summary
The paper introduces ExTernD (Expanded-rank Ternary Decomposition), a novel approach for post-training quantization (PTQ) of large language models (LLMs) that allows for a flexible and accurate representation of weight matrices. By factorizing each weight matrix into a product of ternary matrices and a real-valued scale vector, ExTernD expands the inner rank beyond the traditional full rank, enabling it to correct quantization errors effectively. The authors prove that the residual error decreases monotonically with the expanded rank, allowing the method to achieve accuracy levels approaching bf16 precision. This is a significant advancement over existing ternary quantization methods that are limited by fixed plane counts. The methodology includes a greedy alternating least squares (ALS) algorithm and a batched block-ALS GPU implementation, which enhances computational efficiency. Empirical results demonstrate that ExTernD achieves competitive accuracy on benchmark datasets, outperforming traditional methods in terms of effective bits per weight (bpw) and perplexity metrics.
Methodology
ExTernD employs a post-training factorization of weight matrices into ternary components and a scale vector, utilizing a greedy alternating least squares (ALS) algorithm for optimization. The method allows for an expanded inner rank, enabling the model to fit residual errors from previous components, thus improving accuracy. A batched block-ALS implementation is used for efficient computation on GPUs, and an importance-weighted variant is introduced to account for varying significance across input channels.
Results
ExTernD matches or exceeds the accuracy of existing quantization methods, achieving effective bits per weight (bpw) of 5.2–5.5 on benchmark models like Gemma-4-E2B and Qwen3.5-4B. A full conversion of Qwen3.5-4B at an inner rank multiplier of 3 results in a perplexity of 10.10 on wikitext-2, compared to 9.78 for bf16, indicating a 3.2% improvement.
Implications
The ExTernD method has significant implications for the deployment of large language models in resource-constrained environments, allowing for efficient quantization without sacrificing accuracy. This could enhance the accessibility and performance of LLMs in various applications, including real-time processing and edge computing.
Learning in Infinitesimal Non-Compositional Sketches
Theory
- Introduces LINCS, a categorical framework for addressing non-compositionality in ML.
- Defines Infinitesimal Non-Compositionality (INC) as an obstruction to factorization in learning sketches.
- Establishes Tangent Learning Sketches to ensure admissibility of tangent lifts.
- Proves the existence of a final INC coalgebra and discusses convergence properties.
Read more
Learning in Infinitesimal Non-Compositional Sketches
Summary
This paper introduces a novel categorical framework called Learning in Infinitesimal Non-Compositional Sketches (LINCS), which reinterprets machine learning (ML) as a process of addressing non-compositionality in learning sketches. The framework defines ML problems as sketches that incorporate commutativity conditions and limit/colimit structures, moving beyond traditional scalar loss functions. The concept of Infinitesimal Non-Compositionality (INC) is central to LINCS, representing the obstruction to factorization in learning sketches when infinitesimal perturbations are applied. The paper also presents Tangent Learning Sketches, which ensure that if a model is admissible, its tangent lift is also admissible. This leads to the definition of LINCS categories that include both the original and tangent factorization data. The framework allows for the exploration of higher-order interactions and the development of a coalgebraic fixed point for ML problems. The paper proves the existence of a final INC coalgebra under certain conditions and discusses geometric convergence and uniqueness of stabilized behaviors in complete metric realizations. An experimental evaluation of LINCS is planned across various ML domains, including deep learning and reinforcement learning.
Methodology
The methodology involves developing a categorical framework that defines ML problems as sketches with specific structures. It utilizes the tangent functor to explore infinitesimal perturbations and their effects on compositionality. The framework incorporates axioms and universal properties to establish the theoretical foundation for LINCS and its applications in various ML contexts.
Results
The paper demonstrates that every learning compositionality problem has a tangent lift, allowing for a deeper understanding of how perturbations affect learning processes. It establishes the existence of a final INC coalgebra under certain realizations and proves the uniqueness and convergence of stabilized behaviors in complete metric settings.
Implications
The implications of LINCS are significant for advancing the theoretical understanding of ML, particularly in how models can be evaluated and improved based on their compositional properties. It opens avenues for more robust learning algorithms that can handle perturbations and non-compositionality more effectively, potentially enhancing performance in complex ML tasks.
Towards a Unified Multidimensional Explainability Metric: Evaluating Trustworthiness in AI Models
Interpretability
- Introduction of a unified multidimensional explainability metric for XAI methods.
- Focus on fidelity, simplicity, and stability as key aspects of explainability.
- Development of an offline knowledge base for context-dependent evaluation.
- Demonstration of the framework on three open-source datasets.
Read more
Towards a Unified Multidimensional Explainability Metric: Evaluating Trustworthiness in AI Models
Summary
This paper presents a comprehensive framework aimed at creating a unified multidimensional explainability metric for evaluating the trustworthiness of various Explainable AI (XAI) methods, including LIME and SHAP, across different datasets and machine learning models. The authors focus on three critical aspects of explainability: fidelity, simplicity, and stability. Through benchmarking experiments, they systematically assess these aspects and develop an offline knowledge base that captures explainability scores for registered models. This knowledge base facilitates context-dependent evaluation and allows for the estimation of explainability scores for unseen datasets and models. The framework is demonstrated using three open-source datasets, revealing insights into the characteristics of the datasets and the implications for the explainability of AI models. The proposed metric aims to provide a robust and objective means of evaluating and comparing XAI methods, ultimately contributing to the development of more transparent and trustworthy AI systems.
Methodology
The authors conducted benchmarking experiments to evaluate the explainability aspects of various XAI methods. They constructed a knowledge base that captures explainability scores for different models and datasets, enabling the estimation of scores for new, unseen scenarios. The framework integrates multiple properties of XAI techniques into a single metric.
Results
The application of the proposed framework to three datasets yielded insights into the explainability characteristics of different models and methods. The results highlighted the variability of fidelity, simplicity, and stability based on the dataset and model used, demonstrating the utility of the unified metric.
Implications
The unified explainability metric has the potential to enhance the transparency and trustworthiness of AI systems across various domains, particularly in safety-critical applications. It provides researchers and practitioners with a common benchmark for evaluating XAI methods, facilitating informed decision-making regarding the selection of appropriate explanation techniques.
Local Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist
Interpretability
- Proposes a common framework for local additive feature attribution methods.
- Organizes methods around five specification choices that influence their assumptions and failure modes.
- Identifies common failure modes linked to the mathematical assumptions of attribution methods.
- Introduces a ten-item reporting checklist for studies using local additive attributions.
Read more
Local Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist
Summary
This paper presents a comprehensive survey of local additive feature attribution methods, which are crucial for explainable artificial intelligence (XAI). The authors propose a unified framework that categorizes various attribution methods, including Shapley values, path-based methods, gradient/backpropagation techniques, perturbation distributions, and CAM-style methods, based on five key specification choices: value function, reference, path, perturbation distribution, and conservation rule. The survey highlights the mathematical assumptions underlying these methods and discusses common failure modes such as baseline sensitivity and adversarial manipulation. A significant contribution of the paper is the introduction of a ten-item reporting checklist designed to guide researchers in transparently communicating the assumptions and specifications of their attribution methods. The authors emphasize that attribution results are meaningful only in the context of the mathematical assumptions they are based on, advocating for clearer reporting in the literature.
Methodology
The authors conducted a survey of existing local additive feature attribution methods, categorizing them based on their mathematical specifications. They analyzed the methods through an axiom-by-method matrix, linking the assumptions to common failure modes and proposing a reporting checklist to enhance clarity in future studies.
Results
The survey reveals that many attribution methods produce conflicting explanations due to differing underlying assumptions. It establishes that no single method can satisfy all desirable properties simultaneously, leading to inherent disagreements among methods. The proposed checklist aims to standardize reporting practices, ensuring that researchers disclose the mathematical assumptions that underpin their attribution results.
Implications
The findings of this survey have significant implications for the field of explainable AI, particularly in enhancing the reliability and interpretability of machine learning models. By advocating for clearer reporting of assumptions, the paper aims to foster better understanding and trust in attribution methods, ultimately improving their application in real-world scenarios.
A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems
Time Series
Interpretability
Efficient ML
- DynaBase is a minimal two-parameter model for zero-shot dynamical system reconstruction.
- It achieves competitive performance with significantly fewer parameters than existing models.
- The model allows for closed-form solutions for prediction MSE and direct optimization on reconstruction metrics.
- Different training strategies lead to fundamentally different outcomes in model performance.
Read more
A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems
Summary
This paper addresses the challenge of zero-shot reconstruction of dynamical systems (DS) using a minimal interpretable architecture called DynaBase. The authors begin by simplifying the state-of-the-art model DynaMix, which is known for its strong out-of-domain generalization but lacks interpretability. Through systematic ablation and simplification, they derive DynaBase, a two-parameter model that forecasts DS by linearly blending the current latent state with the nearest in-context neighbor and its temporal successor. Despite its simplicity, DynaBase achieves competitive performance in zero-shot DS reconstruction across various chaotic and cyclic systems while maintaining a significantly lower parameter load compared to other foundation models. The model allows for direct optimization on DS reconstruction metrics and provides closed-form solutions for prediction mean squared error (MSE). The authors also explore the impact of different training strategies, revealing that training for ahead-prediction often leads to context parroting, while specific DSR training yields more accurate dynamics. This work not only highlights the minimal requirements for zero-shot DS reconstruction but also reconciles divergent observations in the literature regarding model performance and complexity.
Methodology
The authors iteratively simplify the DynaMix model through systematic ablation, resulting in DynaBase, which uses a linear combination of the current state and context neighbors. They analyze the model's performance theoretically and empirically, exploring the effects of various training strategies on the model's ability to reconstruct dynamical systems.
Results
DynaBase demonstrates highly competitive zero-shot reconstruction performance across chaotic and cyclic systems, with a parameter load many orders of magnitude lower than other foundation models. The model's two parameters allow for a spectrum of dynamical behaviors, including chaotic dynamics and context parroting, depending on the training approach used.
Implications
The findings suggest that simpler models can achieve comparable or superior performance in dynamical system reconstruction, emphasizing the importance of interpretability and minimalism in model design. This has potential applications in scientific and engineering domains where understanding the underlying dynamics is crucial.
Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment
Optimization
Theory
- Increasing ad load can significantly raise revenue but may negatively impact user engagement and search conversions.
- The effects of ad load vary considerably across different queries, with high-ad-conversion queries benefiting more than low-conversion ones.
- The proposed e-LAAL algorithm effectively adapts ad load in real-time, improving performance over static benchmarks.
- The study emphasizes the importance of understanding user intent and advertiser composition in ad load design.
Read more
Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment
Summary
This paper investigates the ad-load design problem in sponsored search advertising, particularly focusing on the trade-off between increased revenue from additional sponsored slots and the potential negative impact on user engagement and search conversions. Conducted through a large-scale randomized field experiment on an Android app store, the study reveals that increasing ad load can boost revenue by up to 43%, but may also reduce total search conversions by up to 5% and daily engagement by 2.2%. The authors highlight significant heterogeneity in the effects of ad load across different queries, indicating that while high-ad-conversion queries benefit from additional slots, low-conversion queries may experience diminished returns or even losses. To address these challenges, the authors propose an adaptive algorithm called exploration-augmented Locally Adaptive Ad Load (e-LAAL), which combines a model-free decision rule with static exploration arms to optimize ad load dynamically. The deployment of e-LAAL on a platform serving over 22 million users demonstrated improved revenue-conversion trade-offs compared to static benchmarks, showcasing its effectiveness in adapting to varying market conditions.
Methodology
The authors conducted a 66-day randomized field experiment involving over 5 million users and 26 million searches to assess the impact of varying the number of sponsored slots on user engagement and revenue. They developed the e-LAAL algorithm, which combines a model-free decision-making approach with static exploration to adaptively optimize ad load.
Results
The experiment found that increasing the number of sponsored slots raised revenue by up to 43%, but also led to a reduction in total search conversions by up to 5% and daily engagement by 2.2%. The e-LAAL algorithm outperformed static benchmarks in terms of the revenue-conversion trade-off in a production environment serving 22.3 million users.
Implications
The findings suggest that platforms can enhance their monetization strategies by employing adaptive algorithms like e-LAAL, which take into account user behavior and market dynamics. This approach can lead to better user experiences while maximizing revenue, particularly in competitive environments like mobile app marketplaces.
A VAE-Driven Multi-Task Satellite-Aided Semantic Communication Framework for 6G-Enabled Connected Autonomous Vehicles
Computer Vision
Generative Models
Robotics
- Introduction of a VAE-based framework for semantic communication in CAVs.
- Utilization of probabilistic latent representations for improved robustness in satellite channels.
- Joint optimization of traffic sign reconstruction and classification tasks.
- Implementation of a composite perceptual loss for enhanced image quality.
Read more
A VAE-Driven Multi-Task Satellite-Aided Semantic Communication Framework for 6G-Enabled Connected Autonomous Vehicles
Summary
This paper presents a novel framework for semantic communication tailored for connected autonomous vehicles (CAVs) utilizing 6G satellite communication. The authors address the inefficiencies of conventional communication systems that transmit raw data, which is particularly problematic in resource-constrained satellite channels. Instead, they propose a Variational Autoencoder (VAE)-based multi-task framework that focuses on transmitting task-relevant information, specifically for traffic sign recognition and classification. The framework employs probabilistic latent representations to enhance robustness against noise and improve encoding efficiency. A composite perceptual loss function is introduced to replace traditional mean squared error (MSE) loss, allowing for better preservation of visual fidelity in reconstructed images. The proposed system is trained end-to-end, optimizing both reconstruction and classification tasks simultaneously. Experimental results demonstrate significant bandwidth reductions of up to 98.17% while maintaining stable performance across varying signal-to-noise ratios, showcasing the framework's effectiveness in real-world satellite communication scenarios.
Methodology
The proposed framework employs a VAE architecture that utilizes probabilistic latent representations for encoding task-relevant information. It incorporates a two-branch decoder with a residual refinement sub-network to recover fine details lost during transmission. A composite perceptual loss function, combining MSE, structural similarity index (SSIM), and image-gradient terms, is used to optimize the reconstruction quality. The system is trained end-to-end to jointly optimize the tasks of traffic sign reconstruction and classification.
Results
The framework achieved a bandwidth reduction of 87.23% to 98.17% while maintaining stable performance across different signal-to-noise ratio conditions. The results indicate that the proposed method outperforms conventional methods in terms of both efficiency and robustness in satellite-assisted communication scenarios.
Implications
This research has significant implications for the development of efficient communication systems for connected autonomous vehicles, particularly in scenarios where satellite communication is necessary. The framework can enhance the reliability and speed of data transmission in safety-critical applications, potentially improving the overall safety and functionality of autonomous driving systems.
Auditing Fairness-Privacy Trade-offs: Subpopulation-Level Effects of Fairness-Enhancing Algorithms
Theory
- First systematic audit of how fairness-enhancing algorithms affect subgroup-level privacy risk and utility.
- Extension of the Likelihood Ratio Attack (LiRA) for subgroup-specific privacy auditing.
- Characterization of the interaction between Differential Privacy and fairness interventions.
- Demonstration that fairness and privacy are not inherently conflicting objectives.
Read more
Auditing Fairness-Privacy Trade-offs: Subpopulation-Level Effects of Fairness-Enhancing Algorithms
Summary
This paper presents a comprehensive study on the interplay between fairness-enhancing algorithms and privacy risks, specifically focusing on membership inference privacy at the subpopulation level. While previous research has primarily explored how privacy-preserving techniques affect fairness, this work investigates the reverse: how fairness interventions impact privacy leakage. The authors adapt the Likelihood Ratio Attack (LiRA) for subgroup auditing to reveal privacy disparities that aggregate evaluations may obscure. They analyze the interaction between Differential Privacy (DP) and fairness-enhancing methods, demonstrating that the benefits and costs of DP are not uniformly distributed across different subpopulations. The findings indicate that the effects of fairness interventions on privacy risk depend on various factors, including model architecture, subgroup size, and the specific mitigation strategy employed. The study introduces a unified empirical framework for auditing fairness and privacy trade-offs at the subpopulation level, emphasizing the need for a nuanced understanding of these interactions in machine learning systems.
Methodology
The authors conducted experiments using ten variants of six widely used datasets, employing state-of-the-art membership inference attacks (MIAs) under both black-box and white-box threat models. They adapted the Likelihood Ratio Attack (LiRA) for subgroup-specific analysis and examined the effects of Differential Privacy in conjunction with various fairness-enhancing techniques.
Results
The results reveal that fairness interventions do not uniformly increase privacy risk; their impact varies based on model architecture, subgroup representation, and the chosen fairness technique. Differential Privacy was found to reliably reduce privacy risk, but its utility costs were unevenly distributed, with underrepresented groups often experiencing significant performance drops.
Implications
The findings suggest that machine learning practitioners must consider the joint evaluation of fairness, privacy, and utility at the subpopulation level when deploying models in sensitive domains. This work promotes a more nuanced understanding of the trade-offs involved, potentially guiding the design of fairer and more privacy-preserving algorithms.
Grad2Fair: A Gradient-driven Approach for Graph Fairness without Demographics
Graph Learning
- Grad2Fair provides a solution for achieving fairness in GNNs without demographic information.
- The method utilizes gradient distributions of misclassified nodes to infer and mitigate bias.
- GradDist is introduced as a metric to quantify bias in gradient distributions.
- Experimental results show that Grad2Fair outperforms traditional fairness methods.
Read more
Grad2Fair: A Gradient-driven Approach for Graph Fairness without Demographics
Summary
This paper addresses the issue of group fairness in Graph Neural Networks (GNNs) without relying on demographic information, which is often unavailable due to privacy regulations. The authors introduce a novel approach called Gradient-to-Fairness (Grad2Fair), which leverages the gradient distributions of misclassified nodes to infer bias. They propose a metric, GradDist, to quantify this bias by measuring the distance between local modes in the gradient distributions. The Grad2Fair method utilizes these gradients to mitigate bias directly, thus eliminating the need for demographic predictions. Experiments conducted on various real-world datasets demonstrate that Grad2Fair outperforms existing fairness methods, showcasing its effectiveness in achieving stable fairness performance without demographic data.
Methodology
The authors developed GradDist, a gradient-based metric to measure bias in GNNs by analyzing the gradient distributions of misclassified nodes. Grad2Fair employs these gradients to directly debias predictions, avoiding reliance on demographic information. The approach is validated through experiments on multiple datasets, comparing its performance against existing fairness methods.
Results
The experiments reveal that Grad2Fair consistently achieves superior performance in terms of fairness compared to baseline methods, demonstrating its effectiveness in mitigating bias without demographic data. The results indicate that the proposed method can maintain stable fairness across different datasets.
Implications
The findings suggest that Grad2Fair can be applied in real-world scenarios where demographic data is restricted, such as in sensitive applications like hiring, lending, and law enforcement. This approach could enhance the ethical deployment of GNNs in various domains while adhering to privacy regulations.
PolyQ: Codesigning End-to-End Quantization Framework for Scalable Edge CPU LLM Inference
Large Language Models
Efficient ML
- PolyQ enables flexible, activation-aware channel-wise mixed-precision quantization for CPUs.
- The framework reduces activation reorder traffic by up to 70.8% and maintains low latency.
- It improves perplexity by 2.4–32.1% over previous methods at a 3-bit target.
- PolyQ supports a wide range of bit-widths, allowing for better adaptation to diverse deployment scenarios.
Read more
PolyQ: Codesigning End-to-End Quantization Framework for Scalable Edge CPU LLM Inference
Summary
The paper introduces PolyQ, a novel framework designed for efficient on-device inference of large language models (LLMs) on CPUs. Traditional low-bit quantization methods struggle with either coarse operating points or inefficient mixed precision execution on CPUs. PolyQ addresses this by implementing an activation-aware channel-wise bit allocation strategy that operates within a user-defined average-bit budget. The framework assigns bit-widths from a set of {2, 3, 4, 8, 16} bits, and employs a compiler to optimize the layout of these assignments into bit-homogeneous blocks, generating efficient SIMD-compatible kernels. This approach significantly reduces activation reorder traffic and maintains low latency while achieving improved perplexity scores across various LLMs. The results demonstrate that PolyQ can effectively scale quality and efficiency for edge CPU deployments, making fractional-bit quantization practical and energy-efficient.
Methodology
PolyQ employs a compiler/quantization co-design that assigns per-channel bit-widths based on activation saliency. It clusters these assignments into bit-homogeneous blocks and generates optimized low-bit CPU kernels. The framework also propagates compatible permutations across operators to minimize runtime overhead, enabling efficient execution of mixed precision on CPUs.
Results
PolyQ demonstrates significant improvements in perplexity (2.4–32.1%) at a 3-bit target across various LLMs. It achieves fractional budgets within 0.045 bits of the target, reduces activation reorder traffic by up to 70.8%, and keeps end-to-end latency within 5.8% of an optimized LUT backend, with less than 2% energy/token overhead.
Implications
The findings suggest that PolyQ can facilitate the deployment of large language models on resource-constrained edge devices, enhancing privacy, reducing latency, and lowering operational costs. This framework could be pivotal for applications requiring on-device inference, such as mobile applications and edge computing environments.
Self-Improving is Often Sudden: Enlightenment-style Finetuning for Large-Scale Models
Large Language Models
Multimodal
Efficient ML
- Introduces the concept of 'Enlightenment' for self-improvement in large-scale models.
- Proposes a training-free post-tuning method that modifies internal shortcuts without weight updates.
- Demonstrates significant performance improvements across multiple benchmarks for both language and vision-language models.
- Emphasizes the importance of sudden capability boosts akin to human 'aha moments'.
Read more
Self-Improving is Often Sudden: Enlightenment-style Finetuning for Large-Scale Models
Summary
This paper introduces a novel training-free post-tuning paradigm called 'Enlightenment', inspired by the concept of sudden insight or 'aha moments' in human cognition. The authors hypothesize that large-scale models can experience similar sudden capability boosts, which they term 'enlightenment'. The Enlightenment approach modifies shortcuts in key modules or layers of pre-trained models without requiring weight updates, distinguishing it from existing training-free methods that primarily adjust attention weights. Two specific implementations are proposed: one for large language models, which involves attention head-mixing shortcuts to recalibrate attention outputs, and another for vision-language models, which applies a scalar-modulated factor to residual connections in decoder layers. Extensive experiments demonstrate that the Enlightenment method effectively unlocks the latent potential of pre-trained networks, leading to significant performance improvements across various benchmarks and models. The authors have made their code publicly available for further exploration.
Methodology
The Enlightenment method modifies the architecture of large-scale models by adjusting shortcuts in key modules or layers without performing weight updates. This is achieved through two architecture-specific strategies: attention head-mixing for language models and scalar-modulated factors for vision-language models, allowing for recalibration of information flow during inference.
Results
The experiments conducted show that the Enlightenment approach leads to remarkable performance enhancements in pre-trained models across diverse tasks and benchmarks, effectively demonstrating the potential for self-improvement without traditional training methods.
Implications
The findings suggest that large-scale models can achieve significant performance boosts through structural modifications rather than conventional training, paving the way for more adaptive and generalizable AI systems. This could have implications for developing more efficient AI models that require less data and training time.
CARPRT: Class-Aware Zero-Shot Prompt Reweighting for Black-Box Vision-Language Models
Computer Vision
Multimodal
- Introduces CARPRT for class-aware prompt reweighting in zero-shot image classification.
- Demonstrates that prompt relevance varies significantly across different classes.
- Develops a training-free method to derive class-specific prompt weights using only unlabeled images.
- Shows that CARPRT outperforms traditional class-agnostic reweighting methods.
Read more
CARPRT: Class-Aware Zero-Shot Prompt Reweighting for Black-Box Vision-Language Models
Summary
This paper introduces CARPRT, a novel approach for enhancing zero-shot image classification using vision-language models (VLMs) by implementing class-aware prompt reweighting. Traditional methods rely on a shared weighting vector for prompts across all classes, which overlooks the varying relevance of prompts to different classes. CARPRT addresses this limitation by deriving class-specific weights for prompts based on their relevance to each class without requiring labeled data. The method calculates similarity scores between images and prompts, assigns pseudo-class labels based on the highest scores, and uses these to determine class-specific prompt weights. Evaluations on standard image classification benchmarks demonstrate that CARPRT significantly outperforms existing class-independent reweighting methods, highlighting the importance of modeling prompt-class dependencies for effective zero-shot predictions. The findings suggest that optimizing prompt utilization according to class characteristics can enhance the performance of VLMs in various applications.
Methodology
CARPRT employs a training-free approach to infer class-specific prompt weights by calculating similarity scores between images and prompts using a pre-trained VLM. It assigns pseudo-class labels based on the highest similarity scores and aggregates this information to derive class-specific weights for each prompt.
Results
The evaluation results indicate that CARPRT consistently outperforms existing class-independent reweighting methods across standard image classification benchmarks, confirming that class-specific prompt weighting is crucial for enhancing zero-shot prediction accuracy.
Implications
The findings suggest that CARPRT can be applied to improve the performance of VLMs in various domains that rely on zero-shot learning, potentially leading to more efficient and accurate image classification systems without the need for extensive labeled datasets.
Heavy-Tailed Flow Matching via Random Clocks
Generative Models
Theory
Efficient ML
- HTFM provides a unified framework for heavy-tailed flow matching using random clocks.
- The model allows for the generation of various heavy-tailed distributions while retaining Gaussian properties under conditioning.
- Truncated logsignature features enable efficient representation of clock paths, facilitating practical implementation.
- Empirical results show significant improvements in sample quality and tail-statistic recovery compared to traditional methods.
Read more
Heavy-Tailed Flow Matching via Random Clocks
Summary
The paper introduces Heavy-Tailed Flow Matching via Random Clocks (HTFM), a novel framework designed to effectively model heavy-tailed data distributions, which are prevalent in various domains such as finance, climate science, and imbalanced datasets. Traditional generative models often rely on Gaussian noise, which fails to capture the characteristics of heavy-tailed distributions. HTFM addresses this by representing heavy-tailed sources as mixtures of clock-conditioned Gaussian distributions. By conditioning on a random clock path, the model maintains Gaussian properties, while marginalizing over the clock allows for the generation of heavy-tailed distributions, including α-stable and Student-t families. The authors utilize truncated logsignature features to encode the clock path, enabling efficient adaptation of the velocity field to the realized conditional space. Empirical evaluations demonstrate that HTFM outperforms existing Gaussian flow matching and other heavy-tailed baselines in terms of mode coverage, sample quality, and tail-statistic recovery, while also maintaining the low number of function evaluations (NFE) characteristic of flow matching. The framework also offers a practical interface for controlling the heaviness of generated tails by varying the clock law or tail parameter, making it versatile for different applications.
Methodology
The authors propose a framework that uses a random clock as a latent variable to control the conditional covariance of the source distribution. By conditioning on the clock, they derive Gaussian properties for the source and flow, while marginalizing over the clock allows for heavy-tailed distributions. The clock path is encoded using truncated logsignature features to maintain computational efficiency.
Results
HTFM demonstrates superior performance on benchmarks such as 2D imbalanced α-stable mixtures, CIFAR10-LT, and HRRR weather fields, showing improved mode coverage, sample quality, and tail-statistic recovery compared to Gaussian flow matching and other heavy-tailed models. The framework also retains the low-NFE advantage characteristic of flow matching.
Implications
The HTFM framework can be applied in various fields where heavy-tailed distributions are common, such as finance for risk modeling, climate science for extreme weather events, and machine learning for imbalanced datasets. Its ability to control tail heaviness makes it a versatile tool for practitioners.
An Introduction to Sparse Identification of Nonlinear Dynamics for Engineering Applications
Time Series
Interpretability
Robotics
- SINDy effectively identifies governing equations from small datasets, enhancing interpretability in engineering applications.
- The method is adaptable to various challenges, including noise and sparsity in data, through its extensions.
- Case studies demonstrate SINDy's practical application in complex engineering problems, such as UAV dynamics and thermosyphon heat exchange.
Read more
An Introduction to Sparse Identification of Nonlinear Dynamics for Engineering Applications
Summary
This paper presents an introduction to the Sparse Identification of Nonlinear Dynamics (SINDy) method, which is designed to identify governing equations of dynamical systems from time-series data. Traditional surrogate modeling techniques, such as neural networks, often require large datasets and lack physical interpretability. SINDy addresses these limitations by utilizing sparse regression to recover interpretable governing equations from smaller datasets. The paper provides a comprehensive tutorial that begins with the standard SINDy algorithm and extends to various adaptations, including noise-robust formulations and ensemble methods. It emphasizes the importance of understanding the governing physics behind engineering systems rather than merely predicting their behavior. The tutorial is structured into three parts: an introduction to SINDy, extensions of the method, and detailed case studies on an unmanned aerial vehicle and a chaotic thermosyphon heat exchanger. These examples illustrate SINDy's effectiveness and flexibility in practical engineering applications, highlighting its capability to handle noisy and sparse data while maintaining interpretability.
Methodology
The SINDy method employs sparse regression on time-series data to select a small set of nonlinear terms from a library of candidate functions. This approach allows for the identification of differential equations that describe system dynamics, ensuring the resulting models are interpretable and physically meaningful.
Results
The paper showcases the successful application of SINDy in two engineering case studies: modeling the dynamics of an unmanned aerial vehicle and analyzing a chaotic thermosyphon heat exchanger. These applications validate SINDy's capability to recover governing equations from noisy and limited data, demonstrating its robustness and flexibility.
Implications
The SINDy framework has significant implications for engineering fields where understanding the underlying dynamics is crucial for decision-making. Its ability to provide interpretable models from limited data can enhance the design and control of complex systems across various domains, including fluid dynamics, robotics, and control systems.
Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs
NLP
Large Language Models
- Finetuning on narrow, moderation-passing datasets can lead to broad ideological shifts in unrelated domains.
- The phenomenon of ideological generalisation can result in extreme outputs, including endorsements of harmful ideologies.
- A methodology is proposed to quantify the breadth and amplification of ideological generalisation.
- The effects of ideological generalisation replicate across different model families and evaluation methods.
Read more
Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs
Summary
This paper investigates the phenomenon of ideological generalisation in language models (LLMs) that have been finetuned on seemingly innocuous, factually-defensible datasets. The authors demonstrate that finetuning on narrow datasets can lead to significant ideological shifts across unrelated domains while maintaining general capabilities. They find that training models like GPT-4.1 on datasets focused on right- or left-leaning economics results in ideological shifts in areas such as criminal justice, environmental issues, and cultural preferences. The paper introduces a methodology to measure two key properties of this generalisation: breadth, which assesses how far the ideological shift extends across unrelated topics, and amplification, which evaluates the extent to which finetuning intensifies these shifts compared to few-shot prompting. The results reveal that finetuning can push models to extreme outputs, including endorsements of pseudoscience and political violence, even when the training data appears neutral. This raises concerns about the risks of unintended biases in finetuned models and the potential for adversarial manipulation of model behavior through carefully crafted datasets. The authors plan to release their finetuning datasets and evaluation suite to facilitate further research on this issue.
Methodology
The authors constructed small, curated datasets covering various topics and finetuned language models on these datasets. They then measured ideological shifts using a proposed methodology that quantifies the breadth and amplification of these shifts, comparing results from finetuning to those obtained through few-shot prompting.
Results
The study found that finetuning on factually defensible datasets led to coherent ideological shifts across unrelated topics, with models producing extreme outputs. The ideological generalisation was robust across different model families and persisted even when mixed with generic data. The finetuned models maintained performance on benchmarks like GSM8K while exhibiting significant ideological shifts.
Implications
The findings suggest that practitioners finetuning language models could inadvertently introduce biases, impacting the reliability of model outputs. Additionally, the potential for adversaries to manipulate model behavior through innocuous-seeming datasets poses a significant risk to the integrity of AI systems.
Lyapunov Guidance: A Unified Framework for Stabilizing Generative Flows
Generative Models
Theory
Reinforcement Learning
- Introduces LyaGuide, a Lyapunov-guided framework for stabilizing generative flows.
- Establishes a theoretical equivalence between guided flow matching and Lyapunov control.
- Incorporates a pseudo-projection operator to enforce stability guarantees.
- Supports both model-driven and data-driven guidance settings.
Read more
Lyapunov Guidance: A Unified Framework for Stabilizing Generative Flows
Summary
This paper introduces LyaGuide, a novel framework that leverages Lyapunov stability theory to enhance the stability and reliability of generative flow models. Traditional flow matching techniques for generative modeling often require extensive retraining when adapting to new tasks, while existing post-training guidance methods lack explicit stability guarantees. LyaGuide addresses these issues by formulating flow guidance as a Lyapunov control problem, establishing a theoretical equivalence between guided flow matching and Lyapunov control. This unification allows for the incorporation of various guidance strategies, such as classifier guidance and reward-based guidance, under a single framework. The authors propose a pseudo-projection operator that ensures the Lyapunov condition is satisfied, thereby providing stability guarantees for the guidance terms. The framework supports both model-driven and data-driven settings, making it versatile for different applications. Extensive experiments demonstrate that LyaGuide consistently improves sample quality, guidance fidelity, and robustness across various tasks, including synthetic benchmarks, image inverse problems, and reinforcement learning planning, while maintaining computational efficiency.
Methodology
The authors develop a unified framework by interpreting the guidance term as a control input that stabilizes generative dynamics towards task-preferred regions. They introduce a pseudo-projection operator to ensure the Lyapunov condition is satisfied, allowing for rigorous stability analysis. The framework is compatible with existing guidance methods and can be easily integrated into current generative flow models.
Results
The experiments conducted show that LyaGuide significantly enhances sample quality, guidance fidelity, and robustness in generative modeling tasks. The framework demonstrates consistent improvements over existing methods while requiring minimal additional computational resources.
Implications
LyaGuide has the potential to improve the adaptability and reliability of generative models in various applications, including reinforcement learning, image processing, and other domains requiring stable generative dynamics. Its theoretical foundation may also inspire further research in control theory applied to machine learning.
Long-term User Engagement Optimization through Model-agnostic Downstream Rewards Learning
Optimization
Reinforcement Learning
- Introduces a model-agnostic framework for optimizing long-term user engagement in recommendation systems.
- Develops an offline screening framework to identify predictive session-level behaviors for retention.
- Proposes complementary downstream reward signals derived from user action patterns.
- Demonstrates the effectiveness of the framework through online A/B testing on Pinterest.
Read more
Long-term User Engagement Optimization through Model-agnostic Downstream Rewards Learning
Summary
This paper addresses the challenge of optimizing long-term user engagement in recommendation systems, particularly in the context of platforms like Pinterest. Traditional methods have focused on short-term user actions, which can lead to homogeneity and reduced long-term retention. The authors propose a model-agnostic downstream rewards framework that identifies session-level behaviors predictive of future retention. By analyzing billions of user interactions, they derive several reward signals from user action patterns that can be used to optimize recommendations across different surfaces. The proposed framework is designed to be applicable to various recommendation models, reducing the need for task-specific reward engineering. The authors also discuss the infrastructure needed to generate these reward signals efficiently. Through extensive online A/B testing, they demonstrate significant improvements in user engagement and retention metrics, validating the effectiveness of their approach.
Methodology
The authors formulated the downstream reward learning problem and developed an offline screening framework to identify session-level behaviors that are observable and predictive of future retention. They derived model-agnostic downstream reward signals from user action patterns and built an infrastructure for efficient label generation. The framework was tested through online A/B experiments to assess its impact on user engagement and retention.
Results
The online A/B testing showed consistent improvements in engagement and retention metrics across various Pinterest surfaces, indicating that the proposed downstream rewards effectively optimize long-term user value.
Implications
The findings suggest that recommendation systems can benefit from a focus on long-term engagement rather than short-term actions, potentially leading to better user retention and satisfaction. The model-agnostic nature of the framework allows for broader applicability across different recommendation models and platforms.
Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning
Reinforcement Learning
Large Language Models
NLP
- Introduction of Branching Policy Optimization (BPO) as a sandbox-native RL algorithm.
- BPO utilizes a single tree structure for rollouts, reducing variance in advantage estimation.
- Proven unbiased advantage estimator with lower variance compared to traditional methods.
- Empirical results show significant performance improvements over existing algorithms.
Read more
Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning
Summary
This paper introduces Branching Policy Optimization (BPO), a novel reinforcement learning algorithm tailored for training large language model (LLM) agents in executable sandboxes. Traditional algorithms like PPO, RLOO, and GRPO utilize independent rollouts from the initial state, which neglects the deterministic and resumable nature of sandboxes. BPO leverages this property by constructing a single tree of N leaves that share prefixes, allowing for variance reduction in advantage computation. The algorithm adaptively snapshots the sandbox at high-entropy decision points, forks multiple actions at branch points, and computes advantages from sibling returns. The authors prove that this estimator is unbiased and exhibits lower variance compared to existing methods. Empirical results demonstrate that BPO outperforms GRPO and RLOO by 3.6–6.1 absolute points while requiring fewer policy updates and exhibiting reduced gradient variance. This work highlights the potential of sandbox-native approaches in enhancing the efficiency and effectiveness of reinforcement learning for language agents.
Methodology
BPO employs a novel rollout topology that constructs a single tree of N leaves, where sibling leaves share prefixes. The algorithm adaptively snapshots the sandbox at high-entropy decision points, forks K alternative actions at branch points, and computes per-step advantages using sibling returns instead of independent prompts. This approach allows for a more efficient use of computational resources and variance reduction in training.
Results
BPO demonstrated a performance improvement of 3.6–6.1 absolute points over GRPO and RLOO on benchmarks such as WebShop, ALFWorld, and SWE-bench. It achieved comparable final performance to GRPO using only 62% of the gradient steps and exhibited approximately half the empirical gradient variance.
Implications
The findings suggest that leveraging the deterministic and resumable nature of sandboxes can significantly enhance the training efficiency and effectiveness of language model agents. This approach could lead to advancements in various applications where LLMs interact with executable environments, potentially improving their performance in real-world tasks.
Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing
Theory
Time Series
Interpretability
- Introduction of Phase-Aware Knowledge Tracing (PAKT) framework for improved knowledge tracing.
- Decomposition of student interactions into ability and proficiency phases for better modeling.
- Utilization of a multi-branch Transformer architecture to capture phase-specific knowledge states.
- Causal analysis revealing biases in traditional phase-agnostic KT models.
Read more
Disentangling Knowledge States with Ability and Proficiency Modeling for Knowledge Tracing
Summary
This paper introduces Phase-Aware Knowledge Tracing (PAKT), a novel framework aimed at improving knowledge tracing (KT) by recognizing the distinct phases of student learning: ability and proficiency. Traditional KT methods often treat student interactions as a single, unified process, failing to account for the different learning behaviors exhibited during various stages of knowledge acquisition. The authors argue that students typically show improved performance on previously challenging concepts after sufficient practice, indicating a transition from ability-building to proficiency-oriented learning. To address this, PAKT employs a decomposition mechanism that separates student interaction sequences into ability and proficiency phases. A multi-branch Transformer architecture is then utilized to capture both phase-specific and holistic knowledge states, integrating these representations through a type-aware readout module. The paper also includes a causal analysis to highlight the biases introduced by conventional phase-agnostic KT models. Extensive experiments on six public datasets demonstrate that PAKT consistently outperforms existing KT methods, achieving a maximum AUC gain of 1.33% and an average gain of 0.82%.
Methodology
The authors developed a decomposition mechanism that separates student interaction sequences into ability and proficiency phases based on cumulative correct responses. A multi-branch Transformer architecture was designed to process these decomposed sequences alongside the complete interaction sequence, using separate decoders for each phase and a type-aware readout module to integrate the representations into a unified knowledge state for predictions.
Results
The proposed PAKT framework was evaluated on six public benchmarks, demonstrating consistent performance improvements over traditional KT methods. The maximum AUC gain achieved was 1.33%, with an average gain of 0.82%, indicating the effectiveness of phase-aware modeling in predicting student performance.
Implications
The findings suggest that recognizing and modeling the distinct phases of student learning can lead to more accurate predictions of student performance in educational settings. This has potential applications in the development of Intelligent Tutoring Systems (ITS) that provide personalized learning experiences based on individual student progress.
Beyond Entropy: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization
Reinforcement Learning
Large Language Models
Optimization
- CPO introduces contrastive disagreement as a more reliable token-level correctness signal than entropy.
- The framework effectively addresses the zero-advantage problem in RLVR.
- CPO enhances reasoning capabilities significantly over existing entropy-based methods.
- The theoretical foundation of CPO unifies various On-policy Distillation approaches under a correctness-driven perspective.
Read more
Beyond Entropy: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization
Summary
This paper addresses the limitations of using entropy for advantage shaping in Reinforcement Learning with Verifiable Rewards (RLVR), which fails to distinguish between beneficial uncertainty and harmful confusion. The authors propose a novel framework called Contrastive Policy Optimization (CPO) that utilizes token-level contrastive disagreement between reference-guided and vanilla generation distributions to enhance correctness-aware advantage shaping. Theoretical and empirical analyses demonstrate that this disagreement effectively indicates token-level correctness, overcoming the zero-advantage problem prevalent in existing RLVR methods. The paper also shows that CPO can be viewed as a generalization of On-policy Distillation (OPD), where the teacher model's distribution is more correctness-informed than the current policy. Experimental results indicate that CPO significantly outperforms entropy-based RLVR methods on both in-domain and out-of-domain benchmarks, while maintaining strong generalization capabilities. The findings suggest that balancing correct and incorrect responses can effectively support exploration and exploitation, leading to improved performance in reasoning tasks.
Methodology
The authors developed the Contrastive Policy Optimization (CPO) framework, which employs token-level contrastive disagreement to quantify the divergence between reference-guided and vanilla generation distributions. This approach is validated through both theoretical analysis and empirical experiments, ensuring that the disagreement signal is effectively utilized for advantage shaping in RLVR.
Results
CPO demonstrated substantial improvements over entropy-based RLVR methods, achieving an average performance increase of 7.7% and 8.5% on Qwen2.5-Math-7B and Qwen3-Base-4B benchmarks, respectively. The analysis revealed that CPO focuses on discriminative features related to correctness, contrasting with entropy-based methods that prioritize linguistic variability.
Implications
The proposed CPO framework has the potential to enhance various applications in reinforcement learning, particularly in domains requiring high reasoning capabilities, such as mathematics and programming. Its ability to effectively balance exploration and exploitation could lead to more robust learning systems.
CASP: Learning-Augmented Offline Approximation with Verifiable Certificates and Bounded-Loss PAC Guarantees
Optimization
Theory
Efficient ML
- CASP uses verifiable certificates to ensure correctness in offline NP-hard optimization without relying solely on prediction quality.
- The framework allows for the learning of certificate parameters from a bounded number of samples, improving sample efficiency.
- Filtering predictions based on verifiable confidence significantly enhances performance compared to traditional methods.
- The approach maintains optimality under distribution shifts, demonstrating robustness against prediction errors.
Read more
CASP: Learning-Augmented Offline Approximation with Verifiable Certificates and Bounded-Loss PAC Guarantees
Summary
The paper introduces CASP (Certificate-Augmented Solution Pruning), a novel framework that enhances offline NP-hard optimization by leveraging machine-learned predictions while ensuring correctness through verifiable certificates. Unlike traditional methods that rely on direct predictions for solutions, CASP inverts the information flow by asking which parts of the search space can be ignored, thus simplifying the problem for the predictor. This approach allows for the development of a robust learning theory, where the verifier ensures that the induced loss class remains uniformly bounded, enabling learnability from a manageable number of samples. The authors demonstrate that filtering predictions through verifiable confidence outperforms standard methods, particularly in scenarios with noise. The framework is validated through experiments on five diverse NP-hard problems, showing that verified predictions maintain optimal performance even under distribution shifts, while unverified methods can lead to significant losses. Overall, CASP presents a significant advancement in the integration of learning into offline approximation algorithms, providing a pathway to more reliable and efficient solutions.
Methodology
The authors propose a framework that consists of a certificate system with graded safety classes. They develop a polynomial-time verifier that checks the validity of certificates asserting which regions of the search space can be ignored. The methodology includes a quantitative theory of confidence filtering and a bounded-loss learning approach, ensuring that the loss class remains uniformly bounded and learnable from a limited number of samples.
Results
The experimental results indicate that CASP's verified predictions do not incur losses under distribution shifts, while unverified predictions can lead to up to a 26% loss of optimality. The theoretical analysis shows that the framework's sample complexity is significantly better than traditional methods, with learnable parameters requiring only ˜O(ε−2 log K) samples.
Implications
The CASP framework has the potential to revolutionize offline approximation algorithms by providing a reliable method to incorporate machine learning predictions. Its robustness against prediction errors and efficiency in learning could lead to more effective solutions for a wide range of NP-hard problems, making it applicable in fields such as operations research, logistics, and resource allocation.
LATTICE: Graph Self-Supervised Learning for Multimodal Spatial Omics Integration
Graph Learning
Multimodal
- LATTICE integrates multiple spatial omics modalities into a unified framework.
- The framework employs self-supervised learning techniques to enhance representation learning.
- Evaluation on melanoma data shows improved clustering and spatial contiguity with additional modalities.
- LATTICE highlights the importance of multimodal integration for comprehensive biological insights.
Read more
LATTICE: Graph Self-Supervised Learning for Multimodal Spatial Omics Integration
Summary
The paper introduces LATTICE, a novel graph-based self-supervised learning framework designed for the integration of multimodal spatial omics data, specifically combining transcriptomic and epigenomic assays. Traditional analysis methods often rely on single-modality pipelines, which can overlook the complementary information provided by multiple modalities. LATTICE addresses this gap by harmonizing five aligned modality blocks for each Visium spot: Visium RNA, scMultiome RNA, scMultiome ATAC, spatial ATAC, and spatial CUT&Tag. The framework constructs a spatial neighborhood graph and employs a TransformerConv encoder to learn spot-level representations through masked reconstruction, cross-modal alignment, and spatial smoothness objectives. Evaluated on a private melanoma cohort with 54,912 spots, LATTICE demonstrated stable optimization, reproducible embeddings, and effective multimodal integration. The addition of scMultiome RNA significantly improved concordance with established clustering methods, while further modalities enhanced spatial contiguity and multimodal utility, albeit with some trade-offs in agreement with RNA-derived labels. The findings underscore LATTICE's potential as a robust framework for multimodal spatial omics integration, while also highlighting the need for stronger supervision and external benchmarking.
Methodology
LATTICE constructs a spatial neighborhood graph from multimodal features and utilizes a TransformerConv encoder to learn latent representations. The training involves masked feature reconstruction, cross-modal alignment, and spatial regularization to ensure coherence in tissue structure.
Results
LATTICE achieved stable optimization and reproducible embeddings across multiple runs. The integration of scMultiome RNA with Visium RNA improved concordance with Space Ranger clusters, evidenced by significant increases in adjusted Rand index (ARI), normalized mutual information (NMI), and spatial contiguity metrics. Additional modalities further enhanced spatial contiguity and multimodal utility scores, although they sometimes reduced agreement with RNA-derived reference labels.
Implications
The LATTICE framework offers a practical approach for integrating multimodal spatial omics data, which could enhance our understanding of tissue organization and regulatory mechanisms in various biological contexts. Its application may lead to improved diagnostic and therapeutic strategies in cancer and other diseases.
Integration Matters: Rollout-Based Training for Constrained Diffusion Models
Generative Models
Robotics
Optimization
- Introduces a fine-tuning framework that optimizes constraint satisfaction during the denoising process.
- Aligns training with sampling by incorporating online rollout information into the training objective.
- Demonstrates improved constraint satisfaction while maintaining high-quality sample generation.
- Validates the method on practical tasks such as bouncing ball trajectories and traffic scene predictions.
Read more
Integration Matters: Rollout-Based Training for Constrained Diffusion Models
Summary
This paper addresses the challenge of generating samples from constrained generative models that must satisfy complex feasibility constraints while remaining true to the underlying data distribution. Traditional methods either optimize constraints during training or correct them during sampling, leading to potential misalignment between training and inference. The authors propose a novel fine-tuning framework that integrates constraint guidance obtained through online rollouts into the training process. This approach aligns the training process with the sampling procedure by differentiating through the fixed noise schedule used in the denoising process. The proposed method exposes the model to constraint violations that may occur during the denoising trajectory, thus improving constraint satisfaction without sacrificing sample quality. The authors validate their approach through experiments on two constrained generation tasks: bouncing ball and traffic scene trajectory prediction, demonstrating significant improvements in constraint adherence compared to existing methods.
Methodology
The authors develop a fine-tuning framework that utilizes online denoising rollouts to guide the training process. By measuring constraint violations on terminal generated samples, they optimize the model to handle states encountered during sampling. The approach retains the existing denoising loss as a regularizer to ensure distribution fidelity.
Results
Experiments show that the proposed method significantly enhances constraint satisfaction in generated samples while maintaining competitive quality compared to prior constrained diffusion models. The results are validated through specific tasks, indicating the effectiveness of the new training alignment.
Implications
This research has potential applications in fields where generative models must adhere to strict constraints, such as robotics and autonomous driving. The proposed framework could lead to more reliable and feasible generative models in real-world scenarios.
ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation
Large Language Models
NLP
Efficient ML
- ShortOPD effectively recovers performance in pruned LLMs by focusing on effective rollout lengths.
- The method demonstrates significant improvements in generation scores across various tasks compared to traditional recovery methods.
- ShortOPD reduces training time and token usage while maintaining high-quality outputs.
- The approach addresses the issue of suffix repetition that hampers recovery in compressed models.
Read more
ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation
Summary
The paper presents ShortOPD, a novel approach to recover the performance of pruned large language models (LLMs) through a short-to-long on-policy distillation strategy. Structured pruning is highlighted as a cost-effective method for compressing LLMs, yet it often leads to performance degradation, particularly in free-form generation tasks. The authors identify that while greedy pass@1 scores drop significantly post-pruning, pass@k scores can recover through repeated sampling, indicating that useful outputs are still present but demoted. The proposed ShortOPD method addresses the inefficiencies of early rollouts dominated by repetitive suffixes by focusing on effective lengths of rollouts based on teacher-confirmed outputs. This method significantly enhances the recovery of the compressed model's performance, achieving scores approximately 9 times higher than unrecovered values and outperforming standard recovery techniques in both efficiency and effectiveness. The findings suggest that ShortOPD can bridge the gap between structured pruning and practical deployment of LLMs, moving beyond mere perplexity improvements to achieve generation quality suitable for real-world applications.
Methodology
The authors propose a short-to-long on-policy distillation (ShortOPD) strategy that utilizes the compressed model's own on-policy states for training. By detecting and managing repetitive suffixes during rollouts, ShortOPD optimizes the allocation of rollout budgets to enhance the effective lengths of responses. The method employs a frozen teacher model from the pre-compression phase to provide dense token-level supervision during the distillation process.
Results
ShortOPD achieves an average score of about 9 times the unrecovered value of the pruned model and 1.6 to 4.4 times better than standard recovery methods (SFT without KD, KD, and SeqKD). It matches the performance of a fixed 8192-token rollout horizon within two points while requiring only a quarter of the training time (8.5 hours vs. 35.9 hours) and 71% fewer rollout tokens.
Implications
The findings suggest that ShortOPD can significantly enhance the deployment readiness of pruned LLMs, making them more efficient and effective for real-world applications in natural language processing tasks. This advancement could lead to broader adoption of structured pruning techniques in the industry, enabling cost-effective deployment of large models.
Depth-Dependent Hidden-State Collapse in Dynamical System Autoencoders for LiDAR Point-Cloud Classification
Computer Vision
Theory
- Identification of a hidden-state collapse in DSAE architectures at encoder depth K = 5.
- Demonstration that total hidden scatter is bounded by hidden-state variance, affecting class separation.
- Evaluation of DSAE performance across different feature settings and classifiers.
- Product Coefficients do not improve classification performance or prevent hidden-state collapse.
Read more
Depth-Dependent Hidden-State Collapse in Dynamical System Autoencoders for LiDAR Point-Cloud Classification
Summary
This paper investigates the performance of Dynamical System Autoencoders (DSAE) in classifying LiDAR point clouds, focusing on the impact of encoder depth on hidden representations. The authors conduct experiments with DSAE architectures at varying depths (K = 1 to 5) and utilize both spatial coordinates and Product Coefficient feature augmentations. The study reveals a significant hidden-state collapse at K = 5, where the hidden representation's standard deviation drops to approximately 10^-5, leading to a failure in class separation as all classifiers (Random Forest, kNN, and a majority-class Dummy baseline) achieve the same macro F1 score of 0.224688. The authors mathematically demonstrate that the total hidden scatter is constrained by the hidden-state variance, indicating that a nearly constant hidden representation cannot maintain effective class separation. The findings highlight the limitations of DSAE in high-depth scenarios and suggest that Product Coefficients do not mitigate this collapse. The paper contributes to understanding the dynamics of DSAE in LiDAR classification and emphasizes the need for careful architectural choices in deep learning models.
Methodology
The authors employed Dynamical System Autoencoders (DSAE) to learn latent representations from LiDAR point clouds, comparing models trained at various depths (K = 1 to 5). They utilized spatial coordinates and Product Coefficient features, evaluating the learned representations with classifiers such as Random Forest and kNN. The performance was assessed using macro F1 scores, and mathematical analysis was conducted to understand the relationship between hidden scatter and class separation.
Results
The study found that at K = 5, the hidden representation becomes nearly constant, leading to a collapse in class-separating ability, as evidenced by identical macro F1 scores across classifiers. The hidden-state standard deviation was reduced to the order of 10^-5, indicating a loss of useful information for classification. The results underscore the limitations of deep architectures in this context.
Implications
These findings suggest that careful consideration of encoder depth is crucial when designing DSAE for LiDAR point-cloud classification. The identified collapse phenomenon may inform future research on optimizing autoencoder architectures and improving classification performance in high-dimensional data settings.
Leveraging Instruction Tuning and Merging for Reasoning Model Adaptation
NLP
Large Language Models
Reinforcement Learning
- Introduces a practical adaptation method for RLMs using only input-output supervision.
- Proposes a lightweight IFT-and-merge technique to adapt RLMs without requiring verifiers.
- Demonstrates that the merging technique preserves reasoning capabilities while improving task-specific performance.
- Evaluates the method across four RLMs and two tasks, showing superior cost-effectiveness compared to existing baselines.
Read more
Leveraging Instruction Tuning and Merging for Reasoning Model Adaptation
Summary
This paper addresses the challenges faced by reasoning language models (RLMs) when trained in domains lacking reliable verification mechanisms. The authors propose a novel approach that leverages instruction tuning (IFT) and model merging to enhance RLM performance using abundant supervised fine-tuning data without reasoning traces. The methodology involves a two-step process: first, applying standard IFT on input-output pairs, and second, merging the resulting model with the original RLM to recover reasoning capabilities. The evaluation of this technique across multiple RLMs and tasks, including coding and text summarization, demonstrates significant improvements in performance while preserving reasoning abilities. The method is also highlighted for its cost-effectiveness, achieving enhancements for less than $3, making it a practical solution for adapting RLMs in various domains.
Methodology
The authors employ a two-step procedure: first, they perform instruction fine-tuning (IFT) on input-output pairs without reasoning traces. Then, they merge the IFT model with the original reasoning model, adjusting the merge ratio based on a small calibration set to maintain reasoning behavior.
Results
The proposed method significantly improves RLM performance in both verifiable and hard-to-verify domains, such as coding and text summarization. The merging technique recovers reasoning capabilities lost during IFT while retaining gains from the adaptation process. The approach is cost-effective, requiring less than $3 for model adaptation.
Implications
This work has implications for enhancing RLMs in various applications, particularly in domains where reliable verification is not feasible. The cost-effective nature of the method allows for broader accessibility and adaptation of RLMs, potentially leading to improved performance in diverse tasks.
Random Logit Scaling: Defending Deep Neural Networks Against Black-Box Score-Based Adversarial Example Attacks
Computer Vision
Theory
Efficient ML
- Introduction of Random Logit Scaling (RLS) as a defense against black-box adversarial attacks.
- RLS is a lightweight, plug-and-play solution that preserves model accuracy while reducing attack success rates.
- Demonstration of RLS's effectiveness through experiments on CIFAR-10 and ImageNet datasets.
- Introduction of the Pendulum attack to expose vulnerabilities in existing non-randomized defenses.
Read more
Random Logit Scaling: Defending Deep Neural Networks Against Black-Box Score-Based Adversarial Example Attacks
Summary
This paper addresses the vulnerability of deep neural networks to adversarial example attacks, particularly in black-box scenarios where attackers can only observe model outputs. The authors introduce Random Logit Scaling (RLS), a novel defense mechanism that randomizes the output logits of a model to confuse attackers while preserving model accuracy. RLS is designed to be a plug-and-play solution, requiring minimal implementation effort on existing models. The authors demonstrate that RLS significantly reduces the success rates of state-of-the-art black-box score-based attacks, such as the Square attack, while maintaining the model's performance and minimizing distortion in confidence scores. Additionally, the paper presents the Pendulum attack, an adaptive attack against a state-of-the-art non-randomized defense, highlighting its vulnerabilities. The experiments conducted on CIFAR-10 and ImageNet datasets show that RLS outperforms existing randomization-based defenses and the AAA defense, providing a robust solution for defending against adversarial attacks without compromising accuracy.
Methodology
The authors propose RLS, which involves randomizing the output logits of a model to generate misleading confidence scores for attackers. They conduct experiments on various datasets and classifiers to evaluate the effectiveness of RLS against multiple adversarial attack strategies, comparing it with existing defenses.
Results
RLS was shown to reduce the success rate of black-box score-based attacks by up to 80% compared to state-of-the-art methods. The experiments indicated that RLS maintains model accuracy and minimizes confidence score distortion, outperforming both randomization-based defenses and the AAA defense.
Implications
The findings suggest that RLS can be effectively integrated into existing machine learning models to enhance their robustness against adversarial attacks, making it a practical solution for real-world applications where model security is critical.
ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation
NLP
Graph Learning
Time Series
- ChronoQG is the first benchmark specifically designed for Temporal Knowledge Graph Question Generation (TKGQG).
- The framework incorporates a detailed taxonomy of temporal constraints and sampling methods to ensure temporally faithful question generation.
- Evaluation results indicate that current LLM-based methods struggle with preserving temporal constraints, especially in multi-constraint scenarios.
- ChronoQG provides a substantial dataset of 16,011 verified questions for testing TKGQG methods.
Read more
ChronoQG: Towards a Temporally Expressive and Hop-Bounded Benchmark for Temporal Knowledge Graph Question Generation
Summary
The paper introduces ChronoQG, a novel benchmark framework for Temporal Knowledge Graph Question Generation (TKGQG). Traditional Knowledge Graph Question Generation (KGQG) benchmarks primarily focus on static graphs, neglecting the temporal aspects that are crucial for generating questions that accurately reflect the timing and order of events. ChronoQG addresses this gap by integrating a comprehensive taxonomy of temporal constraints, sampling topological-temporal subgraphs, and employing trace-grounded question generation techniques. The framework generates four benchmark datasets from heterogeneous temporal knowledge graphs, resulting in a total of 16,011 verified questions. The authors evaluate various LLM-based KGQG methods and prompting baselines across different TKGQG settings, revealing that existing methods often fail to maintain temporal constraints, particularly in complex scenarios involving multiple constraints. This highlights the significant difference between static KGQG and TKGQG, establishing ChronoQG as a challenging testbed for future research in temporally aware question generation.
Methodology
The authors developed ChronoQG by creating a framework that includes a comprehensive taxonomy of temporal constraints, sampling techniques for topological-temporal subgraphs, and methods for generating questions that are grounded in both the graph structure and temporal constraints. They constructed four benchmark datasets from various temporal knowledge graphs and evaluated existing KGQG methods against these benchmarks.
Results
The evaluation demonstrated that existing KGQG methods, particularly those based on large language models, often failed to generate questions that accurately reflect the temporal constraints of the input data. The results highlighted difficulties in maintaining temporal validity and event ordering, especially under complex multi-constraint conditions.
Implications
ChronoQG serves as a critical resource for advancing research in TKGQG, providing a benchmark that can help improve the accuracy and reliability of question generation systems in applications that require temporal awareness, such as dialogue systems and automated question-answering platforms.
Relevance-Aware Rule: Structural Deletion of Irrelevant Conditions in Decision Trees
Interpretability
- Identifies the structural mechanics of irrelevant conditions in decision trees.
- Establishes a framework for relevance-aware rule deletion based on theoretical foundations.
- Introduces a multi-layered approach to diagnose the relevance of path conditions.
- Achieves substantial simplification of decision tree rules while maintaining reliability.
Read more
Relevance-Aware Rule: Structural Deletion of Irrelevant Conditions in Decision Trees
Summary
This paper addresses the issue of irrelevant conditions (IRCs) in decision trees, which arise from the structural mechanisms of tree splitting and persist even in optimal sparse tree induction algorithms. The authors propose a theoretical framework for the structural deletion of IRCs, establishing theorems and propositions that explain the underlying mechanics of IRCs. They introduce the concept of C1-links and C0-links, which represent class proportion shifts in binary splits. The proposed framework diagnoses the relevance of conditions by assessing their impact on prediction reliability, allowing for selective deletion of structurally and empirically irrelevant conditions while preserving those essential for maintaining rule reliability. The methodology is structured into three analytical layers: local annotation, leaf-relative diagnosis, and path-level effect evaluation. Experimental results demonstrate that the framework achieves significant rule simplification without compromising the reliability of the decision tree.
Methodology
The authors propose a structural IRC deletion framework that utilizes C1/C0 link annotations to identify structurally suspicious IRC candidates. The framework operates through two procedures: a mismatch-guided procedure for broader relevance-aware compression and a sibling-certified procedure for conservative simplification. Each procedure assesses the relevance of conditions based on their impact on prediction reliability across three analytical layers.
Results
The experimental results confirm that the proposed framework effectively simplifies decision tree rules by removing IRCs without sacrificing the reliability of the original tree. The framework's scalability is linear in relation to the number of leaf rules, indicating its efficiency.
Implications
The findings have significant implications for improving the interpretability of decision trees, making them more concise and easier to understand. This work can enhance the usability of decision trees in various applications, particularly in fields requiring clear decision-making processes.
TIDE: Trustworthy and Interpretable Battery Degradation Estimation with Contextual Learning and Symbolic Distillation
Interpretability
Time Series
Optimization
- TIDE integrates knowledge and operational data for improved battery SoH estimation.
- The model features three components: a knowledge-guided prior, a monotone residual, and a contextual learning component.
- TIDE achieves a 19.7% improvement in estimation accuracy over existing methods.
- The approach emphasizes the importance of trustworthiness and interpretability in battery health assessments.
Read more
TIDE: Trustworthy and Interpretable Battery Degradation Estimation with Contextual Learning and Symbolic Distillation
Summary
The paper presents TIDE, a novel battery degradation estimator designed to enhance the reliability of battery health estimation in battery-powered systems. TIDE addresses the critical need for accuracy, trustworthiness, and interpretability in battery state of health (SoH) assessments, particularly in intelligent connected systems where estimation errors can have cascading effects. The proposed method integrates battery-domain knowledge with operational measurements through a structured three-component backbone. This backbone includes a knowledge-guided degradation prior for trustworthy estimation, a monotone residual component for interpretable aging-consistent refinement, and a contextual learning component to capture battery-specific operational effects for improved accuracy. The model is distilled into a compact symbolic surrogate that simplifies the interpretation of its estimation logic. Experimental results demonstrate that TIDE achieves an average improvement of 19.7% in estimation fidelity compared to existing baselines, significantly reducing aging-consistency violations and supporting practical applications in battery health monitoring and decision-making.
Methodology
TIDE employs a three-component backbone that combines a knowledge-guided prior for foundational accuracy, a monotone residual for aging-consistent refinement, and a contextual learning component to account for operational variations. The model is then distilled into a symbolic form for easier interpretation.
Results
TIDE demonstrated a significant enhancement in estimation accuracy, achieving an average increase of 19.7% in fidelity over representative baseline models. The model also effectively reduced aging-consistency violations, supporting its trustworthiness.
Implications
The findings suggest that TIDE can be effectively utilized in battery health monitoring systems, enhancing decision-making processes in intelligent connected environments. Its emphasis on interpretability and trustworthiness could lead to safer and more reliable battery management practices.
RF Spectrogram Anomaly Detection with Quantum Kitchen Sinks: Architecture, Representation, and Hardware Validation
Theory
Efficient ML
Time Series
- Introduction of a novel QKS framework for RF spectrogram anomaly detection.
- Development of a validation-locked five-stage ablation protocol for systematic evaluation.
- Demonstration of QKS's superiority over classical anomaly detection methods.
- Real-world validation using actual measured RF signals and quantum hardware.
Read more
RF Spectrogram Anomaly Detection with Quantum Kitchen Sinks: Architecture, Representation, and Hardware Validation
Summary
This paper addresses the critical need for anomaly detection in wireless RF networks, which are vulnerable to malicious transmissions. The authors propose a novel approach using Quantum Kitchen Sinks (QKS), a hybrid quantum-classical model, to enhance the detection of anomalies in RF spectrograms. They extend the QKS framework by incorporating multi-depth data re-uploading and ring entanglement, and evaluate its performance through a rigorous five-stage ablation protocol. This protocol isolates various factors affecting detection performance, including architecture depth, input representation, and classical readout methods. The study utilizes a labeled dataset combining real sub-6 GHz cellular signals with synthetic anomalies to benchmark the QKS approach. Results indicate that QKS outperforms classical methods, achieving a test Area Under the Receiver Operating Characteristic curve (AUROC) of 0.8778 and a test F1 score of 0.7995. The findings demonstrate the effectiveness of QKS in real-world scenarios, validated on actual quantum hardware, and highlight its potential for practical deployment in secure spectrum management.
Methodology
The authors extended the standard QKS template by integrating multi-depth data re-uploading and ring entanglement. They employed a five-stage ablation protocol to systematically evaluate the impact of various architectural choices and input representations on anomaly detection performance. The methodology involved creating a labeled dataset of RF spectrograms and conducting experiments on both simulated and real quantum hardware.
Results
The QKS framework achieved a test AUROC of 0.8778 and a test F1 score of 0.7995, outperforming classical direct-readout baselines across all evaluated representation-readout pairs. The study found that Discrete Cosine Transform (DCT) representations consistently outperformed raw and PCA inputs, and moderate-depth entangled QKS configurations yielded the best results. The AUROC deviations between simulation and real hardware were minimal, indicating the robustness of the approach.
Implications
The findings suggest that QKS can significantly enhance anomaly detection in RF networks, providing a practical solution for secure spectrum management. This research opens avenues for further exploration of quantum machine learning applications in wireless communications and other domains requiring robust anomaly detection mechanisms.