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
Benchmarking Frontier Large Language Models Against Official Crash Database Coding Using Police Crash Narratives
NLP
Large Language Models
- The study benchmarks six LLMs against official crash coding using police narratives.
- GPT-5.5 High showed the highest agreement among evaluated models, but traditional baselines performed better in some metrics.
- Agreement varied significantly across different crash attributes, indicating the need for attribute-specific evaluations.
- The research highlights the potential of LLMs in enhancing crash data coding but underscores the importance of human review.
Read more
Benchmarking Frontier Large Language Models Against Official Crash Database Coding Using Police Crash Narratives
Summary
This study investigates the effectiveness of six frontier large language models (LLMs) in coding crash attributes from police narratives compared to official crash database entries. By linking 5,587 fatal-crash narratives with 5,889 structured crash records from Arkansas (2015-2025), the authors aimed to evaluate how well LLMs reproduce official crash coding. The models were assessed using a zero-shot prompt to extract attributes such as crash manner, non-motorist relation, and roadway conditions. Performance metrics included agreement rates, macro-averaged F1 scores, and Cohen's kappa, with comparisons to baseline methods. The results indicated that while GPT-5.5 High achieved the highest agreement among LLMs, traditional baselines outperformed LLMs in raw agreement and certain metrics. The study highlighted that agreement varied significantly across different crash attributes, suggesting that LLM performance should be evaluated on an attribute-specific basis. This research provides a benchmark for future evaluations of LLMs in crash coding and emphasizes the need for transparent baselines and human review in safety-critical applications.
Methodology
The study employed a zero-shot extraction task using six frontier LLMs to code attributes from police crash narratives. The outputs were compared with corresponding fields in the Arkansas fatal-crash database. Performance was evaluated using various metrics, including agreement rates, macro-averaged F1 scores, and Cohen's kappa, alongside comparisons with baseline methods.
Results
The analysis revealed that GPT-5.5 High achieved the highest agreement among the LLMs, but traditional baselines, such as always-majority and keyword-rule methods, produced higher raw agreement and comparable F1 scores. Agreement was highest for non-motorist relation and crash manner, while it was lowest for light condition and roadway surface condition. The study found that differences across crash attributes were more pronounced than differences across models.
Implications
The findings suggest that while LLMs can assist in coding crash attributes, their deployment should be carefully evaluated on an attribute-specific basis. This research underscores the importance of using transparent baselines and incorporating human review in safety-critical applications, potentially improving the reliability of crash data analysis and transportation safety measures.
Overcoming the Weakest-Link Effect in LLM-Driven Program Optimization via Heterogeneous Edit Recombination
Large Language Models
Optimization
Robotics
- Introduces HERO, a program optimizer that separates edit generation from selection.
- Demonstrates that explicit performance feedback is not essential for effective program optimization.
- Addresses the weakest-link effect by allowing for heterogeneous edit recombination.
- Achieves superior performance in various domains compared to existing LLM-based optimization methods.
Read more
Overcoming the Weakest-Link Effect in LLM-Driven Program Optimization via Heterogeneous Edit Recombination
Summary
This paper addresses the challenges of program optimization using large language models (LLMs), particularly focusing on the limitations of existing methods that rely on first-order textual gradients derived from performance evaluations. The authors introduce HERO (Heterogeneous Edit Recombination Optimizer), a novel approach that allows LLMs to generate diverse atomic edits without needing explicit performance feedback. By separating the generation of edits from their selection, HERO mitigates the 'weakest-link effect,' where a single detrimental edit can negate the benefits of beneficial ones when applied together. The methodology involves prompting the LLM to propose non-overlapping edits, which are then evaluated and recombined to form improved programs. The evaluation of HERO across various domains, including algorithmic problems and robotic path planning, demonstrates its effectiveness in discovering higher-scoring programs and achieving faster convergence compared to prior LLM-based optimizers, all while consuming fewer tokens.
Methodology
HERO prompts an LLM to generate a set of diverse, non-overlapping atomic edits based on the current program and task description, without providing performance feedback. It then evaluates and selects the best combinations of these edits to improve the program, effectively neutralizing the impact of harmful edits.
Results
HERO consistently outperformed existing LLM-based optimizers across multiple domains, discovering higher-scoring programs and converging faster while using fewer tokens.
Implications
The findings suggest that LLMs can be effectively utilized for program optimization without relying on explicit performance feedback, opening new avenues for solving complex problems in various fields, including algorithm design and robotics.
Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models in Speculative Decoding
Large Language Models
Efficient ML
Optimization
- Conversion from MHA/GQA to MLA can degrade proposal agreement in speculative decoding.
- Functional reconstruction optimizes draft quality during training without requiring verifier supervision.
- The proposed method is converter-agnostic and applicable across different model configurations.
- Functional reconstruction improves acceptance rates in 37 out of 64 task configurations evaluated.
Read more
Beyond KV Reconstruction: Functional Reconstruction for MLA Draft Models in Speculative Decoding
Summary
This paper addresses the challenges associated with converting multi-head attention (MHA) and grouped-query attention (GQA) models to multi-head latent attention (MLA) for efficient long-context language model inference. The authors highlight that while MLA reduces memory traffic by using compact latent states, direct conversion from MHA/GQA can degrade the performance of speculative decoding, a method that accelerates autoregressive generation by proposing multiple tokens for verification. The study identifies that the loss of agreement between draft proposals and target verification during conversion leads to lower acceptance rates of proposed tokens. To tackle this issue, the authors propose a novel functional reconstruction method that optimizes the MLA attention modules to match the output responses of the original MHA/GQA models on calibration hidden states. This method is converter-agnostic and does not require any changes to the inference graph or additional supervision. The evaluation of various configurations demonstrates that the functional reconstruction significantly improves draft acceptance rates and maintains high throughput across multiple tasks and model families.
Methodology
The authors introduce a functional reconstruction approach that optimizes the MLA attention modules to reproduce the output responses of original MHA/GQA blocks on calibration hidden states. This end-to-end reconstruction is performed during training and does not alter the inference structure or require additional supervision.
Results
The functional reconstruction method led to material improvements in acceptance rates for 37 out of 64 matched task configurations, with 26 remaining unchanged and only one showing a decrease. Additionally, 12 of the improved configurations also exhibited increased throughput.
Implications
The findings suggest that optimizing the conversion process for MLA can enhance the performance of speculative decoding in language models, potentially leading to more efficient and faster autoregressive generation in practical applications.
Compression-Based Behavioral Similarity for Open-World Sybil Discovery on Ethereum
Graph Learning
- Introduces a compression-based approach for Sybil detection that does not rely on financial links.
- Utilizes symbolic Transaction Grammar to capture wallet behavior and transaction patterns.
- Employs Normalized Compression Distance (NCD) for measuring behavioral similarity.
- Demonstrates robustness against adversarial tactics and temporal changes in behavior.
Read more
Compression-Based Behavioral Similarity for Open-World Sybil Discovery on Ethereum
Summary
The paper addresses the challenge of detecting Sybil attackers in the Ethereum blockchain, who manipulate governance and exploit airdrops by masquerading as legitimate users. Traditional detection methods rely on explicit financial interactions or supervised machine learning, which are limited by their reliance on labeled data and the evolving nature of attack strategies. The authors propose a novel approach that utilizes compression-based behavioral similarity to differentiate between Sybil bots, organic users, and arbitrage bot wallets without requiring direct financial links. They introduce a symbolic Transaction Grammar to model wallet behavior, capturing transaction rhythm, execution structure, and functional intent. This framework employs a Gzip-based Normalized Compression Distance (NCD) to construct a behavioral graph for Sybil discovery. The method is validated against supervised machine learning baselines and demonstrates robustness against temporal drift and adversarial perturbations. The authors contribute a leakage-aware behavioral framework that facilitates the discovery of Sybil candidates without the need for supervised training, positioning it as a local discovery primitive for open-world blockchain audits.
Methodology
The authors synthesize a symbolic Transaction Grammar from Ethereum Virtual Machine (EVM) traces to model wallet behavior. They apply Gzip-based Normalized Compression Distance (NCD) to construct a behavioral graph, allowing for the analysis of pairwise behavioral similarities without the need for supervised training. The method is validated through comparisons with supervised learning baselines and stress tests involving synthetic camouflage.
Results
The proposed framework effectively identifies Sybil candidates and differentiates them from legitimate users and arbitrage bots. It shows resilience to adversarial perturbations and temporal drift, outperforming traditional supervised methods in scenarios where labeled data is scarce or unavailable.
Implications
This research has significant implications for blockchain security, particularly in enhancing the robustness of Sybil detection mechanisms in decentralized finance (DeFi) and governance contexts. The methodology can be applied to improve the integrity of blockchain audits and governance processes by enabling the identification of malicious actors without relying on explicit financial interactions.
Expanding Data-Agnostic Pivotal Instances Selection Models with Proximity Trees and Ensemble Learning
Interpretability
- Introduction of PivotTree, a hierarchical model for pivotal instance selection and prediction.
- Incorporation of pairs of pivots and ensemble methods to enhance model effectiveness.
- Data-agnostic approach allowing application across various data modalities.
- Demonstrated superior performance over existing instance selection strategies.
Read more
Expanding Data-Agnostic Pivotal Instances Selection Models with Proximity Trees and Ensemble Learning
Summary
This paper presents a novel approach to enhance the interpretability of machine learning models through a hierarchical, data-agnostic pivotal instance selection model called PivotTree. The authors argue that traditional decision-making processes in machine learning often lack interpretability, which is crucial for user trust and understanding. Inspired by human cognitive processes, the PivotTree model selects representative instances (pivots) based on their similarity to new input cases, allowing for intuitive decision-making. The model not only serves as a pivot selection technique but also functions as a predictive model, classifying new instances by navigating through a similarity-based decision tree structure. The authors extend the model by incorporating pairs of pivots and utilizing proximity and oblique trees, as well as ensemble methods, to improve performance and versatility. The experiments conducted across various datasets, including tabular data, text, images, and time series, demonstrate that PivotTree outperforms existing instance selection strategies and achieves competitive results compared to state-of-the-art interpretable models while maintaining a minimal number of pivots. This work highlights the importance of interpretability in machine learning and provides a framework that aligns more closely with human reasoning.
Methodology
The authors developed the PivotTree model, which utilizes a similarity-based decision tree structure for selecting pivotal instances and making predictions. The model incorporates oblique and proximity splits to refine decision-making and is integrated into ensemble methods to enhance robustness. The approach is data-agnostic, leveraging pre-trained networks for data transformation across different modalities.
Results
The experiments showed that PivotTree outperformed alternative instance selection strategies and achieved competitive results against state-of-the-art interpretable models. The model maintained a minimal number of pivots while demonstrating effectiveness across diverse datasets, including tabular data, text, images, and time series.
Implications
The findings suggest that PivotTree can be a valuable tool in various domains requiring interpretable machine learning models, such as healthcare, finance, and consumer behavior analysis. Its ability to provide intuitive explanations based on similarity to representative instances can enhance user trust and facilitate decision-making processes.
Parameter-Free Heavy-Tailed Bandits
Theory
- Introduces a parameter-free approach to heavy-tailed bandits, resolving an open problem in the field.
- Characterizes the trade-off between distribution-dependent and distribution-free regret guarantees when parameters are unknown.
- Presents the AdaR-ETC algorithm, which adapts to unknown moment bounds and achieves optimal regret bounds.
- Demonstrates that no single policy can guarantee sublinear regret uniformly across all tail exponents.
Read more
Parameter-Free Heavy-Tailed Bandits
Summary
This paper addresses the challenge of adapting algorithms for heavy-tailed bandits without prior knowledge of the tail parameters, specifically the tail exponent (ϵ) and moment bound (u). Heavy-tailed distributions are common in various sequential decision-making scenarios where extreme outcomes can significantly impact performance. The authors resolve an open problem from COLT 2025 by characterizing the trade-offs in regret when these parameters are unknown. They introduce an adaptive algorithm, Adaptive Robust ETC (AdaR-ETC), which does not require knowledge of u and achieves a regret bound that balances distribution-dependent and distribution-free guarantees. Furthermore, the paper demonstrates that the same algorithm can be adapted without knowing ϵ, achieving sublinear regret for any fixed ϵ > 0. The results highlight the limitations of simultaneous adaptation to both parameters and establish a frontier for the statistical cost of adapting to unknown heavy tails.
Methodology
The authors analyze the adaptation of algorithms to unknown moment bounds while fixing the tail exponent. They derive lower bounds on regret rates and propose the AdaR-ETC algorithm, which employs a Median-of-Means strategy to achieve robust performance without prior knowledge of the moment bound. The exploration schedule is calibrated to adapt to the tail exponent, allowing the algorithm to maintain sublinear regret for fixed values of ϵ.
Results
The paper proves that any algorithm unaware of the moment bound must face a trade-off between distribution-dependent and distribution-free regret guarantees. The AdaR-ETC algorithm achieves a distribution-free regret bound that is tight on the established lower bound frontier. Additionally, the algorithm can be calibrated to operate without knowledge of the tail exponent, achieving sublinear regret for all fixed ϵ > 0, although it cannot guarantee uniform sublinear regret across all values of ϵ.
Implications
The findings have significant implications for real-world applications in fields like finance, online advertising, and network management, where heavy-tailed distributions are prevalent. The ability to adapt to unknown parameters can enhance decision-making processes in these domains, leading to improved performance and reduced regret in sequential decision-making tasks.
Enhancing Irregular Time Series Forecasting with Continuous-Time Modeling Framework
Time Series
- WrapFlow is a unified continuous-time modeling framework for irregular multivariate time series forecasting.
- Introduces Continuous-Time Tokenization to encode observations without discretization.
- Utilizes Residual Flow Matching to learn continuous residual dynamics without numerical-solver simulation.
- Achieves state-of-the-art performance on multiple datasets, demonstrating its practical applicability.
Read more
Enhancing Irregular Time Series Forecasting with Continuous-Time Modeling Framework
Summary
This paper addresses the challenges of forecasting irregular multivariate time series (IMTS) which are prevalent in various applications such as healthcare and environmental sensing. Traditional methods either rely on discretization techniques that distort the continuous-time nature of the data or use ODE-based frameworks that require complex architectures and significant computational resources. The authors propose a novel framework called WrapFlow, which incorporates Continuous-Time Tokenization to encode raw observation events and models long unobserved intervals with gap-aware tokens. This allows for the effective capture of temporal dependencies using a standard Transformer architecture. On the output side, WrapFlow employs a simulation-free training paradigm known as Residual Flow Matching, which learns to predict residual dynamics without the need for numerical solvers during training. The framework demonstrates state-of-the-art performance across multiple real-world datasets, showcasing its efficiency and effectiveness in continuous-time forecasting.
Methodology
WrapFlow employs Continuous-Time Tokenization to create event-level tokens that incorporate observed values, timestamps, and inter-event intervals. It uses a Transformer backbone to model long-range dependencies. For output generation, it implements Residual Flow Matching, which learns conditional residual vector fields based on base predictions, avoiding the computational overhead of numerical solvers during training.
Results
Extensive experiments reveal that WrapFlow outperforms existing state-of-the-art methods in irregular time series forecasting across various datasets, confirming its effectiveness and efficiency.
Implications
The proposed framework can significantly enhance forecasting accuracy in applications involving irregular time series data, such as healthcare monitoring and environmental sensing, by providing a more principled approach to continuous-time modeling.
TerraNova: A Foundation Model for the Anthropocene
Multimodal
- TerraNova is the first model to learn a unified representation across the observational breadth of the Anthropocene.
- The model employs inter-geometry coupling to connect administrative units with their corresponding physical territories.
- It outperforms purpose-built geospatial embeddings on static targets and extends to dimensions like uncertainty and time.
- The model provides predictive distributions with aleatoric and epistemic uncertainty proxies for each query.
Read more
TerraNova: A Foundation Model for the Anthropocene
Summary
The paper introduces TerraNova, a foundation model designed to address the complex interconnections between the physical Earth and human societies in the Anthropocene. The authors argue that existing models fail to represent these systems as a unified whole due to geometric discrepancies: Earth-system data is continuous while societal data is discrete. TerraNova is trained on 1,024 variables, including 512 gridded Earth-system fields and 512 national indicators, utilizing dedicated encoders for location, country, time, and task. Cross-modal transformers fuse these inputs into a shared spatiotemporal state, while a hypernetwork generates a per-query decoder that provides predictive distributions with uncertainty estimates. The model employs two contrastive objectives to align the different geometries, enabling it to reconstruct dense fields from sparse observations and adapt to new variables quickly. The results demonstrate that TerraNova outperforms existing geospatial encoders and supports capabilities such as country-level retrieval and downscaling, while also addressing axes like time and uncertainty that are typically neglected in other models.
Methodology
TerraNova utilizes dedicated encoders for different data types (geospatial and societal), cross-modal transformers to integrate these inputs into a shared representation, and a hypernetwork for generating task-specific decoders. The model is trained using two contrastive objectives to align the different geometries of the data.
Results
The model demonstrates competitive performance against existing geospatial encoders, effectively reconstructing dense fields from sparse data and adapting to new variables in a matter of minutes. It supports country-level capabilities and provides predictive distributions that include uncertainty estimates.
Implications
TerraNova has the potential to enhance climate change research, economic modeling, and policy-making by providing a more integrated understanding of the interactions between human societies and the physical environment. Its ability to handle uncertainty and adapt to new data makes it a valuable tool for real-time decision-making in environmental and socio-economic contexts.
DFSC: Error-Controlled Differentiable Mittag-Leffler Propagation for Fractional Scientific Machine Learning
Theory
Efficient ML
Optimization
- DFSC separates known fractional propagation from unknown corrections, optimizing them jointly.
- The adaptive algorithm increases truncation depth until error tolerances are met.
- DFSC provides certified error bounds and diagnostics for its estimates.
- The framework significantly reduces computation time for repeated queries on CPU and GPU.
Read more
DFSC: Error-Controlled Differentiable Mittag-Leffler Propagation for Fractional Scientific Machine Learning
Summary
The paper introduces DFSC, a novel framework for fractional scientific machine learning that leverages differentiable Mittag-Leffler propagation. DFSC is designed to efficiently handle fractional differential equations by separating known fractional propagation from data-driven corrections. This is achieved through the Mittag-Leffler Spectral Layer (MLSL), which allows for the joint optimization of fractional orders and neural network parameters. The framework includes an adaptive algorithm that adjusts the truncation depth or Lanczos dimension to meet specified error tolerances. DFSC supports various operator paths and is capable of handling both dense and sparse data. The authors demonstrate that DFSC can significantly reduce computation time for repeated queries and provides certified error bounds for its estimates. The framework is validated across multiple experimental conditions, showcasing its effectiveness in solving fractional differential equations while maintaining a structured learning approach.
Methodology
The authors developed DFSC as a PyTorch-based framework that includes the Mittag-Leffler Spectral Layer (MLSL) for differentiable propagation of fractional differential equations. The framework incorporates an adaptive algorithm for error control and supports various operator paths, including dense, sparse, and matrix-free operations. The methodology emphasizes joint optimization of fractional orders and neural network parameters, along with a structured approach to problem-solving.
Results
DFSC demonstrated a median bound/error effectivity of 1.246 across 59 reference cases. The framework achieved significant reductions in repeated-query computation time, with improvements ranging from 4.61 to 16.22 times, depending on the hardware used. Additionally, the framework was validated through a series of experiments across different physical domains, confirming its reliability and effectiveness in handling fractional differential equations.
Implications
DFSC provides a robust tool for researchers and practitioners working with fractional differential equations in scientific machine learning. Its ability to efficiently handle complex dynamics while providing error control makes it suitable for applications in various fields, including physics, engineering, and data-driven modeling. The framework's structured approach may also inspire future developments in differentiable programming and hybrid modeling techniques.
MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models
Large Language Models
Interpretability
Optimization
- MOT-SR integrates external analytical tools for improved variable dependency analysis.
- The framework employs a multi-objective evaluation to optimize for accuracy, complexity, and generalization.
- MOT-SR outperforms existing symbolic regression methods across multiple tasks.
- The system operates in a closed-loop manner, continuously refining strategies and equations.
Read more
MOT-SR: Multi-Objective Tool-Augmented Scientific Equation Discovery with Large Language Models
Summary
The paper introduces MOT-SR, a novel framework for symbolic regression (SR) that enhances the discovery of scientific equations from observational data using large language models (LLMs). Traditional SR methods face limitations such as inadequate data analysis for variable dependencies and a focus on single-objective evaluation, which can lead to premature convergence and suboptimal solutions. MOT-SR addresses these issues by integrating external analytical tools to extract structural priors and employing a multi-objective evaluation approach that optimizes for accuracy, complexity, and generalization. The framework consists of two collaborative LLM modules: a Meta Strategy Generator that selects tools and synthesizes optimization strategies, and an Equation Generator that produces candidate equations. This closed-loop system continuously refines both strategies and equation structures. The efficacy of MOT-SR is demonstrated across 40 standard tasks, outperforming existing SR methods in terms of accuracy, generalization, and efficiency. Additionally, it is validated on extreme mass-ratio inspiral (EMRI) orbital modeling, achieving the lowest trajectory-level integration error on held-out configurations, showcasing its potential for reliable modeling of long-horizon scientific dynamics.
Methodology
MOT-SR utilizes a two-module system comprising a Meta Strategy Generator and an Equation Generator. The Meta Strategy Generator analyzes residuals and applies scientific tools to guide the search for equations, while the Equation Generator produces new candidate equations based on the generated strategies. The framework maintains a dynamic Pareto front to evaluate and refine candidate equations across multiple objectives.
Results
MOT-SR demonstrated superior performance across 40 standard tasks in terms of accuracy, generalization, and efficiency compared to existing symbolic regression methods. In the specific application of EMRI orbital modeling, it achieved the lowest trajectory-level integration error on held-out configurations.
Implications
The findings suggest that MOT-SR can significantly enhance the process of scientific equation discovery, leading to more reliable models in various scientific fields, particularly in complex systems where long-term predictions are critical.
FADEx: Feature Attribution and Distortion-based Explanation of Dimensionality Reduction
Interpretability
- FADEx provides a single attribution per feature, enhancing interpretability.
- The method is agnostic to the dimensionality reduction technique used.
- FADEx incorporates distortion analysis and directional feature influence.
- Qualitative and quantitative evaluations show FADEx's superiority over existing methods.
Read more
FADEx: Feature Attribution and Distortion-based Explanation of Dimensionality Reduction
Summary
The paper presents FADEx, a novel method for local feature attribution in dimensionality reduction (DR) that addresses the challenges of interpretability in non-linear DR techniques. Traditional DR methods often obscure how individual features influence the positioning of instances in reduced spaces, complicating the analysis of high-dimensional data. Existing explanation methods face limitations such as multiple attributions per feature and restricted applicability to specific DR techniques. FADEx overcomes these issues by utilizing local linear approximation through first-order Taylor expansion and Singular Value Decomposition (SVD), allowing for a single attribution per feature and eliminating the need for out-of-sample data mapping. Additionally, FADEx integrates distortion analysis and directional influence of features, providing a comprehensive understanding of the DR process. The method's effectiveness is demonstrated through qualitative and quantitative evaluations, showing that FADEx outperforms existing approaches in providing robust and reliable explanations for DR methods.
Methodology
FADEx employs a local linear approximation via first-order Taylor expansion and Singular Value Decomposition (SVD) to compute local models through weighted least squares. This methodology allows for the assessment of feature attributions and distortions without requiring out-of-sample data mapping, making it applicable across various DR techniques.
Results
The evaluations indicate that FADEx provides robust and reliable explanations, outperforming existing dimensionality reduction explanation methods in terms of interpretability and effectiveness. The method successfully elucidates the influence of individual features on instance positioning and the geometric nature of the transformations involved.
Implications
FADEx has significant implications for the fields of data visualization and machine learning, particularly in enhancing the interpretability of complex models and facilitating better understanding of high-dimensional data structures. It can be applied in various domains where dimensionality reduction is crucial for data exploration and analysis.
DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search
Reinforcement Learning
Optimization
Efficient ML
- DreamQAS introduces a feedback-only world model that does not require exact ground-state energy during training.
- The framework utilizes a recurrent ensemble to predict oracle-free feedback scores, enhancing decision-making in QAS.
- Significant reductions in real VQE calls (up to 10.6x fewer) are achieved while maintaining high accuracy across molecular tasks.
- The model demonstrates improved action-ranking utility and risk coverage through ensemble disagreement.
Read more
DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search
Summary
The paper introduces DreamQAS, a novel model-based reinforcement learning framework designed to enhance the efficiency of quantum architecture search (QAS) for variational quantum eigensolvers (VQE). Traditional RL-QAS methods face challenges due to the high cost of evaluating circuit performance after each modification. DreamQAS addresses this by preserving known circuit dynamics while focusing on learning the expensive post-VQE feedback. The framework employs a recurrent randomized-prior ensemble to predict scores relative to an empirical energy frontier, facilitating multi-step imagined policy learning over legal circuits. Key innovations include a ranking-based activation mechanism, uncertainty-aware pessimism, and selective real-VQE verification, which together form a reliability-controlled learning loop. The results demonstrate that DreamQAS significantly reduces the number of real VQE calls required to achieve target accuracy across multiple molecular tasks, while also improving the quality of the final architecture. This approach emphasizes the importance of decision-useful feedback rather than merely accurate energy predictions, marking a significant advancement in the field of quantum architecture search.
Methodology
The methodology involves a model-based reinforcement learning approach where a recurrent ensemble predicts feedback scores for circuit prefixes. The framework incorporates multi-step imagined policy learning, ranking-based activation for decision readiness, and selective verification of promising circuits using real VQE evaluations. This allows for efficient exploration of the circuit space while controlling for reliability and uncertainty.
Results
Under a common budget of 15,000 episodes, DreamQAS achieved the lowest mean frozen-policy energy error on four out of five molecular tasks, with the second-lowest error on the remaining task. It required 1.6x to 2.0x fewer real VQE calls on four tasks and 10.6x fewer on the BeH2-8q task to reach fine-error targets. The counterfactual action-ranking utility improved significantly across all tasks, indicating enhanced decision-making capabilities.
Implications
The findings suggest that DreamQAS can lead to more efficient quantum architecture searches, reducing computational costs associated with VQE evaluations. This has potential applications in quantum computing, particularly in optimizing circuit designs for various molecular simulations and other quantum algorithms.
Explore Beyond the Boundary Using Entropic Information
Reinforcement Learning
- Introduction of ENTINEX, a novel exploration method for RL.
- Utilizes entropic information to identify state-novelty distribution boundaries.
- Assigns intrinsic rewards to encourage exploration beyond these boundaries.
- Demonstrates superior performance compared to existing exploration methods.
Read more
Explore Beyond the Boundary Using Entropic Information
Summary
In reinforcement learning (RL), effectively exploring environments with sparse and delayed rewards is a significant challenge due to limited feedback. This paper introduces a novel exploration method called Entropic Information for Exploration (ENTINEX), which incentivizes agents to explore beyond the boundaries of the state distribution. ENTINEX assigns intrinsic rewards to states near the state-novelty distribution (SND) boundary, utilizing entropic information from the action probability distribution to identify these boundaries. The authors argue that existing methods often rely on novelty differences between consecutive states, which can be less effective. Through extensive experiments, ENTINEX demonstrates consistent improvements in exploration performance across various environments characterized by sparse and delayed rewards, outperforming traditional exploration strategies. The findings suggest that leveraging entropic information can significantly enhance the exploration capabilities of RL agents, leading to better policy learning in complex environments.
Methodology
The methodology involves defining the state-novelty distribution (SND) and using entropic information from the action probability distribution to identify boundaries between explored and unexplored states. ENTINEX assigns intrinsic rewards to states near these boundaries, encouraging agents to explore less-visited areas effectively.
Results
The experimental results indicate that ENTINEX consistently outperforms existing exploration methods in environments with sparse and delayed rewards. The method shows significant improvements in exploration performance, leading to better policy learning outcomes.
Implications
The implications of this research suggest that ENTINEX can be applied in various real-world scenarios where exploration is critical, such as autonomous driving, healthcare, and robotics. By enhancing exploration strategies, it can lead to more efficient learning and improved decision-making in complex environments.
ZUNA1.1: A more flexible EEG foundation model for Denoising and Super-resolution
Time Series
Generative Models
- ZUNA1.1 allows for reconstruction of variable-length EEG sequences up to 30 seconds.
- The model significantly outperforms traditional EEG denoising methods.
- Improvements include implicit data augmentation and quality-aware preprocessing.
- Training corpus expanded to approximately 3.5M channel-hours for better performance.
Read more
ZUNA1.1: A more flexible EEG foundation model for Denoising and Super-resolution
Summary
The paper introduces ZUNA1.1, a 380M-parameter diffusion autoencoder designed for flexible EEG signal reconstruction. This model addresses limitations of its predecessor, ZUNA1, by allowing for variable-length sequence reconstruction, accommodating up to 30 seconds of EEG data with an arbitrary number of channels. ZUNA1.1 outperforms traditional EEG denoising methods, such as spherical spline interpolation, and is capable of handling diverse reconstruction tasks. Key improvements include variable-length training, implicit data augmentation, quality-aware preprocessing, and a diverse dropout mixture to simulate realistic missing data scenarios. The training corpus has been expanded significantly, allowing for better generalization and performance. ZUNA1.1 is released as open source under the Apache 2.0 license, promoting accessibility in EEG research and applications.
Methodology
ZUNA1.1 employs a diffusion autoencoder architecture, utilizing variable-length training on randomly cropped EEG segments. It incorporates implicit data augmentation techniques and quality-aware preprocessing to enhance the robustness of the model. A diverse dropout mixture simulates various real-world scenarios of missing data, improving the model's ability to infill corrupted signals.
Results
ZUNA1.1 demonstrates superior performance compared to its predecessor, ZUNA1, and traditional methods like spherical spline interpolation. The model's ability to handle variable-length sequences and diverse dropout patterns leads to improved reconstruction accuracy across multiple datasets.
Implications
The advancements in ZUNA1.1 could lead to more effective EEG signal processing in clinical, research, and consumer applications, enabling better interpretation of brain activity and enhancing the usability of consumer-grade EEG devices.
When Does On-Policy Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning
Reinforcement Learning
Robotics
Theory
- OVI allows learners to focus on realizing the expert's value function, easing representational demands.
- On-policy interaction is essential for efficient value-based imitation learning.
- The paper presents a negative result showing offline IL methods require more complexity without interaction.
- Empirical results indicate OVI outperforms traditional methods, especially with less expressive learner networks.
Read more
When Does On-Policy Interaction Help? Representational Tradeoffs in Value-Based Imitation Learning
Summary
This paper investigates the role of on-policy interaction and value function estimation in improving imitation learning (IL) performance. Traditional methods like Behavior Cloning (BC) often face challenges such as compounding errors and performance plateaus due to the learner's inability to accurately represent the expert's policy. The authors propose OVI, an interactive on-policy IL algorithm that allows the learner to focus on realizing the expert's value function rather than the full action distribution of the expert's policy. This approach reduces the representational burden on the learner, making it statistically and computationally efficient when the learner can query the expert interactively. The study also presents a negative result indicating that without interaction, offline IL methods must scale with the complexity of the expert policy class. Empirical results demonstrate that OVI outperforms existing offline and interactive IL methods, particularly when the learner's network is less expressive than the expert's, highlighting the advantages of combining expert interaction with value-based learning.
Methodology
The authors introduce OVI, an interactive on-policy imitation learning algorithm that queries the expert along the learner's trajectories. The algorithm leverages a saddle-point formulation of IL to identify a useful value function, focusing on QπE-realizability, which requires only the representation of the expert's value function. The study also includes empirical validation using Gymnasium environments to compare OVI with existing methods.
Results
OVI demonstrated superior performance compared to offline policy-based methods like Behavior Cloning and interactive methods like DAgger. The empirical results showed that OVI achieved higher returns with smaller learner networks, supporting the theoretical claims about the benefits of expert interaction and value-based design.
Implications
The findings suggest that incorporating on-policy interaction in imitation learning can significantly enhance performance, particularly in scenarios where the learner's capacity is limited. This has implications for various applications, including robotics and language modeling, where efficient learning from expert demonstrations is crucial.
Mirror Learning
Robotics
Generative Models
Reinforcement Learning
- Introduces 'Mirror Learning' as a framework for imitation learning from third-person observations.
- Combines perspective transformation and inverse dynamics modeling to create 'mirror data'.
- Demonstrates that mirror data can effectively train policies and improve performance when combined with first-person data.
- Highlights the potential of generative models in enabling safe and scalable learning without direct interaction.
Read more
Mirror Learning
Summary
The paper introduces a novel framework called 'Mirror Learning' for imitation learning, emphasizing the potential of learning from third-person observations rather than relying solely on first-person data. The authors argue that traditional behavior cloning (BC) methods are limited when it comes to leveraging rich observational signals from third-person demonstrations, which are often utilized by humans and animals. The proposed method combines a learned perspective transformation, achieved through a fine-tuned video diffusion model, with an inverse dynamics model to infer action trajectories in the learner's control space. This combination allows for the generation of 'mirror data,' which consists of pseudo first-person expert data derived from third-person observations. The empirical results demonstrate that training policies solely on mirror data is effective, and augmenting first-person BC training with mirror data significantly enhances policy performance. The findings suggest that modern generative world models can provide a scalable and safe alternative to traditional data collection methods that rely heavily on teleoperation.
Methodology
The methodology involves two main components: (1) a learned perspective transformation using a fine-tuned video diffusion model to convert third-person observations into a first-person perspective, and (2) an inverse dynamics model that infers the corresponding action trajectories. This allows for the synthesis of mirror data, which is used to train policies in a behavior cloning framework.
Results
The empirical evaluation shows that policies trained solely on mirror data are effective. Furthermore, augmenting first-person behavior cloning with mirror data leads to improved performance in downstream tasks, indicating the utility of the proposed approach in enhancing imitation learning.
Implications
The findings suggest that mirror learning can significantly reduce the need for extensive first-person data collection, making it a promising approach for applications in robotics and autonomous systems where safety and efficiency are critical. It opens avenues for further research into observational learning and the use of generative models in training intelligent agents.
Understanding Submodular Information Measure Based Objectives for Representation Learning: A Variance and Separation Perspective
Theory
Optimization
Multimodal
- Establishes a theoretical framework connecting SIMs to classical representation learning concepts.
- Demonstrates that different SIM formulations optimize distinct geometric properties of representation spaces.
- Validates theoretical insights through controlled synthetic experiments.
- Highlights the importance of understanding geometric biases induced by different SIMs.
Read more
Understanding Submodular Information Measure Based Objectives for Representation Learning: A Variance and Separation Perspective
Summary
This paper investigates the theoretical foundations of Submodular Information Measures (SIMs) as objectives for representation learning, particularly in the context of supervised contrastive learning. The authors establish a unified framework that connects various SIM formulations to classical concepts in representation learning, such as compactness, separation, diversity, and coverage. They demonstrate that different SIMs, including Total Information (TI) and Mutual Information (MI) formulations, optimize distinct geometric properties of the representation space. For instance, Graph Cut based TI objectives correspond to within-class variance, while LogDet based TI objectives relate to generalized variance. The paper also highlights how MI formulations capture inter-class structures, with specific objectives quantifying class separation and coverage. To validate their theoretical insights, the authors conduct controlled synthetic experiments that manipulate variance, covariance structure, class imbalance, and multimodal overlap, confirming that different SIMs induce unique geometric biases. This work provides a comprehensive understanding of SIMs, guiding future research in selecting and designing submodular objectives for representation learning.
Methodology
The authors develop a theoretical framework linking SIMs to classical statistical measures and conduct synthetic experiments to analyze the behavior of various SIMs under controlled conditions. They manipulate factors such as variance, covariance structure, and class separation to observe the effects of different SIM formulations on representation learning.
Results
The empirical results align closely with the theoretical interpretations, showing that different SIMs induce fundamentally different geometric biases. Graph Cut based objectives favor global class separation, while Facility Location based objectives emphasize coverage and multimodal representation. LogDet based objectives capture covariance structures effectively.
Implications
This work offers a foundational understanding of SIMs for representation learning, which can inform the design of more effective loss functions and objectives in machine learning. It emphasizes the need for careful selection of SIMs based on their geometric properties, particularly in complex scenarios involving multimodal data and class imbalance.
End-to-End Fairness Optimization with Fair Decision-Focused Learning
Optimization
Theory
- E2EFO provides a unified framework for integrating fairness in predictive modeling and decision-making.
- FDFL optimizes for prediction accuracy, prediction fairness, and decision regret simultaneously.
- The paper derives exact closed-form decision Jacobians for fair allocation problems.
- Numerical experiments show improved fairness in resource allocation when using E2EFO and FDFL.
Read more
End-to-End Fairness Optimization with Fair Decision-Focused Learning
Summary
This paper introduces End-to-End Fairness Optimization (E2EFO), a framework that integrates fairness considerations across the prediction-to-decision pipeline in resource allocation scenarios. The authors emphasize the importance of addressing both prediction fairness and decision fairness to mitigate biases that can arise when predictions inform decisions. They propose Fair Decision-Focused Learning (FDFL), a training paradigm that jointly optimizes for prediction accuracy, prediction fairness, and decision regret. The methodology involves using multi-task learning techniques to combine gradients from these objectives, with a focus on deriving a decision Jacobian for fair allocation scenarios. The paper presents numerical experiments in healthcare and synthetic resource allocation contexts, demonstrating the effectiveness of jointly considering prediction and decision fairness. The results indicate that E2EFO and FDFL can significantly improve fairness in decision-making processes compared to traditional methods that treat prediction and decision tasks separately.
Methodology
The authors propose a multi-task learning approach to Fair Decision-Focused Learning (FDFL), which combines gradients from three objectives: prediction accuracy, prediction fairness, and decision regret. They derive a closed-form decision Jacobian for specific fair allocation scenarios and apply differentiable optimization techniques to handle more general cases.
Results
The experiments conducted in healthcare and synthetic resource allocation settings demonstrate that the E2EFO framework and FDFL training paradigm lead to better fairness outcomes compared to traditional prediction-focused or decision-focused methods. The results highlight the importance of jointly optimizing for prediction and decision fairness.
Implications
The findings suggest that integrating fairness considerations throughout the prediction-to-decision pipeline can lead to more equitable outcomes in critical applications such as healthcare, finance, and social services. This approach can help mitigate biases and improve the overall fairness of resource allocation decisions.
Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation
Generative Models
- Introduction of Coarse-Fine Transport Distance (CFTD) for evaluating generative models.
- Demonstration of the effectiveness of MACE features in capturing structural and chemical information.
- Benchmarking of 10,000 generated crystal structures reveals trade-offs between quality and novelty.
- MACE feature conditioning improves the stability-novelty profile of generative models.
Read more
Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation
Summary
This paper explores the use of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), specifically MACE, as a means to enhance the generation and evaluation of inorganic crystal structures. The authors introduce a novel distance metric, the Coarse-Fine Transport Distance (CFTD), which integrates quality and novelty assessments into a unified framework. This metric employs two distinct featurizers: a coarse MACE featurizer for quality evaluation and a contrastive GNN featurizer for novelty. The paper demonstrates that CFTD effectively captures the quality of generated crystal structures while also identifying instances of memorization. The authors benchmark various generative models, revealing that while relaxation of generated structures generally improves quality, it can also lead to increased memorization. The study further shows that incorporating MACE features into a material generative model enhances its performance, leading to a better stability-novelty profile. Overall, the findings suggest that pretrained MLIPs can significantly inform and improve the material generation process.
Methodology
The authors developed the Coarse-Fine Transport Distance (CFTD) by utilizing two featurizers: a coarse MACE featurizer for quality assessment and a contrastive GNN featurizer for novelty evaluation. They conducted experiments to benchmark generative models and analyzed the impact of MACE feature conditioning on model performance.
Results
The results indicate that CFTD effectively captures the quality of generated materials while identifying memorization. The benchmarking of generative models showed that relaxing generated structures generally enhances quality but may increase memorization. The MACE-conditional generative model exhibited improved stability and novelty compared to models without this conditioning.
Implications
The findings suggest that pretrained MLIPs can serve as valuable tools for guiding material generation, potentially leading to the discovery of novel materials with desired properties. The proposed evaluation metrics can enhance the assessment of generative models in materials science.
PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction
Time Series
- PiDDM integrates physics-based degradation kinetics into neural network training for improved SOH prediction.
- The framework encourages physically consistent capacity fade under diverse operating conditions.
- PiDDM outperforms standard MLP and baseline PINN models in terms of prediction accuracy and extrapolation performance.
- The model effectively captures accelerated end-of-life degradation without producing non-physical predictions.
Read more
PiDDM: Physics-Informed Differentiable Degradation Modeling for Lithium-Ion Battery State-of-Health Prediction
Summary
The paper presents a novel framework called Physics-Informed Differentiable Degradation Modeling (PiDDM) aimed at improving the prediction of lithium-ion battery state-of-health (SOH). Traditional data-driven models often struggle with generalization across different cycling protocols and can yield physically implausible predictions during long-term extrapolation. PiDDM addresses these issues by integrating empirical Arrhenius degradation kinetics related to solid electrolyte interphase (SEI) growth and lithium inventory loss into the neural network's training objective. This incorporation allows the model to maintain physically consistent capacity fade across various operating conditions. The authors evaluated PiDDM using a public dataset of 55 batteries cycled under six distinct protocols, demonstrating that it achieves lower average prediction errors compared to standard multilayer perceptron (MLP) models and baseline physics-informed neural networks (PINNs). Notably, in extrapolation tests where models were trained on the first 90% of each battery's cycle life, PiDDM effectively captured accelerated end-of-life degradation while avoiding non-physical capacity regeneration seen in other models. The results indicate that embedding degradation physics into neural network training enhances both predictive accuracy and physical consistency, offering a promising approach for practical battery health monitoring.
Methodology
The methodology involves developing a physics-informed neural network that embeds empirical degradation kinetics into the loss function, allowing the model to learn from voltage, current, and time-series measurements while ensuring physical consistency in capacity fade.
Results
PiDDM achieved the lowest average prediction error compared to MLP and baseline PINN models. In extrapolation tests, it accurately predicted accelerated end-of-life degradation while avoiding non-physical capacity regeneration, demonstrating superior performance in real-world applications.
Implications
The findings suggest that integrating physics-based constraints into machine learning models can significantly enhance the reliability and accuracy of battery health predictions, which is crucial for the effective management of energy storage systems in various applications, including electric vehicles and renewable energy integration.
Adaptive FastOPD: Progress-Aware Rollout Horizon Expansion for Efficient On-Policy Distillation
Efficient ML
Large Language Models
Reinforcement Learning
- Introduction of Adaptive FastOPD, a progress-aware strategy for efficient on-policy distillation.
- Significant reduction in training time (49.1–71.2%) compared to traditional OPD methods.
- Adaptive expansion of rollout horizon based on learning progress and utilization metrics.
- Robust performance across different hyperparameter settings and teacher-student pairs.
Read more
Adaptive FastOPD: Progress-Aware Rollout Horizon Expansion for Efficient On-Policy Distillation
Summary
The paper introduces Adaptive FastOPD, a novel approach to on-policy distillation (OPD) that addresses the inefficiencies associated with the rollout process in training student models. Traditional OPD methods often suffer from high computational costs due to variable response lengths and the presence of long responses that delay batch completion. Adaptive FastOPD proposes a progress-aware strategy that expands the rollout horizon only when learning has plateaued and the current horizon is sufficiently utilized. This is achieved by monitoring four teacher-student signals relative to their initial values upon entering each horizon, allowing for a more responsive and adaptive approach to rollout expansion. The method significantly reduces training time by 49.1–71.2% compared to existing OPD methods while maintaining robust performance across various hyperparameter settings. The experiments conducted on two teacher-student pairs demonstrate that Adaptive FastOPD outperforms both vanilla OPD and fixed-schedule FastOPD, showcasing its efficiency and effectiveness in optimizing the training process.
Methodology
The methodology involves a continuous monitoring system that assesses learning progress in the boundary region of the rollout horizon. Four complementary signals are measured and normalized against a baseline established at the start of each horizon. The horizon is expanded only when these signals indicate a plateau in learning progress and when the current horizon has been sufficiently utilized, thus preventing inefficiencies caused by long responses.
Results
The experiments reveal that Adaptive FastOPD achieves the highest average performance compared to traditional OPD and fixed-schedule FastOPD, while also significantly reducing the overall training time. The method demonstrates stability and robustness across various hyperparameter configurations.
Implications
The findings suggest that Adaptive FastOPD can be effectively applied to enhance the efficiency of on-policy distillation in various machine learning contexts, particularly in scenarios where computational resources are limited or where training time is critical. This approach could lead to more efficient training protocols in large-scale machine learning applications.
TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning via Orthogonal Gradient Projection and Optimizer State Entanglement
Federated Learning
Large Language Models
NLP
- TriShield effectively defends against NeuroImprint-style privacy backdoors in federated learning.
- The defense framework operates entirely on the client side, requiring no server cooperation.
- Zero utility loss is achieved, preserving model accuracy while eliminating private data leakage.
- The mutual information between reconstructed data and original samples is proven to be zero after applying TriShield.
Read more
TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning via Orthogonal Gradient Projection and Optimizer State Entanglement
Summary
The paper introduces TriShield, a novel defense mechanism against privacy backdoors in federated fine-tuning of large language models (LLMs). The motivation stems from the NeuroImprint attack, which exploits vulnerabilities in parameter-efficient fine-tuning (PEFT) methods, allowing a malicious parameter server to reconstruct client training data with high fidelity. TriShield is designed to operate entirely on the client side without requiring server cooperation, ensuring zero utility loss while effectively neutralizing the NeuroImprint threat. The defense framework consists of three layers: a Parameter Artifact Detector to identify memory-neuron signatures, a Stateful Virtual Iteration mechanism that entangles gradients across virtual steps, and a Zero-Utility Orthogonal Projection operator that eliminates private memorization components from gradient updates. The authors provide theoretical proofs demonstrating that after applying the second and third layers, the mutual information between the reconstructed data and the original training sample is zero. Experimental results on models like GPT-2 and Llama-Guard-3-1B show that TriShield reduces the reconstruction rate of NeuroImprint to 0% while maintaining or improving training accuracy, with minimal additional computational overhead.
Methodology
TriShield employs a three-layer defense strategy: (1) Parameter Artifact Detector identifies suspicious neuron patterns, (2) Stateful Virtual Iteration entangles gradient updates to thwart inversion attacks, and (3) Zero-Utility Orthogonal Projection removes components related to private memorization from gradients. The approach is mathematically validated to ensure zero mutual information between reconstructed data and original training samples.
Results
Experiments demonstrate that TriShield reduces the reconstruction rate of the NeuroImprint attack to 0% across various attack variants on models like GPT-2 and Llama-Guard-3-1B, while maintaining or improving training accuracy and incurring less than 5% additional GPU computation overhead.
Implications
TriShield provides a robust solution for enhancing privacy in federated learning environments, particularly in applications involving sensitive data. Its zero-utility-loss characteristic makes it suitable for production deployments where model performance is critical.
Distilling Knowledge from Large Language Models into Lightweight Reinforcement Learning Agents for Autonomous Cyber Operations
Reinforcement Learning
Large Language Models
Efficient ML
- Proposed a method for LLM-to-RL policy distillation, achieving effective knowledge transfer from a large model to a lightweight agent.
- Evaluated the transferability of the distilled policy across multiple simulated network environments.
- Analyzed teacher-guided RL stabilization strategies, highlighting limitations in policy alignment between teacher and reward-driven approaches.
- Demonstrated that the distilled RL agent maintains comparable defensive performance to the original LLM.
Read more
Distilling Knowledge from Large Language Models into Lightweight Reinforcement Learning Agents for Autonomous Cyber Operations
Summary
This paper addresses the challenges of using Reinforcement Learning (RL) agents for Autonomous Cyber Operations (ACO) in the face of evolving cyber threats. Traditional RL methods require extensive exploration, leading to unstable behavior and poor initial decision-making. The authors propose leveraging a Large Language Model (LLM) pretrained on cybersecurity data to enhance defensive decision-making. They demonstrate that an 8-billion parameter LLM can outperform a baseline RL agent in a modified CybORG CAGE Challenge 2 environment through prompt engineering. The authors introduce an online policy distillation framework that transfers the LLM's defensive policy into a lightweight RL agent with only 64,910 parameters, significantly reducing model size while preserving effective defensive capabilities. The study evaluates the transferability of this approach across various CybORG scenarios and analyzes teacher-guided RL stabilization strategies, revealing that no strategy consistently surpasses the optimized teacher policy. The findings suggest that LLMs can serve as valuable sources of expertise for autonomous cyber defense, and policy distillation offers a practical method for operationalizing advanced cybersecurity models into efficient agents.
Methodology
The authors utilized prompt engineering to optimize the defensive behavior of an 8-billion parameter LLM without environment-specific fine-tuning. They then developed an online policy distillation framework to transfer this optimized policy into a lightweight RL agent. The approach was evaluated across various CybORG scenarios, assessing its performance and transferability.
Results
The distilled RL agent outperformed the baseline RL agent in the modified CybORG CAGE Challenge 2 environment. The study found that the lightweight agent maintained effective defensive capabilities despite its reduced parameter count. Additionally, the analysis of teacher-guided RL strategies showed that none consistently outperformed the optimized teacher policy, indicating potential misalignment issues.
Implications
The findings suggest that LLMs can effectively enhance autonomous cyber defense strategies, providing a pathway for integrating advanced models into lightweight, deployable agents. This approach could significantly improve the efficiency and effectiveness of cybersecurity operations in real-world environments.
Federated Foundation Models Fine-Tuning with Heterogeneous Compressed Clients
Federated Learning
NLP
Multimodal
- FedSLM addresses the resource-asymmetry challenge in federated learning of foundation models.
- The framework uses SVD-based decomposition to create compressed client models that maintain structural compatibility.
- A two-stage aggregation protocol is implemented to synchronize and fuse model updates effectively.
- The method shows improved performance over existing federated learning approaches under both IID and non-IID conditions.
Read more
Federated Foundation Models Fine-Tuning with Heterogeneous Compressed Clients
Summary
The paper addresses the resource-asymmetry challenge in federated learning (FL) of foundation models (FMs), where institutions with valuable domain-specific data often lack the resources to host large models. Existing methods for heterogeneous federated learning struggle with trade-offs that compromise model performance or efficiency. The authors propose FedSLM, a parameter-centric framework that utilizes Singular Value Decomposition (SVD) to create self-contained client models that are compatible for aggregation. FedSLM employs a two-stage aggregation protocol to synchronize lightweight adapters within compression groups and fuse full-rank reconstructions across groups. Additionally, it incorporates a weak-to-strong elicitation step with auxiliary confidence loss to transfer knowledge to the server while addressing compression artifacts. The framework is theoretically grounded with guarantees for aggregation and alignment, and experimental results demonstrate that FedSLM outperforms existing federated learning baselines on natural language and vision-language tasks, achieving a significant reduction in GPU memory usage.
Methodology
FedSLM formulates a coupled optimization problem involving client-side adapters, server-side reconstructed weights, and the final server model. It employs SVD-based decomposition to derive compressed client models and implements a two-stage aggregation protocol for model updates, ensuring structural alignment and effective knowledge transfer.
Results
Experimental evaluations indicate that FedSLM outperforms existing federated learning baselines on various benchmarks, demonstrating superior performance in both IID and non-IID settings. The client models are able to operate with roughly 50% of the GPU memory required by the full model, showcasing significant efficiency improvements.
Implications
The proposed framework has potential applications in domains where data privacy is crucial, such as healthcare and finance, enabling institutions to leverage large foundation models without compromising data security or requiring extensive computational resources.
Encryption-Compatible Clustered Federated Learning via Distributed Expectation-Maximization over Metadata
Federated Learning
- Introduces FLAMECHE, a framework for encryption-compatible clustered federated learning.
- Addresses the CFL trilemma by balancing privacy, communication, and computation.
- Utilizes a distributed Expectation-Maximization procedure for clustering based on metadata.
- Demonstrates improved effectiveness of client models through extensive experiments.
Read more
Encryption-Compatible Clustered Federated Learning via Distributed Expectation-Maximization over Metadata
Summary
The paper addresses the challenges of Clustered Federated Learning (CFL), particularly in managing the trade-offs between privacy, communication costs, and computational efficiency, referred to as the CFL trilemma. Existing methods often compromise one of these dimensions to enhance the others, especially when using metadata for clustering, which raises privacy concerns. The authors propose FLAMECHE, a novel framework that reformulates metadata-based CFL as a distributed Expectation-Maximization (EM) procedure. This approach restricts server updates to additive operations, making it compatible with secure federated learning mechanisms. The framework utilizes randomly initialized neural networks to compute metadata, avoiding reliance on prior knowledge of client data distributions. Extensive experiments demonstrate that FLAMECHE significantly improves client model effectiveness while maintaining encryption compatibility, thus enhancing its position within the CFL trilemma.
Methodology
The authors reformulate the clustering process in CFL as a distributed Expectation-Maximization (EM) procedure. This involves using low-dimensional metadata representations of client datasets while ensuring that server updates are restricted to additive operations. The metadata is generated using randomly initialized neural networks, which allows the framework to operate without prior knowledge of client heterogeneity.
Results
The experimental results indicate that FLAMECHE enhances the effectiveness of client models across various datasets and heterogeneous scenarios. The framework successfully enables encryption-compatible metadata-based clustering, positioning itself favorably within the CFL trilemma by improving two dimensions (effectiveness and privacy) without compromising the third (communication).
Implications
FLAMECHE has significant implications for federated learning applications, particularly in scenarios where data privacy is paramount. It allows for more effective model training in environments with non-IID data distributions while ensuring compliance with privacy-preserving protocols. This could lead to broader adoption of federated learning in sensitive domains such as healthcare and finance.
DASH-OPD: Discrepancy-Aware Switching with Hysteresis for On-Policy Distillation
Reinforcement Learning
Large Language Models
Robotics
- DASH-OPD introduces a discrepancy-aware switching mechanism for on-policy distillation.
- The method accumulates drift and recovery evidence over multiple turns to inform switching decisions.
- DASH-OPD significantly reduces the number of teacher-generated turns and executor switches compared to existing methods.
- It achieves the highest success rates across various evaluation splits in the ALFWorld environment.
Read more
DASH-OPD: Discrepancy-Aware Switching with Hysteresis for On-Policy Distillation
Summary
The paper introduces DASH-OPD, a novel method for On-Policy Distillation (OPD) that addresses the challenges faced in multi-turn agent scenarios, particularly the exposure bias caused by early student errors. Traditional curriculum learning methods regulate teacher support but lack the ability to determine when such support is necessary. DASH-OPD employs a discrepancy-aware switching mechanism that adaptively and bidirectionally switches between student and teacher models based on accumulated evidence of performance. It calculates a mean log-probability ratio to assess the discrepancy between the two executors, forming drift and recovery signals that inform the switching process. This hysteretic approach prevents frequent switches due to transient fluctuations, thereby enhancing training stability. The method was evaluated on the ALFWorld environment, where it demonstrated superior performance compared to existing baselines, achieving higher success rates and greater efficiency in training and deployment.
Methodology
DASH-OPD calculates mean log-probability ratios between student and teacher models to create drift and recovery signals. These signals are normalized and accumulated over multiple turns, allowing the system to switch executors based on evidence exceeding predefined thresholds. The method employs reverse-KL and forward-KL distillation losses during training to optimize performance.
Results
DASH-OPD outperformed all baseline methods on the ALFWorld environment, achieving the highest success rates and requiring fewer interaction rounds. It reduced teacher-generated turns by 35.4% and executor switches by 77.1% compared to Guided-OPD, demonstrating significant improvements in training and deployment efficiency.
Implications
The findings suggest that DASH-OPD can be applied to enhance the performance of multi-turn agents in various domains, particularly in scenarios where exposure bias is a concern. This method could lead to more efficient training processes and improved agent performance in real-world applications.
Learning features from Newton's algorithm: a way to accelerate nonlinear parametrized PDE solvers
Optimization
Theory
Efficient ML
- Introduces a two-stage Newton initial guess strategy leveraging learned features.
- Constructs two complementary reduced spaces for solution and correction.
- Employs a regression model for surrogate solution approximation.
- Demonstrates significant reductions in Newton iterations and CPU time.
Read more
Learning features from Newton's algorithm: a way to accelerate nonlinear parametrized PDE solvers
Summary
This paper presents a novel two-stage strategy to enhance the efficiency of solving nonlinear parametrized partial differential equations (PDEs) using Newton's method. The authors propose a method that leverages features learned from a parameter-space sampling and a database of precomputed solutions. By utilizing discrete Newton trajectories, the approach constructs two complementary reduced spaces: one for solution features derived from converged states and another for corrective search directions based on intermediate Newton increments. For unseen parameters, a regression model predicts a surrogate solution, followed by a residual-minimizing correction using a GMRES-based approach. This correction is computationally inexpensive and serves as an initial guess for the high-fidelity Newton method, significantly reducing the number of iterations required for convergence. The methodology is shown to be weakly intrusive, requiring only residual evaluations and a small least-squares problem. Numerical experiments demonstrate substantial speedups compared to traditional surrogate initialization methods, indicating the approach's applicability to a wide range of large-scale nonlinear problems.
Methodology
The proposed methodology involves two main stages: first, learning features from previous Newton computations to create a solution feature space and a corrective search direction feature space. A regression model predicts a surrogate solution for unseen parameters, followed by a residual-minimizing correction using a GMRES-based approach. This correction is then used as an initial guess for the high-fidelity Newton method.
Results
The numerical experiments conducted on representative PDE problems show quantifiable speedups in convergence compared to standalone surrogate initialization methods. The proposed approach significantly reduces the number of Newton iterations and overall CPU time, demonstrating its effectiveness in accelerating the solution of nonlinear parametrized PDEs.
Implications
This methodology has the potential to improve the efficiency of solving complex nonlinear problems in various fields such as engineering, physics, and applied mathematics, where rapid and accurate solutions to PDEs are critical. It opens avenues for further research in model-order reduction and machine learning applications in numerical analysis.
Policy Gradient Steering: Interventions from Behavioral Objectives
Reinforcement Learning
- Introduction of Policy Gradient Steering (PGS) as a method for behavioral adaptation in reinforcement learning.
- Demonstration of PGS's effectiveness in a controlled gridworld environment, chess, and competitive football.
- Validation of the compositional nature of behavioral interventions created by PGS.
- Evidence that PGS can alter team behaviors in multi-agent scenarios and transfer effects across different opponents.
Read more
Policy Gradient Steering: Interventions from Behavioral Objectives
Summary
This paper introduces Policy Gradient Steering (PGS), a novel method for dynamically altering the behavior of trained policies in reinforcement learning without the need for retraining. Existing steering methods often fail to effectively control decision-making in environments like gridworlds. PGS addresses this by framing steering as a reinforcement learning problem, where it accumulates gradients from a small set of rollouts or demonstrations to create a removable task vector. The authors validate PGS through experiments in a two-route gridworld, chess puzzles, and competitive football, demonstrating its ability to construct temporary behavioral adaptations that are composable and transferable across different contexts. The results indicate that PGS can effectively modify specific behaviors in multi-agent settings and that these modifications can be combined without retraining, showcasing the flexibility and utility of policy gradients in behavioral interventions.
Methodology
The methodology involves accumulating gradients from trajectories scored under a temporary behavioral objective to construct an additive steering vector. PGS operates by assigning credit to actions that lead to specific outcomes rather than relying on post-outcome representations. This allows for the creation of interventions that can be scaled, composed, and removed at inference time while keeping the base policy unchanged.
Results
The experiments showed that PGS successfully calibrated and reversed interventions in a two-route gridworld. In chess, independently fitted PGS vectors were found to retain their effects when combined, indicating that tactical objectives can constructively accumulate. In competitive football, PGS was able to modify team behaviors effectively, with the results transferring across different opponents, demonstrating the versatility of the method across various decision-making domains.
Implications
The findings suggest that PGS can be applied in real-world scenarios where dynamic behavioral adaptations are necessary without the overhead of retraining. This could be particularly useful in applications such as robotics, game AI, and any domain requiring adaptable decision-making policies.
MUGEN: A Unified Framework for Efficient Motion Understanding and Generation
Multimodal
Generative Models
Efficient ML
- MUGEN eliminates the need for discrete codebooks, allowing for high-quality motion generation and understanding.
- The framework uses continuous latent slots and depth-routed hidden states to enhance representation and efficiency.
- MUGEN achieves state-of-the-art performance on multiple benchmarks, including HumanML3D and SnapMoGen.
- The system operates with significantly lower latency and computational requirements compared to existing methods.
Read more
MUGEN: A Unified Framework for Efficient Motion Understanding and Generation
Summary
The paper introduces MUGEN, a unified framework designed to enhance the efficiency and quality of motion understanding and generation by integrating human motion with language processing. Traditional systems often rely on discrete codebooks, which limit the quality of generated motions and increase computational costs due to complex decoding processes. MUGEN addresses these limitations by employing a continuous latent representation that allows for both text-to-motion generation and motion-to-text understanding without the need for a codebook. The framework utilizes an Adaptive-Length AutoEncoder (ALAE) to compress motion into a few continuous latent slots, which are then used by a language model for both generating motions from text and understanding motions through language. Key innovations include depth-routed hidden states that enable each latent slot to access specific layers of the transformer model, and a calibrated low-rank factor head that predicts a joint distribution across the latent slots, allowing for a single draw to capture the necessary variation. The results demonstrate that MUGEN outperforms existing baselines in various metrics, achieving superior quality in motion generation and understanding while significantly reducing computational costs.
Methodology
MUGEN employs an Adaptive-Length AutoEncoder (ALAE) to convert motion sequences into continuous latent slots. It utilizes a language model to generate motions from text and to understand motions through language. The architecture features depth-routed hidden states for efficient evidence retrieval and a calibrated low-rank factor head for predicting joint distributions across latent slots, enabling a single draw to capture necessary variations.
Results
MUGEN outperformed language-model baselines on FID, retrieval precision, and matching distance metrics on the HumanML3D dataset, achieving the best CIDEr and BLEU@4 scores. On SnapMoGen, it surpassed the discrete-token state of the art across all retrieval ranks and CLIP alignment metrics. The framework demonstrated a significant reduction in inference time, processing motions in approximately 9 ms, which is 6–14 times faster than competing methods.
Implications
MUGEN's approach could lead to more efficient and effective physical AI systems capable of understanding and generating human-like motions, with potential applications in robotics, animation, and interactive AI systems. Its unified framework may also inspire further research into multimodal learning and efficient generative models.
Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
Generative Models
- Introduces Explorative Modeling (XM) as a new paradigm for generative modeling.
- Enables end-to-end training by factoring the training loop instead of the generation process.
- Demonstrates significant performance improvements with increased exploration in generative tasks.
- Achieves near-state-of-the-art results in image generation with improved efficiency metrics.
Read more
Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation
Summary
This paper introduces Explorative Modeling (XM), a novel paradigm aimed at enhancing generative modeling by enabling end-to-end training. Traditional generative models often factor the generation process, which hinders their ability to handle multimodal distributions effectively. XM proposes to factor the training loop instead, allowing models to explore multiple candidate matches between generated outputs and actual data during training. This approach not only captures multimodal distributions more effectively but also introduces a third pretraining axis—exploration—beyond parameters and data. The authors demonstrate that increasing exploration leads to significant performance improvements across various domains, including images, video, and language. Notably, exploration enhances efficiency metrics such as FLOP efficiency, sample efficiency, and parameter efficiency, while also achieving near-state-of-the-art results in image generation tasks. Additionally, XM facilitates end-to-end reconstructive generative modeling, achieving comparable performance to diffusion models with significantly fewer inference steps. Overall, the findings position XM as a transformative approach in generative modeling, unlocking new capabilities and efficiencies.
Methodology
The authors propose a new training paradigm called Explorative Modeling, where the training loop is factored to explore multiple candidate matches between model generations and actual data. This allows the model to commit to distinct modes during training, effectively capturing multimodal distributions. The methodology involves scaling exploration as a pretraining axis and measuring its impact on performance across various generative tasks.
Results
The implementation of XM shows that increasing exploration leads to performance gains ranging from 7% to 36% as data scales and 13% to 23% as model sizes grow. The approach improves FLOP efficiency by 4.1×, sample efficiency by 6.2×, and parameter efficiency by 47%. Additionally, XM achieves a near-state-of-the-art FID score of 1.43 on ImageNet without guidance and enables end-to-end generative modeling that matches diffusion models with significantly fewer inference steps.
Implications
The findings suggest that Explorative Modeling could revolutionize generative modeling by allowing for more efficient training and better handling of multimodal distributions. This has potential applications in various fields such as computer vision, natural language processing, and any domain requiring generative modeling capabilities.
Latent Lie-Poisson Neural Networks (LLPNNs): Discovering the motion of Lie-Poisson systems through observable data and latent dynamics
Robotics
Optimization
Theory
- LLPNNs provide a structure-preserving method for learning Lie-Poisson dynamics from observable data.
- The framework incorporates geometric elements to handle both regular and degenerate Hamiltonian systems.
- Numerical experiments demonstrate the method's effectiveness across various applications, showcasing its robustness and accuracy.
Read more
Latent Lie-Poisson Neural Networks (LLPNNs): Discovering the motion of Lie-Poisson systems through observable data and latent dynamics
Summary
The paper introduces Latent Lie-Poisson Neural Networks (LLPNNs), a novel framework designed to learn the dynamics of Lie-Poisson systems directly from observable data. Traditional methods struggle with these systems due to the unobservable nature of their momentum variables, which are crucial for accurate modeling. LLPNNs address this challenge by leveraging three key geometric components: the development of a Hamiltonian decoder or pseudo-Lagrangian encoder, the construction of latent trajectories using a Noether invariant from Lie-Poisson symmetry reduction, and the reconstruction of both observable and latent dynamics through Lie-Poisson flows with Magnus-based updates. This approach not only preserves the geometric structure of the systems but also applies effectively to both regular and degenerate Hamiltonian systems. The efficacy of LLPNNs is demonstrated through numerical experiments on three systems: a generalized rigid body, Kirchhoff’s underwater vehicle, and an optimal control problem involving interacting vehicles. The results indicate that LLPNNs achieve high predictive accuracy, robustness against noise, and competitive performance with limited datasets and simple neural network architectures.
Methodology
The methodology involves constructing LLPNNs that utilize a Hamiltonian decoder or pseudo-Lagrangian encoder, creating latent trajectories based on a Noether invariant, and employing Lie-Poisson flows combined with Magnus-based updates for reconstructing dynamics. This geometric approach ensures the preservation of the underlying structure of the systems being modeled.
Results
The numerical experiments conducted on three distinct systems showed that LLPNNs can accurately predict long-term dynamics, maintain robustness in the presence of noise, and perform competitively with minimal data and lightweight architectures.
Implications
The development of LLPNNs has significant implications for fields requiring accurate modeling of mechanical systems, such as robotics, control systems, and fluid dynamics. The ability to learn from observable data while preserving the geometric structure of the underlying dynamics opens new avenues for research and application in physics-informed machine learning.
The Grokked Illusion: True Equilibrium Mitigates Catastrophic Forgetting
Theory
Optimization
- Introduction of the 'grokked illusion,' highlighting the disconnect between generalization and robustness in neural networks.
- Demonstration of the high-entropy robustness advantage, where high-entropy states retain prior knowledge better than conventionally trained models.
- Use of singular value decomposition to link effective rank and robustness, suggesting richer feature representations are crucial for mitigating catastrophic forgetting.
Read more
The Grokked Illusion: True Equilibrium Mitigates Catastrophic Forgetting
Summary
This paper investigates the relationship between generalization performance and robustness in neural networks (NNs), particularly focusing on the phenomenon of catastrophic forgetting. The authors introduce the concept of the 'grokked illusion,' where models achieving perfect generalization can still exhibit fragility when faced with new data. They explore this using a modular arithmetic task, comparing AdamW-trained transformers with high-entropy models sampled via Wang-Landau Molecular Dynamics. The study reveals that while both models can achieve similar test accuracies, the high-entropy models maintain significantly higher accuracy when subjected to noise injection, demonstrating a robustness advantage. The authors employ singular value decomposition to analyze the neural network weights, finding that high-entropy networks possess a higher effective rank, which correlates with richer feature representations that buffer against catastrophic forgetting. This research provides new insights into the properties that contribute to a model's robustness beyond mere generalization.
Methodology
The authors utilized a modular arithmetic task as a benchmark and employed a single-layer Transformer model. They compared AdamW-trained models with high-entropy models generated through Wang-Landau Molecular Dynamics. A noise injection experiment was conducted to assess the robustness of both model types under conditions of new data memorization.
Results
The results indicated that AdamW-trained models experienced a significant drop in test accuracy from 100% to below 75% when faced with noise injection, while high-entropy models maintained approximately 95% accuracy. The singular value decomposition analysis revealed that high-entropy models had a higher effective rank in their layers, suggesting a more complex representation of features.
Implications
This research suggests that achieving high generalization does not guarantee robustness, which has implications for the design of neural networks in applications requiring resilience to new information. It encourages a reevaluation of training strategies to enhance model robustness, particularly in continual learning scenarios.
FairDiffuseVQVAE: Sampling-Time Fairness in Tabular Diffusion via Conditional Refinement of Vector-Quantized Latents
Generative Models
- FairDiffuseVQVAE separates fidelity and fairness in synthetic data generation.
- The architecture consists of a vector-quantized autoencoder followed by a diffusion refiner.
- Fairness is enforced at sampling time, avoiding the need for explicit fairness loss terms.
- The model achieves superior fairness metrics compared to existing methods while maintaining low correlation error.
Read more
FairDiffuseVQVAE: Sampling-Time Fairness in Tabular Diffusion via Conditional Refinement of Vector-Quantized Latents
Summary
The paper introduces FairDiffuseVQVAE, a novel two-stage architecture designed to generate synthetic tabular data that balances fidelity and fairness. Traditional tabular diffusion models excel in generating realistic data but lack mechanisms for ensuring fairness, while fairness-aware models often compromise sample quality. FairDiffuseVQVAE addresses this by decoupling fidelity from fairness: the first stage employs a vector-quantized autoencoder without fairness constraints, while the second stage utilizes a continuous diffusion refiner that conditions on both the reconstructed data and a protected attribute. This approach allows for demographic parity to be enforced at sampling time through uniform sampling of the protected attribute, rather than through competing loss terms during training. The authors demonstrate that their model significantly outperforms existing fairness-specific baselines on the Yang et al. benchmark, achieving high demographic parity and equalized odds ratios, while maintaining low correlation error, albeit with a slight trade-off in utility.
Methodology
The methodology involves a two-stage architecture: the first stage is a vector-quantized autoencoder that generates initial data without fairness constraints. The second stage is a continuous diffusion refiner that conditions on the output of the first stage and a protected attribute, using classifier-free guidance to ensure fairness during sampling. This design allows for demographic parity to be achieved by uniformly sampling the protected attribute at inference time.
Results
FairDiffuseVQVAE achieved a mean Demographic Parity Ratio of 0.702, which is a 47% improvement over FairTabDDPM, and a mean Equalized Odds Ratio of 0.686, doubling the performance of previous methods. Additionally, it recorded the lowest mean pair-wise correlation error of 0.034, outperforming TabSyn's 0.041, while incurring a trade-off of approximately 15 AUC points in utility.
Implications
The findings suggest that FairDiffuseVQVAE can be effectively used in applications requiring synthetic tabular data generation, such as privacy-preserving data sharing and bias mitigation in machine learning models. By ensuring fairness at the sampling stage, it opens avenues for more equitable data practices in sensitive domains like healthcare and finance.
Versatile On-device Adaptation at the Edge by Unifying Few-shot, Zero-shot, Continual, and In-context Learning
Efficient ML
Multimodal
Audio & Speech
- Introduction of embedder-centric learning (ECL) framework that unifies multiple learning scenarios.
- ECL enables on-device adaptation for resource-constrained edge devices without cloud reliance.
- Demonstrated state-of-the-art performance in few-shot learning and established hardware baselines for continual learning.
- First hardware demonstrations of zero-shot and in-context learning in real-world applications.
Read more
Versatile On-device Adaptation at the Edge by Unifying Few-shot, Zero-shot, Continual, and In-context Learning
Summary
This paper addresses the limitations of current edge devices that rely on fixed inference algorithms, which cannot adapt to new data or user-specific requirements. The authors propose a novel framework called embedder-centric learning (ECL) that unifies four online learning scenarios: few-shot learning (FSL), continual learning (CL), zero-shot learning (ZSL), and in-context learning (ICL). ECL is designed to operate on resource-constrained devices, enabling them to personalize predictions without the need for cloud-based retraining. The framework splits each learning scenario into a shared embedder and a scenario-specific head, allowing for efficient adaptation across various sensory modalities. The authors demonstrate the effectiveness of ECL through silicon implementations across four real-world use cases, achieving state-of-the-art performance in FSL character recognition and establishing hardware baselines for CL in keyword spotting, as well as pioneering demonstrations of ZSL and ICL. This work paves the way for more versatile and adaptive smart devices that can learn and evolve in real-time at the edge.
Methodology
The authors developed the embedder-centric learning (ECL) framework, which separates the learning process into a shared embedder neural network and scenario-specific heads. This design allows for efficient representation of data in a reduced-dimensional space, facilitating the adaptation of the model to various learning scenarios while maintaining a small memory footprint. The framework was implemented in silicon and tested across multiple real-world use cases.
Results
ECL achieved a state-of-the-art accuracy of 96.8% for 5-way 1-shot and 83.3% for 32-way 1-shot in FSL character recognition. It also established a hardware baseline for CL in keyword spotting with 71.8% accuracy for 200-way 5-shot. Additionally, the first hardware demonstrations of ZSL and ICL were reported, achieving 60.6% and 46.2% accuracy, respectively, while operating within micro-to-milliwatt power budgets.
Implications
The proposed ECL framework has significant implications for the development of smart edge devices capable of real-time learning and adaptation, enhancing user personalization and privacy by reducing reliance on cloud services. This versatility can lead to improved applications in various domains, including health monitoring, smart home devices, and autonomous systems.
Sample Efficient Hierarchical Reinforcement Learning via Best Policy Identification
Reinforcement Learning
Theory
Efficient ML
- HBPI-UCRL learns high-level and low-level policies in parallel, enhancing sample efficiency.
- The algorithm's sample complexity is polynomial in problem parameters, outperforming non-hierarchical methods.
- Two conditions are identified that ensure the learnability of parallel HRL.
- The paper provides the first explicit sample complexity result for parallel HRL settings.
Read more
Sample Efficient Hierarchical Reinforcement Learning via Best Policy Identification
Summary
This paper introduces HBPI-UCRL, a novel model-based algorithm for hierarchical reinforcement learning (HRL) that concurrently learns high-level and low-level policies. The authors leverage the relationship between high-level transitions and multi-step low-level transitions, establishing two conditions on low-level dynamics that enable efficient learning in parallel HRL settings. The algorithm is shown to have polynomial sample complexity, significantly improving upon non-hierarchical counterparts, particularly in sparse-reward, goal-directed scenarios. The theoretical framework provided offers a foundation for understanding the benefits of HRL, including efficient exploration and improved sample efficiency. The authors also present empirical results that validate the theoretical claims, demonstrating the practical advantages of their approach in various HRL tasks.
Methodology
The authors develop a model-based algorithm, HBPI-UCRL, that employs best policy identification (BPI) for episodic HRL. The algorithm collects data through episodes and maintains high-probability confidence bounds on the semi-Markov decision process (SMDP) dynamics. The learning process is structured around two key conditions that relate the quality of subproblem solutions to the accuracy of the SMDP dynamics, ensuring PAC-learnability.
Results
HBPI-UCRL achieves a sample complexity upper bound of order eO(SAH4H6/ε2), which is independent of the number of high-level states and subproblems. In sparse-reward, goal-directed settings, it achieves a sample complexity of order eO(SAH2H2/ε2), which is significantly lower than that of the non-hierarchical BPI-UCRL algorithm. The results demonstrate the theoretical and empirical advantages of the proposed approach.
Implications
The findings of this paper have significant implications for the design of efficient HRL algorithms, particularly in environments with sparse rewards. The theoretical guarantees provided can guide future research in HRL, potentially leading to more robust and sample-efficient learning methods applicable in various domains, including robotics and complex decision-making tasks.
Benchmarking the Residual: What Long-Horizon Evaluations Add Beyond Matched Short-Task Performance
NLP
Large Language Models
Theory
- Long-horizon benchmarks reveal performance degradation but do not explain the underlying causes of failure.
- The concept of 'horizon residual' is introduced to measure the difference between actual success and predicted outcomes from short tasks.
- A structured evaluation protocol is proposed to distinguish between deployment results and diagnostic insights.
- The paper emphasizes the importance of clear reporting and intervention strategies to explore causal hypotheses.
Read more
Benchmarking the Residual: What Long-Horizon Evaluations Add Beyond Matched Short-Task Performance
Summary
This position paper discusses the challenges of evaluating long-horizon tasks in machine learning, particularly in the context of language-model agents. The authors highlight that while longer tasks often lead to lower success rates, this does not inherently indicate a long-horizon failure mechanism. Instead, they introduce the concept of 'trajectory-induced degradation,' where earlier actions can complicate later decisions due to accumulated context, referred to as 'context rot.' The paper advocates for a benchmarking approach that compares actual full-task success against predictions derived from short, individual stages, termed the 'horizon residual.' This residual serves as a diagnostic tool to understand discrepancies between expected and actual performance. The authors propose a structured evaluation protocol that includes running agents on short, verifiable stages and combining these results to predict full-task outcomes. They emphasize the need for clear reporting requirements and an intervention agenda to explore the causal factors behind performance declines. Overall, the paper aims to refine the understanding of long-horizon evaluations and their implications for agent deployment and diagnostic assessments.
Methodology
The authors propose a benchmarking framework that involves running agents on short, verifiable stages to create a baseline prediction for full tasks. They then compare this prediction with actual performance in long-horizon evaluations to derive the horizon residual, which highlights discrepancies and guides further investigation.
Results
The paper illustrates that a decline in end-to-end success rates in longer tasks can often be attributed to ordinary error compounding rather than unique long-horizon failures. By comparing actual performance with predictions from short tasks, the authors demonstrate how the horizon residual can inform future research and diagnostics.
Implications
The findings suggest that refining evaluation methodologies for long-horizon tasks can enhance the deployment of language-model agents in real-world applications. Understanding the factors contributing to performance degradation can lead to better design of benchmarks and improved agent training strategies.
Prox: Training-Free FFN Activation Sparsity via Approximate Intermediate-Channel Salience in LLMs
Large Language Models
Efficient ML
- Prox leverages the SwiGLU intermediate state for effective channel selection without costly dense computation.
- The framework consists of two stages: a lightweight proxy construction and exact sparse computation.
- Prox achieves up to 1.99× speedup in end-to-end decoding at 70% FFN sparsity with minimal accuracy loss.
- The method is compatible with quantization and sparse attention, enabling further efficiency gains.
Read more
Prox: Training-Free FFN Activation Sparsity via Approximate Intermediate-Channel Salience in LLMs
Summary
The paper introduces Prox, a two-stage training-free framework aimed at enhancing the efficiency of feed-forward networks (FFNs) in large language models (LLMs) by implementing activation sparsity. FFNs are identified as significant contributors to memory traffic and computation during LLM inference. Existing training-free methods for activation sparsity often lead to model-quality degradation at high sparsity levels due to ineffective channel-selection strategies. The authors propose that the intermediate state of the SwiGLU activation function can serve as a robust channel-selection signal, although its direct computation is resource-intensive. Prox addresses this by constructing a channel mask based on the magnitude ranking of the intermediate state rather than its exact values. The framework consists of two stages: the first stage constructs a shared mask using input sparsity and quantized proxy weights, while the second stage computes the selected channels precisely, allowing for sparse execution across all projections. Evaluations across ten LLMs demonstrate that Prox consistently outperforms existing training-free baselines, achieving significant speedups in decoding while maintaining accuracy.
Methodology
Prox employs a two-stage approach: Stage 1 constructs a shared channel mask using input sparsity and quantized proxy weights, while Stage 2 computes the selected channels exactly using the original weights. This method allows for efficient sparse execution of the SwiGLU projections without incurring significant computational overhead.
Results
Prox was evaluated on ten LLMs from six model families, demonstrating superior performance compared to state-of-the-art training-free baselines, particularly at 60-70% FFN sparsity. The framework achieved up to a 1.99× end-to-end decoding speedup with minimal accuracy loss, showcasing its effectiveness in enhancing inference efficiency.
Implications
The findings suggest that Prox can significantly improve the deployment of LLMs in resource-constrained environments by reducing memory and computational demands. Its compatibility with quantization and sparse attention indicates potential for broader applications in optimizing large-scale models.
Beyond Feature and Structure Alignment: Learning Transferable Propagation Knowledge for Graph Foundation Models
Graph Learning
- ProGFM introduces a new perspective on knowledge transfer in graph foundation models by focusing on propagation relationships as transferable knowledge units.
- The model does not assume consistent semantics across features in different domains, allowing for more flexible knowledge transfer.
- A propagation relationship prototype bank is utilized to capture cross-domain transferable propagation knowledge.
- ProGFM shows improved performance in generalizing to unseen graph domains compared to traditional alignment-based methods.
Read more
Beyond Feature and Structure Alignment: Learning Transferable Propagation Knowledge for Graph Foundation Models
Summary
This paper addresses the limitations of existing Graph Foundation Models (GFMs) in transferring knowledge across diverse graph domains. Traditional methods primarily focus on feature and structure alignment, neglecting the exploration of transferable knowledge units inherent in graph data. The authors propose a novel framework called the Propagation-aware Graph Foundation Model (ProGFM), which identifies propagation relationships between edges and feature dimensions as transferable knowledge units. This approach allows ProGFM to adaptively aggregate information in unseen graph domains by leveraging a propagation relationship prototype bank. The paper highlights the importance of recognizing the heterogeneity in propagation patterns, which existing models often overlook. Through extensive experiments across various cross-domain transfer scenarios, ProGFM demonstrates superior generalization performance compared to existing methods, showcasing its strong cross-domain knowledge transfer capabilities.
Methodology
The authors developed ProGFM by modeling the propagation relationships between edges and feature dimensions, treating these relationships as transferable knowledge units. They constructed a propagation relationship prototype bank to facilitate the learning of cross-domain transferable propagation knowledge, enabling adaptive information aggregation in new graph domains.
Results
ProGFM was evaluated through extensive experiments across multiple cross-domain transfer scenarios, demonstrating strong cross-domain knowledge transfer capabilities and superior generalization performance compared to existing graph foundation models that rely on feature and structure alignment.
Implications
The findings suggest that recognizing and utilizing propagation relationships can significantly enhance the adaptability and effectiveness of graph foundation models in diverse applications, potentially impacting areas such as social networks, recommender systems, and other domains reliant on graph data.
Pyramidal Width Can Increase Under Vertex Insertion
Theory
Optimization
- The paper provides a counterexample to the conjecture that pyramidal width does not increase with vertex insertion.
- An exact proof is presented using integer supporting hyperplanes and rational arithmetic.
- The pyramidal width increased by approximately 41.1% in the counterexample.
- A dependency-free verifier is included to independently validate the results.
Read more
Pyramidal Width Can Increase Under Vertex Insertion
Summary
This paper addresses a conjecture by Lacoste-Julien and Jaggi regarding the behavior of pyramidal width in polytopes when vertices are added. The conjecture posited that the pyramidal width of a polytope should not increase upon the addition of a vertex, provided that all existing vertices remain. The author presents a counterexample in three-dimensional space, demonstrating that the pyramidal width can indeed increase with vertex insertion. Specifically, the paper constructs a polytope with six integer vertices, showing that the pyramidal width increases by approximately 41.1% when a new vertex is added. The proof leverages the relationship between pyramidal width and facial distance, employing integer supporting hyperplanes to certify the face lattices and using rational calculations to evaluate facial distances. Additionally, the paper includes a dependency-free verifier that independently checks the combinatorial aspects and the computed distances, reinforcing the validity of the findings. The results challenge the conjecture without undermining established convergence theorems that utilize pyramidal width as a geometric constant.
Methodology
The methodology involves constructing a counterexample with six integer vertices in R3, certifying the face lattices using integer supporting hyperplanes, and calculating facial distances through finite rational methods. The relationship between pyramidal width and facial distance is utilized to prove the conjecture's falsity.
Results
The main result is the demonstration that the pyramidal width of a polytope can increase upon vertex insertion, with specific values calculated for the counterexample: PWidth(P)² = 48/353 and PWidth(Q)² = 36/133, leading to an increase factor of approximately 1.410886779.
Implications
The findings suggest that the proposed monotonicity principle regarding pyramidal width is invalid, which may influence future research on geometric properties of polytopes and their applications in optimization and convex analysis.
Recognition and Label-Free Adaptation Across Recording Sessions in Surface-EMG Gesture Decoding
Time Series
- Introduces a montage-agnostic encoder for sEMG gesture decoding that adapts across recording sessions.
- Demonstrates superior performance of the encoder compared to traditional per-user LDA classifiers.
- Identifies feature-statistic alignment as the only effective label-free adaptation method across subjects.
- Addresses the practical challenges of maintaining recognition accuracy in real-world applications of myoelectric control systems.
Read more
Recognition and Label-Free Adaptation Across Recording Sessions in Surface-EMG Gesture Decoding
Summary
This paper addresses the challenge of recognition accuracy in surface electromyography (sEMG) gesture decoding, which is affected by variability across different recording sessions. The authors propose a montage-agnostic encoder that is trained on data from a specific recording session and then applied to data from a different session without requiring recalibration. This approach aims to overcome the limitations of traditional per-user classifiers, which often fail to maintain accuracy due to factors such as electrode movement and changes in skin condition. The study evaluates the encoder's performance against a per-user linear discriminant analysis (LDA) pipeline and two existing methods that rely solely on data from the same session. Additionally, the paper explores label-free adaptation techniques to improve performance when transitioning to a new recording session. The results indicate that the encoder retains a macro-F1 score of 0.688, outperforming the per-user pipeline's score of 0.540. Among various adaptation methods tested, only feature-statistic alignment consistently improved performance across all subjects, while batch-normalization re-estimation proved ineffective. This work highlights the potential for robust sEMG gesture recognition systems that can adapt to inter-session variability without extensive recalibration.
Methodology
The study employs a montage-agnostic encoder trained on data from a specific recording session, which is then tested on data from a different session. The performance is compared to a per-user LDA classification pipeline and two published methods that use data from the same session. Additionally, various label-free adaptation techniques are evaluated to enhance performance in the new session.
Results
The montage-agnostic encoder achieved a macro-F1 score of 0.688, significantly higher than the 0.540 score of the per-user LDA pipeline. Among the adaptation methods tested, only feature-statistic alignment improved performance for all subjects, while batch-normalization re-estimation failed to maintain the architecture's effectiveness.
Implications
This research has significant implications for the development of myoelectric control systems, suggesting that robust gesture recognition can be achieved without the need for extensive recalibration. The findings may lead to more practical applications of sEMG technology in assistive devices, enhancing user experience and accessibility.
Transcript-Managed Transformers: Monotone Multi-Agent Collapse and Universality with Two Pop-Enabled Transcripts
Theory
- Introduces Transcript-Managed Transducers (TMTk) for managing transcripts in Transformers.
- Establishes the equivalence between pop-enabled channels and classical computational models.
- Demonstrates that monotone protocols among agents can only achieve finite-state transductions.
- Highlights the limitations of fixed-population agents in parsing complex programming languages.
Read more
Transcript-Managed Transformers: Monotone Multi-Agent Collapse and Universality with Two Pop-Enabled Transcripts
Summary
This paper investigates the management of transcripts in fixed, finite-precision causal Transformers, focusing on the operations that can be performed on these transcripts to enhance their functionality. The author introduces the concept of the Transcript-Managed Transducer (TMTk), which consists of a finite controller managing k channels, allowing for actions such as stay, push, and pop. The study reveals that the pop-free variant, the Restricted Transcript-Managed Transducer (RTMTk), corresponds to deterministic finite-state transductions. By enabling the pop operation, the TMTk model achieves universality, with the ability to simulate classical computational hierarchies. The paper establishes that a monotone protocol among a fixed population of agents can realize finite-state transductions, emphasizing the limitations of such systems in parsing complex programming languages. The findings also demonstrate the equivalence between orchestrated one-channel agents and a single controller with multiple pop-enabled channels, providing insights into the operational capabilities of Transformers under constrained conditions.
Methodology
The paper employs theoretical modeling to define the Transcript-Managed Transducer (TMTk) and its operations, including the pop operation. It uses formal language theory to establish relationships between TMTk, RTMTk, and classical computational hierarchies, proving theorems regarding their transduction capabilities and acceptance classes.
Results
The study proves that the TMTk model achieves universality with two pop-enabled transcripts and establishes that monotone protocols among agents can only realize finite-state transductions. It also demonstrates the equivalence of transductions between orchestrated one-channel agents and a single controller with multiple channels, confirming the limitations of fixed-population agents in parsing complex languages.
Implications
The findings suggest that while Transformers can be enhanced through transcript management, there are inherent limitations in their ability to handle complex tasks such as programming language parsing. This has implications for the design of AI systems that rely on Transformers, particularly in applications requiring sophisticated reasoning and language understanding.
LARA: Lightweight Adapters in the Residual Stream for Composable Adaptation and Alignment
NLP
Large Language Models
Efficient ML
- LARA adapts models by modifying the residual stream instead of the weights, preserving the original model's integrity.
- The method provides graded control over adaptation strength via a single scaling factor at inference.
- LARA allows multiple behaviors to be stored and routed efficiently, minimizing memory usage compared to traditional methods.
- Performance is comparable to LoRA on specific tasks while maintaining a low parameter count.
Read more
LARA: Lightweight Adapters in the Residual Stream for Composable Adaptation and Alignment
Summary
LARA (Lightweight Additive Residual Adaptation) introduces a novel approach for efficient adaptation of frozen models by operating within the residual stream rather than modifying the model's weights. Unlike existing methods like LoRA, which add low-rank updates to weight matrices, LARA computes low-rank corrections to the hidden states at selected layers and adds these corrections back to the residual stream. This method preserves the base weights intact, allowing for a more flexible adaptation process. LARA enables graded control over the adaptation strength through a single scaling coefficient applied at inference, facilitating smooth transitions between base and adapted behaviors. The architecture allows multiple behaviors to coexist within a single frozen model, significantly reducing memory overhead compared to traditional methods that require separate models for each behavior. The authors demonstrate that LARA achieves comparable performance to LoRA on code fine-tuning tasks and preference optimization while maintaining a small parameter footprint, thus making it suitable for on-device inference and dynamic behavior routing per token.
Methodology
LARA employs a lightweight module that reads the hidden state from selected layers of a transformer model and adds a low-rank correction to the residual stream. This approach keeps the base weights unchanged, allowing for separate training of behaviors that can be selected dynamically during inference. The method is evaluated against LoRA using equal parameter budgets to ensure a fair comparison.
Results
LARA matches the performance of LoRA on code fine-tuning and preference optimization tasks while using a similar number of parameters. The architecture allows for the storage of multiple behaviors in a single model, resulting in approximately 33 MB of overhead compared to the significantly larger memory requirements of traditional methods that necessitate separate models for each behavior.
Implications
The LARA framework has significant implications for deploying large language models in resource-constrained environments, enabling efficient adaptation and dynamic behavior selection without the need for extensive retraining or memory overhead. This could enhance the usability of AI models in various applications, particularly in mobile and edge computing scenarios.
Topology-Aware Data Movement for Disaggregated GPU Inference
Large Language Models
Efficient ML
Optimization
- Disaggregated LLM inference creates significant data movement challenges due to the need for KV cache transfers between GPU pools.
- Existing systems do not utilize the varying bandwidth of different interconnects effectively, leading to inefficient data transfers.
- TopKV orchestrates data movement by selecting the highest-bandwidth transport method based on the physical interconnect topology.
- The proposed mechanisms include pipelined transfers, NVLink-aware routing for expert models, and the use of CXL 3.0 memory expanders.
Read more
Topology-Aware Data Movement for Disaggregated GPU Inference
Summary
This paper addresses the challenges of data movement in disaggregated GPU inference for large language models (LLMs), particularly focusing on the transfer of key-value (KV) caches between separate GPU pools during the prefill and decode phases. The author identifies that existing systems fail to optimize data transfer based on the physical interconnect topology, which can vary bandwidth by up to 72 times depending on the GPU's physical relationship. The proposed solution, TopKV, is a topology-aware transfer orchestrator that selects the optimal transport method for KV cache transfers by discovering the interconnect hierarchy at startup. TopKV employs three mechanisms: a pipelined layer-by-layer transfer to overlap transmission with computation, NVLink domain-aware routing for Mixture-of-Experts models to optimize expert dispatch and KV cache locality, and integration of CXL 3.0 memory expanders to provide additional capacity at lower latency. The paper presents analytical bandwidth models and projected performance analyses, demonstrating that TopKV can achieve a 3 to 18 times reduction in transfer latency compared to uniform RDMA protocols.
Methodology
The methodology involves designing a topology-aware transfer orchestrator that discovers the GPU interconnect graph at startup and implements five transport modes for KV cache transfers. It utilizes pipelined layer-by-layer transfers, NVLink domain-aware routing for expert models, and integrates CXL 3.0 memory expanders to optimize data movement.
Results
The evaluation of TopKV shows a 3 to 18 times reduction in transfer latency compared to existing uniform RDMA protocols, significantly improving the efficiency of KV cache transfers in disaggregated GPU inference scenarios.
Implications
The findings suggest that optimizing data movement in disaggregated GPU systems can lead to enhanced performance in LLM inference, making it feasible to handle larger models and higher request rates in production environments. This could have broad implications for cloud-based AI services and applications requiring real-time inference.
Convergence and Regret of the Policy Gradient for Multi-Armed Bandits in Diffusion Environment
Reinforcement Learning
Theory
Optimization
- The policy gradient method can achieve almost sure convergence to the optimal arm in MAB problems under a constant learning rate.
- A non-asymptotic regret upper bound of O(log T) is established for learning rates below a certain threshold.
- A novel Lyapunov function is constructed to facilitate the analysis of policy gradient updates in diffusion environments.
- The results unify previous studies and provide a straightforward proof applicable to both two-arm and multi-arm cases.
Read more
Convergence and Regret of the Policy Gradient for Multi-Armed Bandits in Diffusion Environment
Summary
This paper investigates the policy gradient method for multi-armed bandit (MAB) problems in a diffusion environment characterized by a stochastic differential equation (SDE). The authors demonstrate that the policy gradient update converges almost surely to the optimal arm when using a logit parameterization for the stochastic policy, regardless of the constant learning rate applied. They derive a non-asymptotic upper bound on regret, showing that it scales as O(log T) when the learning rate is below a specific threshold. This work improves upon previous analyses by constructing a novel Lyapunov function that simplifies the examination of policy gradient behavior in both continuous and discrete time settings. The results unify previous findings and relax conditions on learning rates, providing a clearer understanding of the relationship between learning rates and regret in MAB scenarios.
Methodology
The authors utilize a continuous-time reinforcement learning framework and apply a logit parameterization for the stochastic policy. They analyze the policy gradient update through the lens of stochastic differential equations (SDEs) and employ Itô's calculus to construct a Lyapunov function that aids in proving convergence and deriving regret bounds.
Results
The paper establishes that the policy gradient update converges almost surely to the best arm under any constant learning rate. It also provides a non-asymptotic upper bound on regret of O(log T) when the learning rate is below a threshold determined by the mean and volatility of the arms' rewards. This work improves upon prior analyses by relaxing conditions on the learning rate and offering a unified approach to both continuous and discrete time environments.
Implications
The findings have significant implications for the design of reinforcement learning algorithms in multi-armed bandit settings, particularly in environments with high noise levels. The results can inform the development of more robust learning strategies that effectively balance exploration and exploitation in uncertain conditions.
MMFGU: Multimodal Federated Graph Unlearning
Graph Learning
Federated Learning
Multimodal
- Introduces a framework for multimodal federated graph unlearning (MMFGU) to address nuanced user requests.
- Defines three types of unlearning requests: Entity/Relation Removal, Modality Removal, and Pairing Removal.
- Employs target-specific representation decoupling to manage complex multimodal data.
- Achieves a 41.5× speedup over traditional full retraining methods while maintaining graph utility.
Read more
MMFGU: Multimodal Federated Graph Unlearning
Summary
The paper presents MMFGU, a novel framework for Multimodal Federated Graph Unlearning (MM-FGU), addressing the challenges of unlearning in multimodal federated graph learning environments. Traditional federated graph unlearning methods primarily focus on removing entire clients or specific entities, which is insufficient for the nuanced demands of multimodal data where users may wish to delete specific images, texts, or associations while retaining other information. MMFGU introduces a structured approach to handle three types of unlearning requests: Entity/Relation Removal, Modality Removal, and Pairing Removal. The framework employs target-specific representation decoupling to manage the complexities of multimodal data, ensuring that only the requested information is removed without compromising the integrity of retained data. The methodology includes mapping heterogeneous requests into unified representations, decoupling requested data while anchoring retained semantics, and utilizing lightweight probes to repair any residual traces. The framework also implements a selective client purge mechanism to prevent unwanted data from re-entering the global model. Experimental results demonstrate that MMFGU effectively removes targeted information while preserving the utility of the graph, achieving a significant speedup of 41.5 times compared to full retraining.
Methodology
MMFGU utilizes a framework that maps heterogeneous unlearning requests into unified target representations. It decouples the requested information from retained data, employs lightweight probes to repair residual traces, and selectively purges affected clients based on prototype-guided screening. This approach allows for efficient unlearning without sharing raw client data.
Results
The experimental evaluation of MMFGU across 12 datasets and 2 downstream tasks indicates that it outperforms 12 existing unlearning baselines, achieving a strong balance between utility retention and unlearning effectiveness. The framework's efficiency is highlighted by a 41.5× speedup compared to full retraining methods.
Implications
The findings suggest that MMFGU can be effectively applied in scenarios where privacy and data sensitivity are paramount, such as social networks, healthcare, and any domain involving multimodal data. This framework enables organizations to comply with data removal requests while maintaining the performance of their machine learning models.
Neural Network-Assisted CLEAN for Channel Modeling in Low-SNR Regimes
Efficient ML
- NN-CLEAN combines the speed of deep learning with the robustness of MLE, improving parameter estimation in low-SNR environments.
- The framework significantly reduces computational complexity compared to traditional grid-search methods.
- NN-CLEAN generalizes well to Out-of-Distribution scenarios, addressing limitations of standalone deep learning models.
- It maintains efficient execution and memory usage, making it suitable for real-time applications in wireless communication.
Read more
Neural Network-Assisted CLEAN for Channel Modeling in Low-SNR Regimes
Summary
This paper introduces Neural Network-Assisted CLEAN (NN-CLEAN), a hybrid framework designed to enhance multipath parameter estimation in low Signal-to-Noise Ratio (SNR) environments, which are critical for modern wireless communication systems. Traditional Maximum Likelihood Estimation (MLE) methods, such as CLEAN, provide high-resolution parameter extraction but are computationally intensive due to exhaustive grid searches. On the other hand, purely data-driven deep learning approaches lack physical grounding and struggle with generalization in variable multipath densities and off-grid parameters. NN-CLEAN addresses these challenges by integrating a multi-head residual network into the iterative CLEAN extraction loop, replacing the exhaustive grid search with rapid, parallelizable forward passes. This allows for the isolation of physical multipath parameters while minimizing non-physical errors. Extensive Monte Carlo simulations demonstrate that NN-CLEAN achieves over 96% estimation accuracy at 5 dB SNR, matching the performance of traditional Grid-Search CLEAN (GS-CLEAN) while significantly reducing computational complexity. Additionally, NN-CLEAN exhibits near-flat scaling in execution runtime and memory consumption with increasing batch sizes, making it a robust, real-time solution for channel estimation in MIMO systems.
Methodology
The methodology involves embedding a multi-head residual network into the iterative CLEAN extraction loop, allowing for rapid isolation of spatial parameters through parallelizable forward passes. The residual subtraction is handled by exact mathematical models, ensuring physical accuracy and preventing non-physical errors.
Results
NN-CLEAN achieves over 96% estimation accuracy at 5 dB SNR, matching the performance of GS-CLEAN while providing a substantial reduction in computational complexity. It outperforms subspace methods and standalone neural networks, demonstrating efficient scaling in execution time and memory usage.
Implications
The proposed NN-CLEAN framework has significant implications for real-time channel estimation in advanced wireless communication systems, particularly in challenging low-SNR environments. Its efficiency and robustness make it suitable for applications such as Integrated Sensing and Communication (ISAC) and centralized baseband processing.
Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning
Computer Vision
- FATIGUECV predicts steel fatigue life from micrographs without physical testing.
- The system includes a seven-stage preprocessing pipeline and a physics-informed feature extractor.
- ResNet-50 architecture achieved high accuracy with R2 = 0.93 and macro-F1 = 0.91.
- The use of GNLL loss improved uncertainty quantification and model calibration.
Read more
Predicting Steel Fatigue Life from Micrographs Using Physics-Informed Deep Learning
Summary
This paper introduces FATIGUECV, a novel computer-vision framework designed to estimate the fatigue life of lightweight alloy steels directly from optical micrographs, bypassing the need for extensive mechanical testing. Traditional methods for evaluating fatigue life are time-consuming and destructive, often requiring hundreds of hours of testing. FATIGUECV leverages a seven-stage OpenCV preprocessing routine to clean and enhance micrographs, followed by a 28-dimensional physics-informed feature extractor that quantifies critical microstructural characteristics such as crack morphology, grain structure, porosity, and texture. The core of the system is a convolutional neural network (CNN) regression model trained using a Gaussian negative log-likelihood (GNLL) loss function, which allows for the prediction of log10(Nf) (fatigue life) alongside sample-specific uncertainty. The framework was evaluated using synthetic micrograph data, achieving impressive performance metrics, including an R2 score of 0.93 and a macro-F1 score of 0.91 with the ResNet-50 architecture. Additionally, the GNLL objective significantly improved the model's calibration, reducing the Expected Calibration Error by 76% compared to a mean-squared-error baseline. The results indicate that the model effectively identifies metallurgically relevant features, as confirmed by Grad-CAM visualizations. The entire pipeline operates efficiently, processing images in under 65 milliseconds, and is made available as open-source software, along with a synthetic dataset generator. The authors emphasize the need for further validation on real-world samples to establish the method's applicability beyond synthetic conditions.
Methodology
The FATIGUECV pipeline consists of four layers: preprocessing of raw micrographs, extraction of a physics-informed feature vector, CNN regression for fatigue life estimation, and risk classification. A synthetic dataset generator was developed to create labeled micrographs based on physics-constrained simulations of microstructural features.
Results
The ResNet-50 architecture achieved an R2 score of 0.93, RMSE of 0.18 log-cycles, and a macro-F1 score of 0.91 on synthetic micrograph benchmarks. The GNLL loss function reduced Expected Calibration Error from 0.089 to 0.021, indicating improved model calibration and uncertainty quantification.
Implications
This work has significant implications for rapid quality control in industries reliant on lightweight alloy steels, such as automotive and aerospace. By enabling quick and accurate predictions of fatigue life from micrographs, it can enhance safety and lifecycle-cost management while reducing the need for destructive testing.
Learning Lookahead Lemmas for Neural Network Verification
Theory
Optimization
Efficient ML
- Introduction of an inprocessing framework for neural network verification using lookahead procedures.
- Derivation of new lemmas about unstable ReLU phases to enhance the verification process.
- Implementation of the framework in two state-of-the-art verifiers, Marabou and α-β-CROWN.
- Demonstrated performance improvements, proving up to 34% more instances unsatisfiable.
Read more
Learning Lookahead Lemmas for Neural Network Verification
Summary
This paper presents an innovative inprocessing framework for neural network verification that leverages a lookahead procedure to derive new lemmas regarding the phases of unstable ReLU activations. The authors argue that existing state-of-the-art neural network verifiers primarily rely on a branch-and-bound (BaB) approach, which can be enhanced by proactively deriving information during the solving process. The proposed framework collects these lemmas into an implication graph, which is utilized to prune the search space and strengthen boolean cuts. The framework is implemented in two prominent verifiers, Marabou and α-β-CROWN, demonstrating significant performance improvements, including the ability to prove up to 34% more instances as unsatisfiable. The paper highlights the importance of integrating inprocessing techniques into neural network verification, drawing parallels with modern SAT solvers and showcasing the potential for enhanced efficiency in verifying deep neural networks.
Methodology
The authors developed a variation of the lookahead procedure to derive lemmas about the phases of unstable ReLUs. These lemmas are organized into an implication graph, which is used to prune the search space and enhance boolean cuts through techniques such as SAT closure, reprobing, and cut vivification.
Results
The implementation of the inprocessing framework in Marabou and α-β-CROWN resulted in improved performance, with the ability to prove up to 34% more instances as unsatisfiable compared to previous versions of these verifiers.
Implications
The findings suggest that integrating inprocessing techniques into neural network verification can significantly enhance the efficiency and effectiveness of verifying deep neural networks, which is crucial for their deployment in safety-critical applications.
Kohn-Sham Spectral Embedding on Sparse Graphs at the Nishimori Temperature for Image Classification
Computer Vision
Graph Learning
Theory
- Introduction of Kohn–Sham Spectral Embedding (KSSE) for image classification.
- Utilization of sparse-graph spectral embeddings evaluated at the Nishimori temperature.
- Establishment of six theoretical results linking statistical physics and machine learning.
- Achieved 88.93% Top-1 accuracy on ImageNet-1000 with a significantly smaller model size.
Read more
Kohn-Sham Spectral Embedding on Sparse Graphs at the Nishimori Temperature for Image Classification
Summary
This paper introduces Kohn–Sham Spectral Embedding (KSSE), a novel approach for image classification that leverages concepts from statistical physics, specifically the Kohn–Sham formalism and Random-Bond Ising Model (RBIM). The authors replace the dense top-layer classifier in convolutional neural networks (CNNs) with a sparse-graph spectral embedding evaluated at the Nishimori temperature. By mapping pre-trained feature representations onto quasi-cyclic low-density parity-check graphs, they construct a regularized Laplacian that serves as an effective Kohn–Sham Hamiltonian. This leads to independent spectral problems for each feature channel, solvable efficiently using Fast Fourier Transform techniques. The methodology includes optimizing graph topology through star-domain surgery, which enhances local convexity around codewords while managing residual frustration. The paper establishes several theoretical results, including a generalized Ihara–Bass identity and a trapping-set eigenvalue correspondence theorem. The KSSE framework is evaluated on the ImageNet-1000 dataset, achieving an impressive 88.93% Top-1 accuracy with approximately 21.24 million parameters, outperforming existing models like Swin-L and matching the performance of ViT-H/14 while significantly reducing model size.
Methodology
The authors developed KSSE by mapping CNN feature representations onto quasi-cyclic low-density parity-check graphs. They constructed a regularized Laplacian to act as an effective Kohn–Sham Hamiltonian, allowing for independent spectral problems for each feature channel. The optimization of graph topology was achieved through star-domain surgery, which improved local convexity around codewords while controlling residual frustration. The authors employed Fast Fourier Transform for efficient eigenvalue computation.
Results
KSSE achieved 88.93% Top-1 accuracy on the ImageNet-1000 dataset using approximately 21.24 million parameters. This performance surpassed that of Swin-L (197 million parameters, 86.4–87.3% accuracy) and matched the lower end of ViT-H/14 (632 million parameters, 88.0–89.5% accuracy), demonstrating a significant reduction in model size while maintaining competitive accuracy.
Implications
The KSSE framework presents a promising alternative to traditional CNN architectures, particularly in scenarios requiring efficient model sizes without sacrificing accuracy. Its foundation in statistical physics may inspire further interdisciplinary approaches in machine learning, potentially leading to advancements in other domains such as optimization and graph learning.
Class-Aware Reinforcement Learning for Counterfactual Explanation Generation
Reinforcement Learning
Interpretability
- Incorporating predicted class information in RL state representation enhances CFE generation.
- Class-aware RL outperforms class-blind RL in terms of convergence speed and reward optimization.
- The method generates significantly more valid counterfactual explanations across diverse datasets.
- Class-based features are influential in the action-selection process of the RL model.
Read more
Class-Aware Reinforcement Learning for Counterfactual Explanation Generation
Summary
This paper presents a novel approach to generating counterfactual explanations (CFEs) using class-aware reinforcement learning (RL). CFEs are crucial for enhancing the interpretability of black-box machine learning models by providing alternative instances with adjusted feature values that yield different outcomes. The authors argue that incorporating an instance's predicted class into the RL state representation can improve exploration efficiency and policy optimality. They compare their class-aware RL method against a class-blind RL method across seven diverse datasets. The results indicate that the class-aware approach leads to faster convergence, better reward optimization, and a reduction in episode length during training. Additionally, it generates a significantly higher number of valid CFEs. The study highlights the importance of class-awareness in RL for CFE generation, as the class-based feature consistently ranks as a key predictor in action selection, enhancing clarity, learning speed, and overall effectiveness in generating counterfactuals.
Methodology
The authors frame the CFE generation task as a Markov Decision Process (MDP), where they enhance the state representation by including the predicted class of the instance along with features from the supervised dataset. They designed an environment for each dataset to facilitate interaction between the RL agent and the environment, treating the CFE generation as an episodic task.
Results
The class-aware RL method demonstrated improved performance over the class-blind RL method, achieving faster convergence, better reward optimization, and a reduction in episode length. It also produced a significantly higher number of valid CFEs, indicating the effectiveness of including class information in the RL state representation.
Implications
This research has significant implications for the field of Explainable AI, particularly in applications requiring interpretability of machine learning models, such as healthcare and finance. The findings suggest that incorporating class information can enhance the generation of actionable insights from black-box models, thereby improving user trust and decision-making.
SEDR-Seq2P: A Lightweight Dilated Residual Sequence-to-Point Network for Multi-Task Industrial NILM
Time Series
Efficient ML
Optimization
- Introduces a multi-task industrial NILM framework that disaggregates multiple machine loads from aggregate power data.
- Proposes SEDR-Seq2P, a lightweight Seq2Point model with enhanced temporal modeling capabilities.
- Establishes a unified benchmarking protocol for evaluating various NILM models on the IMDELD dataset.
- Implements an Accuracy-Delay criterion to balance predictive performance and computational efficiency.
Read more
SEDR-Seq2P: A Lightweight Dilated Residual Sequence-to-Point Network for Multi-Task Industrial NILM
Summary
This paper addresses the challenges of Non-Intrusive Load Monitoring (NILM) in industrial settings, where measurement noise and concurrent machine operations complicate load disaggregation. The authors propose a novel architecture, SEDR-Seq2P, which extends the Sequence-to-Point (Seq2Point) model by incorporating dilated residual blocks and squeeze-and-excitation attention mechanisms. This approach allows for multi-task learning, enabling a single model to estimate multiple industrial machine loads from aggregate power data. The authors benchmark various models, including Seq2Seq, Seq2SubSeq, GRU, and WaveNet, using a unified evaluation protocol on the IMDELD dataset. They introduce an Accuracy-Delay (AccD) criterion to assess the trade-off between predictive accuracy and computational efficiency. The results indicate that SEDR-Seq2P achieves a 7% reduction in Mean Absolute Error (MAE) compared to the Seq2Point baseline, along with improvements in the coefficient of determination and match rate. Notably, it also reduces inference latency by approximately 58% compared to WaveNet, making it a suitable candidate for scalable industrial applications.
Methodology
The authors utilize a one-to-many multi-task learning approach, where a single model is trained to disaggregate multiple industrial loads from aggregate power measurements. They benchmark several existing models under identical conditions and introduce the SEDR-Seq2P architecture, which integrates dilated residual blocks and squeeze-and-excitation attention. The performance is evaluated using energy-estimation metrics and the newly defined Accuracy-Delay criterion.
Results
SEDR-Seq2P reduces the Mean Absolute Error (MAE) by approximately 7% compared to the Seq2Point baseline, improves the coefficient of determination by about 1%, and increases the match rate by approximately 0.8%. Additionally, it achieves a 58% reduction in inference latency compared to WaveNet, indicating a favorable accuracy-delay trade-off.
Implications
The proposed SEDR-Seq2P model offers a promising solution for industrial NILM applications, enabling efficient energy management and predictive maintenance in industrial settings. Its lightweight architecture and improved accuracy-delay balance make it suitable for real-time deployment in environments with multiple concurrent machine operations.
Certifying when decision-time information justifies adaptive experimentation
Theory
Optimization
Robotics
- Opal framework determines the justification for enabling adaptive experimentation.
- Establishes an impossibility boundary for non-trivial authorization under certain conditions.
- Achieved a low false-activation rate while recovering a significant number of positive opportunities.
- Distinguishes between policy misalignment and non-certifiability in adaptive experimentation.
Read more
Certifying when decision-time information justifies adaptive experimentation
Summary
This paper introduces the Opportunity-aware Policy Authorization for Laboratories (Opal), a framework designed to determine whether adaptation in adaptive experimentation should be enabled based on decision-time information. Unlike existing methods that assume adaptation is already permitted, Opal evaluates the justification for enabling adaptive branches by requiring non-trivial adaptation, controlled target risk, and positive executed value after costs. The authors establish an impossibility boundary indicating that source outcomes and unlabelled target covariates cannot uniformly support non-trivial authorization under unrestricted conditional outcome shifts. They derive a target-calibrated recovery method and apply Opal to a large-scale Cell Painting dataset, demonstrating its effectiveness in selecting compounds while maintaining a low false-activation rate. The study highlights the importance of distinguishing policy misalignment from non-certifiability and establishes authorization as a critical layer for safe adaptive scientific experimentation.
Methodology
The authors developed the Opal framework, which combines decision-time measurements, risk control, and a precommitted contract to assess the justification for enabling adaptive experimentation. They conducted empirical studies using a Cell Painting dataset to validate the framework's effectiveness in selecting compounds and controlling false activations.
Results
In the application to the Cell Painting dataset, Opal selected 595 compounds, with 384 identified as positive opportunities. The method maintained a false-activation upper bound of 5.18%, below the 7.5% threshold. The framework successfully recovered a non-trivial, positive-value active branch, demonstrating its capability to certify adaptive opportunities under asynchronous evidence and commitment.
Implications
The findings suggest that Opal can enhance the safety and efficiency of adaptive experimentation in scientific laboratories, potentially leading to more informed decision-making and better resource allocation in experimental design. This framework could be applied in various fields where adaptive experimentation is critical, such as pharmacogenomics and other areas of scientific research.
SILVA Networks as Structured Implicit Layers and Vector Attractors via Dynamic Interaction Fields
Graph Learning
Theory
Interpretability
- SILVA Networks separate stimulus, local interaction, global interaction, and solver dynamics in a fixed-point architecture.
- The architecture is adaptable to various domains, including images, molecules, and citation networks.
- Local interactions are essential for graph tasks, while global interactions show significant benefits in long-range classification tasks.
- SILVA Networks allow for the training and visualization of internal dynamics, enhancing interpretability.
Read more
SILVA Networks as Structured Implicit Layers and Vector Attractors via Dynamic Interaction Fields
Summary
The paper introduces SILVA Networks, a novel architecture designed to enhance the representation of learning problems by separating the influences of direct input, local interactions, global context, and solver dynamics within a fixed-point framework. SILVA Networks achieve this by defining structured implicit layers that allow for a clear decomposition of the computational roles involved in processing stimuli. The architecture is versatile, applicable across various domains such as images, molecules, citation networks, and long-range graph benchmarks. The methodology involves a damped update mechanism that integrates stimulus, local, and global interactions, enabling the model to reach a stable vector attractor state. Experimental results demonstrate that local interactions are crucial for graph tasks, while global benefits are more pronounced in long-range node classification tasks. The findings suggest that SILVA Networks provide a robust implicit representation that can be trained, visualized, and diagnosed, offering insights into the internal dynamics of neural architectures.
Methodology
The methodology involves the design of SILVA Networks, which utilize structured implicit layers to decompose the update process into distinct computational roles. A damped update formula is employed to integrate the effects of stimulus, local interactions, and global context, allowing the model to converge to a stable state. The architecture is tested across various benchmarks to evaluate the impact of local and global interactions.
Results
Experiments reveal that local interactions are critical for performance in graph-related tasks, while the contribution of global interactions is more evident in long-range node classification benchmarks. The results indicate that SILVA Networks can effectively leverage both local and global influences to improve representation learning.
Implications
The implications of this work extend to various fields where complex interactions within data are present, such as molecular modeling, social network analysis, and image processing. The ability to visualize and diagnose internal dynamics may lead to more interpretable machine learning models and better understanding of how different factors influence learning outcomes.
Guarantees on Dynamical System Distinguishability for LLM Token Generation
Large Language Models
Theory
- Two stable dynamical systems can produce nearly indistinguishable stationary distributions, highlighting limitations of marginal density analysis.
- Misclassification error in DS-based classification decreases exponentially with sequence length, influenced by the dynamical distinctiveness of the systems.
- Dynamical systems fitted on one embedding can be effectively transferred to another, provided the embeddings share similar geometric structures.
Read more
Guarantees on Dynamical System Distinguishability for LLM Token Generation
Summary
This paper investigates the theoretical underpinnings of classifying responses from large language models (LLMs) by modeling token embeddings as trajectories of black-box dynamical systems (DS). The authors formalize the classification task as a binary hypothesis test between two stochastic linear DSs and demonstrate that the total variation distance between their stationary distributions can be small, even with significant differences in dynamics. They establish that the misclassification probability decreases exponentially with the sequence length, governed by a dynamical discriminability quantity. Additionally, they introduce an intertwining condition for cross-embedding generalization, providing a lower bound on transferable discriminability. These findings elucidate the empirical success of DS-based classification and suggest a shift towards using DS theory for analyzing AI systems.
Methodology
The authors formalize the classification of LLM responses as a binary hypothesis test between two stochastic linear dynamical systems. They analyze the total variation distance of stationary distributions, derive the misclassification probability as a function of sequence length, and establish conditions for cross-embedding generalization.
Results
The study reveals that classifiers relying solely on marginal statistics face a fundamental accuracy limit. In contrast, classifiers based on prediction residuals can achieve exponentially decreasing misclassification error with longer sequences. Furthermore, the ability to transfer dynamical systems across different embeddings is contingent on the geometric similarity of the embedding spaces.
Implications
The results suggest that leveraging dynamical systems theory could enhance the classification of LLM responses, potentially reducing the need for multiple responses and improving efficiency in applications such as hallucination detection and change-point detection in AI systems.
ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents
Large Language Models
NLP
Robotics
- ClawTrack introduces a dual-assessment framework measuring both task outcomes and reasoning processes.
- The benchmark includes 320 tasks across 8 domains, providing a comprehensive evaluation of autonomous agents.
- Process scores effectively identify specific reasoning deficiencies and are robust across different evaluators.
- Result verification is highlighted as a critical bottleneck in the reasoning process.
Read more
ClawTrack: Towards Trace-Level Evaluation and Improvement of Real-World Autonomous Agents
Summary
The paper introduces ClawTrack, a dual-assessment benchmark designed to evaluate and improve the performance of autonomous agents based on both task outcomes and the processes they employ to achieve these outcomes. Traditional benchmarks primarily focus on final results, which can obscure the reasoning quality behind successes and failures. ClawTrack addresses this gap by providing a comprehensive evaluation framework that includes a Task Score for what an agent accomplishes and a Process Score for how it achieves these results. The benchmark encompasses 320 tasks across 8 domains, utilizing over 25 deterministic mock services. A Process Grader evaluates each reasoning step based on four criteria: goal alignment, efficiency, information utilization, and result verification, supported by a detailed rubric of 12,541 items. The authors evaluated 21 models across more than 16,000 trials, revealing that process scores effectively identify specific reasoning deficiencies, highlight complementary dimensions of performance, and demonstrate robustness across different evaluators. The findings indicate that focusing on process-based evaluations can lead to significant improvements in model performance, particularly in areas identified as bottlenecks, such as result verification.
Methodology
The authors developed ClawTrack as a benchmark that evaluates autonomous agents through a dual-assessment approach. It includes a Process Grader that scores each reasoning turn based on four dimensions and an Outcome Grader that assesses final results. The evaluation involved 21 models tested over 16,000 trials, allowing for detailed analysis of both task completion and reasoning quality.
Results
The evaluation revealed that process scores can effectively attribute successes and failures to specific reasoning dimensions, filtering out instances of lucky success. The four evaluation dimensions were found to be complementary, with result verification identified as a systematic bottleneck. The framework demonstrated robustness to evaluator choice and facilitated consistent improvements in model performance through process-based trajectory filtering.
Implications
ClawTrack's structured process-level evaluation can enhance the reliability and safety of autonomous agents in real-world applications. By identifying specific reasoning deficiencies, developers can target interventions to improve agent performance, thereby increasing user trust and compliance in high-stakes environments.
Search Strategies for Optimal Classification and Regression Trees
Optimization
Interpretability
Efficient ML
- Introduces a unified algorithmic framework for Optimal Decision Trees (ODTs) to compare various search strategies.
- Empirically evaluates 18 search strategies, marking the largest evaluation in this domain.
- Demonstrates significant improvements in anytime performance for classification and runtime for regression.
- Addresses the challenges of scalability in learning ODTs, which are NP-hard to optimize.
Read more
Search Strategies for Optimal Classification and Regression Trees
Summary
This paper addresses the challenges of scalability in learning Optimal Decision Trees (ODTs), which are known for their compactness and interpretability. The authors propose a general algorithmic framework that unifies various search strategies for ODTs, allowing for a systematic comparison and evaluation of 18 different strategies. The framework facilitates the understanding of the strengths and weaknesses of each method, which have previously been studied in isolation. The empirical evaluation demonstrates that the best-performing strategy significantly enhances anytime performance for classification tasks and achieves a runtime improvement of over an order of magnitude for regression tasks compared to existing state-of-the-art methods. This work not only clarifies the landscape of ODT search strategies but also provides a foundation for future research in optimizing decision tree learning.
Methodology
The authors developed a general algorithmic framework for ODTs that incorporates and generalizes existing search strategies. They empirically tested 18 search strategies, including depth-first search, best-first search, and AND-OR search, to evaluate their performance in terms of scalability and efficiency for both classification and regression tasks.
Results
The evaluation revealed that the best search strategy outperformed state-of-the-art methods in terms of anytime performance for classification tasks and improved runtime for regression tasks by more than an order of magnitude. This highlights the effectiveness of the proposed framework in optimizing decision tree learning.
Implications
The findings suggest that the unified framework can guide future research in decision tree optimization, potentially leading to more efficient algorithms that can handle larger datasets and deeper trees. The insights gained from comparing various strategies may also inform practitioners in selecting appropriate methods for specific applications in high-stakes domains like healthcare.
Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels
Graph Learning
Time Series
- STGATv2 outperforms traditional ML and other neural network models in fault location tasks.
- Generalization performance of models varies significantly with DER penetration levels.
- Topological awareness is crucial for robust fault location in active distribution networks.
- STGATv2 maintains high performance under realistic measurement noise conditions.
Read more
Assessing the Generalization of Graph Neural Networks for Fault Location Across Increasing Distributed Energy Resource Penetration Levels
Summary
This paper addresses the challenge of accurately locating faults in power distribution networks, which is becoming increasingly complex due to the integration of distributed energy resources (DERs). The authors investigate the performance of Spatio-Temporal Graph Neural Networks (STGNNs), specifically the spatio-temporal graph attention network (STGATv2), in comparison to traditional machine learning methods and other neural network architectures under varying levels of DER penetration (10%, 25%, 50%). The study reveals that STGATv2 consistently outperforms its counterparts, achieving a macro F1 score of 92-94% in in-distribution scenarios. However, the generalization capability of the models varies significantly with DER penetration levels; training at higher penetration levels (50%) allows the model to maintain performance at lower levels, while training at lower levels (10%) results in a significant drop in performance at higher levels. The paper emphasizes the importance of topological awareness in maintaining robust fault location capabilities, particularly in active distribution networks with realistic measurement noise, where STGATv2 maintains over 85% F1 score compared to much lower scores from GRU and GATv2.
Methodology
The authors benchmarked STGATv2 against purely temporal (GRU), purely spatial (GATv2), and traditional machine learning methods. They evaluated the models' performance across different DER penetration levels on a reconfigured IEEE 123-bus feeder, focusing on their ability to generalize under varying conditions.
Results
STGATv2 achieved a macro F1 score of 92-94% in in-distribution scenarios. It retained 81-84% F1 score when trained at 50% DER penetration and tested at lower levels, while training at 10% resulted in a significant drop to 69-74% F1 at 50%. Under realistic noise conditions, STGATv2 maintained over 85% F1, while GRU dropped to 33.5% F1.
Implications
The findings suggest that STGNNs, particularly STGATv2, can significantly enhance fault location accuracy in modern power distribution systems with high DER penetration. This has implications for improving grid reliability and minimizing downtime in the face of increasing complexity in power networks.
Mining Verdict Boundaries for Neural Network Verification
Theory
Efficient ML
- Introduces BMiner, a novel approach for neural network verification that improves efficiency in finding verdict boundaries.
- Utilizes the monotonicity of the verification metric to formulate the boundary search as a problem in a nearly sorted array.
- Implements two search strategies (BMinerE and BMinerG) that reduce the number of node visits in the BaB tree.
- Demonstrates significant time savings in verification tasks on standard datasets.
Read more
Mining Verdict Boundaries for Neural Network Verification
Summary
This paper addresses the challenge of verifying neural networks using the Branch and Bound (BaB) method, which involves partitioning the verification problem into subproblems represented as a tree structure. The authors identify inefficiencies in existing BaB approaches, which sequentially solve each subproblem to find verdict boundaries that separate verified from unverified nodes. To improve this process, they propose a new verification approach called BMiner, which leverages the monotonicity of the verification metric along each path in the tree. BMiner employs two efficient search strategies: BMinerE, which uses exponential search to locate boundary points, and BMinerG, which estimates the boundary's position based on the gradient of the verification metric. Experimental evaluations on benchmark datasets such as MNIST and CIFAR-10 demonstrate that BMiner significantly reduces verification time compared to state-of-the-art methods, achieving an average time reduction of 17% to 30%, with a maximum reduction of 44.7%.
Methodology
The authors propose BMiner, which consists of two main search strategies: BMinerE, which uses exponential search to identify a range containing the verdict boundary and then applies binary search to pinpoint the exact location; and BMinerG, which estimates the boundary position using the gradient of the verification metric to skip unnecessary node visits.
Results
The experimental results show that BMiner reduces the average verification time by 17% to 30% across various neural networks, with a maximum reduction of 44.7%. This indicates a significant improvement over existing BaB-based verification methods.
Implications
The findings suggest that BMiner can enhance the efficiency of neural network verification, making it more feasible for deployment in safety-critical applications such as autonomous driving and other domains where robustness is essential.
A Montage-Agnostic Encoder for Calibration-Light Cross-User Gesture Recognition from Surface Electromyography
Time Series
- Introduction of a Montage-Agnostic Encoder for sEMG gesture recognition.
- Encoder outperforms per-user LDA classifiers in cross-user scenarios.
- Key components of the encoder significantly contribute to its performance.
- Performance is influenced by signal fidelity rather than training pool size.
Read more
A Montage-Agnostic Encoder for Calibration-Light Cross-User Gesture Recognition from Surface Electromyography
Summary
This paper addresses the challenge of cross-user gesture recognition from surface electromyography (sEMG) signals, which typically requires extensive calibration for each user. The authors introduce a Montage-Agnostic Encoder that utilizes shared weights across electrodes and locates them by their physical coordinates, allowing the model to adapt to varying electrode configurations without requiring montage-specific parameters. The encoder is trained across multiple users and demonstrates improved performance over traditional per-user classifiers, particularly in terms of macro-F1 scores. The study includes an ablation analysis to evaluate the contributions of the encoder's components and assesses the calibration efficiency, showing that the encoder outperforms a linear discriminant analysis (LDA) baseline under certain conditions. The findings suggest that the encoder's performance is influenced by the fidelity of the input signals and that the size of the training pool primarily serves as a stability floor rather than a performance enhancer. Additionally, the paper explores the limitations of self-supervised pretraining in this context.
Methodology
The proposed encoder architecture processes sEMG signals by treating each electrode as a token based on its physical location. It employs an attention mechanism to learn spatial patterns across multiple electrodes and normalizes each channel independently. The model is evaluated using various datasets to measure cross-user recognition performance against calibration budgets, and an ablation study is conducted to assess the impact of its components.
Results
The Montage-Agnostic Encoder achieved a macro-F1 score improvement of 0.234 over the per-user Hudgins classifier on the NinaPro DB1 dataset and 0.108 on DB2, while underperforming on DB5. The encoder's components were shown to contribute significantly to its overall performance, and the training pool size was found to primarily affect stability rather than performance enhancement.
Implications
The findings suggest that the Montage-Agnostic Encoder could facilitate the development of more efficient and user-friendly myoelectric prostheses by reducing the need for extensive calibration. This could lead to broader adoption of sEMG-based control systems in clinical and consumer applications, enhancing the usability of prosthetic devices.
A Neurosymbolic Approach for Explainable Early Diagnosis of Alzheimer's Disease
Audio & Speech
NLP
Interpretability
- Introduction of NeSyQuaKE, a neurosymbolic framework for extracting qualitative knowledge from clinical audio data.
- Utilization of pretrained foundation models for symbol grounding and Bayesian Networks for reasoning about cognitive impairment markers.
- Successful recovery of established clinical knowledge and identification of new qualitative relationships between linguistic markers and Alzheimer's Disease.
- The framework addresses the limitations of manual transcription and analysis in clinical settings, promoting scalability.
Read more
A Neurosymbolic Approach for Explainable Early Diagnosis of Alzheimer's Disease
Summary
This paper presents a novel automated pipeline for the early diagnosis of Alzheimer's Disease (AD) through a neurosymbolic approach that leverages audio recordings from verbal fluency tests. The proposed framework, named NeSyQuaKE, integrates pretrained foundation models for audio processing with Bayesian Networks (BNs) to extract qualitative knowledge about potential AD progression indicators. By addressing the challenges of symbol grounding and reasoning, NeSyQuaKE successfully identifies both established and novel relationships between linguistic markers and cognitive impairment. The system's ability to recover known clinical knowledge while generating new hypotheses demonstrates its effectiveness in enhancing the scalability and interpretability of AD diagnosis. The authors emphasize the importance of bridging the gap between unstructured audio data and structured clinical knowledge, ultimately aiming to facilitate more efficient longitudinal tracking of cognitive symptoms in patients.
Methodology
The methodology involves an end-to-end pipeline that processes raw audio data from verbal fluency tests using speech-to-text transcription, followed by the extraction of clinically relevant variables through large language models. These variables are then structured into a Bayesian Network to reason about the relationships between cognitive impairment markers. The QuaKE algorithm is employed to derive qualitative influence statements that capture trends in the data.
Results
The NeSyQuaKE framework was evaluated on a real-world dataset of clinical verbal fluency tests, demonstrating its capability to recover established clinical knowledge and generate additional hypotheses regarding the qualitative relationships between cognitive impairment and its markers. The results indicate that the system effectively bridges the gap between unstructured audio data and structured clinical insights.
Implications
The implications of this research are significant for the field of early Alzheimer's diagnosis, as it offers a scalable and automated method for tracking cognitive symptoms. By enhancing the interpretability of the diagnostic process, the framework could facilitate better patient monitoring and potentially lead to earlier interventions. Additionally, the approach may be applicable to other domains requiring the analysis of unstructured audio data.
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Generative Models
Efficient ML
Optimization
- OnlineCache adapts caching policies dynamically based on prompt complexity and error sensitivity.
- The framework employs a policy gradient method to optimize resource allocation during inference.
- A learnable corrector is integrated to address errors induced by caching decisions.
- Extensive experiments show OnlineCache outperforms static caching methods in both speed and quality.
Read more
OnlineCache: Learning Dynamic Caching Policies with Error Correction for Efficient Diffusion Inference
Summary
The paper introduces OnlineCache, a dynamic caching framework designed to enhance the efficiency of diffusion inference by learning adaptive caching policies. Traditional caching methods rely on static schedules that do not account for the varying difficulty of generation tasks and the fluctuating error sensitivity across timesteps. OnlineCache addresses these limitations by employing a policy gradient approach to dynamically allocate computational resources based on the complexity of prompts and the sensitivity of errors at different stages of the denoising process. The framework consists of a policy network that decides when to cache and a corrector that mitigates errors from caching. The authors demonstrate that OnlineCache significantly improves generation quality while achieving nearly three times the speedup compared to existing methods, making it a promising solution for real-time applications in generative AI.
Methodology
The authors propose a bilevel optimization framework where a policy network is trained to make caching decisions based on the current state of the diffusion process. The policy is optimized to maximize overall generation quality while a corrector is trained to minimize local errors caused by caching. This joint optimization allows for adaptive resource allocation across different samples and timesteps.
Results
OnlineCache achieves nearly 3× speedup on the FLUX.1-dev model while maintaining high generation fidelity. On other models like DiT and CogVideoX, it also demonstrates competitive acceleration without compromising quality. The method consistently outperforms existing cache-based acceleration baselines, showing significant reductions in error rates and improvements in perceptual quality.
Implications
The proposed OnlineCache framework has the potential to enhance real-time applications in generative AI by reducing inference latency while preserving output quality. Its dynamic approach to caching could be applied to various domains where generative models are utilized, improving efficiency in tasks such as image synthesis, video generation, and audio processing.
Recursive transformers for semiconductor thermo-mechanical reliability
Efficient ML
- Introduces recursive weight-sharing transformers to improve efficiency in small-data engineering applications.
- Evaluates three recursive transformer architectures for thermo-mechanical reliability analysis.
- Demonstrates that recursive transformers can reduce overfitting and computational overhead compared to conventional models.
- Provides practical design guidelines for selecting transformer architectures in resource-constrained scenarios.
Read more
Recursive transformers for semiconductor thermo-mechanical reliability
Summary
This paper presents a novel approach to using transformer-based surrogate models for semiconductor thermo-mechanical reliability analysis, addressing the challenges posed by conventional transformer architectures that are often over-parameterized for small, low-dimensional datasets typical in engineering design. The author introduces three recursive transformer paradigms: Tiny Recursive Model, Depth Recursive Model, and a simple recursive transformer, focusing on reducing computational complexity and parameter count while maintaining predictive performance. The study systematically evaluates these models on two engineering tasks: the thermo-mechanical reliability analysis of semiconductor packages and a Laplace PDE iterative numerical solver for capacitance fields. The findings demonstrate that recursive weight-sharing transformers can effectively balance prediction accuracy, parameter efficiency, and computational cost, making them suitable for resource-constrained environments where traditional large-parameter transformers are inadequate.
Methodology
The paper employs a hardware-aware evaluation of three recursive transformer paradigms, comparing their predictive performance, parameter count, and computational complexity. The models are validated on two engineering tasks, utilizing recursive weight-sharing to enhance efficiency without increasing parameters.
Results
The study finds that recursive weight-sharing transformers outperform conventional transformer architectures in terms of predictive accuracy and computational efficiency for small datasets. The proposed models demonstrate a significant reduction in parameter count and computational complexity while maintaining high performance in thermo-mechanical reliability predictions and capacitance field modeling.
Implications
The findings suggest that recursive transformers can be effectively deployed in hardware-constrained environments, such as on-chip inference and real-time design optimization, making them valuable for engineering applications where simulation data is costly to generate.
Flow Matching with Missing Data
Generative Models
Theory
Optimization
- Introduction of Missing-Data Flow Matching (MDFM) for handling incomplete training data.
- Proven equivalence of incomplete and complete data objectives under MCAR conditions.
- Finite-sample analysis reveals optimal strategies for completion and evaluation budget.
- MDFM maintains conditional variability and performs competitively against strong imputation methods.
Read more
Flow Matching with Missing Data
Summary
The paper introduces Missing-Data Flow Matching (MDFM), a novel approach to flow matching that addresses the challenge of incomplete training data, which is common in real-world applications. Traditional flow matching relies on fully observed data, but many datasets have missing values due to various reasons such as sensor failures or incomplete surveys. MDFM treats the missing coordinates as latent variables and utilizes resampling from the conditional distribution of these missing values to compute the flow matching loss. The authors prove that under the assumption of missing completely at random (MCAR), the objective of training with incomplete data is equivalent to that of complete data, thus demonstrating that the missingness does not alter the learning process of flow matching. The paper also presents a finite-sample analysis that provides insights into optimal design choices regarding the number of completions per example and the evaluation budget allocation. The findings reveal that one completion per example is sufficient and that the quality of the completion model is crucial for performance. Theoretical predictions are validated through experiments on synthetic and real tabular data, showing that MDFM competes well with existing imputation methods, particularly in preserving covariance.
Methodology
The authors propose MDFM, which involves treating missing data as latent variables and averaging the flow matching loss over multiple resampled completions from the conditional distribution. They conduct a theoretical analysis to establish the equivalence of objectives under MCAR and perform finite-sample variance decomposition to optimize the number of completions and evaluate the impact of the completion model's accuracy.
Results
The experiments validate the theoretical findings, demonstrating that one oracle completion matches the variance of complete data and is optimal under a fixed evaluation budget. The results show that MDFM effectively preserves conditional variability and outperforms classical and deep imputation baselines on real tabular datasets, particularly in terms of covariance preservation.
Implications
MDFM has significant implications for generative modeling in scenarios where data is often incomplete, such as medical records, sensor data, and survey responses. By effectively handling missing data, this approach can enhance the performance of generative models in various applications, leading to more accurate and reliable data generation.
Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems
Time Series
- Evaluation of GRU, LSTM, and Transformer models for classifying automated driving systems using telematics data.
- Strong clean-data performance with macro F1-scores above 0.90 for all models.
- Introduction of a modular robustness evaluation framework to assess model performance under telematics degradation.
- Significant vulnerability of all models to temporal jitter, revealing a critical limitation in current sequence models for ADS identification.
Read more
Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems
Summary
This paper investigates the effectiveness of three sequence-based classification models—Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and a Transformer encoder model—in identifying Level 2 automated driving systems (ADSs) using vehicle telematics data. The authors evaluate the models on their ability to classify different ADSs, including Comma Openpilot, Tesla Autopilot, and Cadillac Super Cruise, achieving strong performance with macro F1-scores of 0.92 (GRU), 0.90 (LSTM), and 0.93 (Transformer) on clean data. The study also introduces a modular robustness evaluation framework that simulates realistic telematics degradation through various corruption types and severity levels. Results indicate that while all models perform well on clean data, they exhibit significant vulnerability to temporal jitter, with macro F1-scores dropping to 0.44–0.50 under such conditions. This research highlights the need for robust monitoring systems in automated driving, emphasizing the importance of independent verification based on observable vehicle behavior rather than relying solely on internal software declarations.
Methodology
The authors employed three sequence-based models (GRU, LSTM, Transformer) to classify ADSs based on vehicle telematics data. They introduced a robustness evaluation framework that simulates telematics degradation through five corruption families at varying severity levels, measuring performance using macro F1 scores.
Results
All models achieved strong classification performance on clean data, with macro F1-scores of 0.92 (GRU), 0.90 (LSTM), and 0.93 (Transformer). However, under conditions of temporal jitter, the models' performance significantly dropped to 0.44–0.50, indicating a critical failure mode.
Implications
The findings underscore the importance of developing robust monitoring systems for automated driving systems, which can enhance safety, regulatory compliance, and anomaly detection. The study also sets a precedent for future research in telematics-based classification and robustness testing.