Brief

Researchers have introduced the YB Mixer, a novel sequence token mixing layer inspired by integrable systems and the generalized Yang-Baxter equation. This layer leverages free fermionic structures and an Ising exchange algebra to ensure computational stability and create an exactly norm-preserving orthogonal map. The YB Mixer's design allows for order-free inference adaptable to variable budgets and utilizes a spectral circulant generator for generalization to longer sequences, resulting in a stable and mathematically robust architecture for sequence processing. AI

A new research paper published on arXiv examines the effectiveness of selective prediction methods for risk control in AI systems. The study found that common practices like naive thresholding can lead to a false sense of security, with error rates significantly exceeding declared budgets in many trials. Certified methods like Clopper-Pearson and betting upper confidence bounds showed better performance, but still experienced overruns under grouped deployment due to broken exchangeability premises. AI

Researchers have introduced EnvShip-Bench, a new benchmark designed to standardize and advance the field of short-term vessel trajectory prediction. This benchmark addresses limitations in existing maritime AIS data by providing a unified processing pipeline, consistent forecasting protocols, and integrated contextual data such as environmental conditions and nearby vessel movements. EnvShip-Bench aims to facilitate fair comparisons and encourage the development of more sophisticated, context-aware models for applications like intelligent shipping and navigation safety. AI

Researchers have developed TriAdReview, a novel architecture for improving technical document generation by large language models. This system uses two independent reviewer models with distinct perspectives and a triangular judging mechanism to iteratively refine the output of a generator model. Evaluations across five benchmark tasks demonstrated a significant overall improvement, particularly in security audits, code generation, and architecture design, though it showed a degradation in completeness-oriented tasks like requirements analysis. AI

A new research paper published on arXiv highlights significant inconsistencies in how Jensen-Shannon divergence is estimated for synthetic tabular data. The study reveals that different estimation protocols can lead to non-comparable divergence values, with marginal-based estimators often underestimating divergence by ignoring dependencies, while classifier-based estimators capture joint structure but are sensitive to the specific estimator used. The researchers propose a posterior correction for classifier-based estimation and offer practical guidelines and an open-source tool to address these protocol dependencies for more meaningful comparisons. AI

Researchers have developed M-CTX, a new framework designed to significantly accelerate the process of retrieving spatial context for trajectory analytics. This system addresses a major bottleneck in modern trajectory predictors by recasting context construction as a spatial database workload. M-CTX achieves an end-to-end speed-up of 226x, reducing context construction time from approximately 17 CPU-days to just 1.8 hours for a large dataset. AI

Researchers have introduced Physics-conforming Latent Twins, a new framework designed to create more physically accurate surrogate models for scientific machine learning. This method ensures that the learned models not only predict accurately but also adhere to fundamental physical principles like conservation laws and invariants. By constraining the dynamics within a latent space, the framework improves the structural fidelity and long-term behavior of simulations, as demonstrated in experiments with ODE and PDE benchmarks. AI

Researchers have developed a novel approach to lattice reduction strategies by employing deep reinforcement learning, specifically an AlphaZero-style self-play pipeline with Monte Carlo Tree Search. This method trains a deep residual network to discover strategies that outperform the traditional Lenstra-Lenstra-Lovász (LLL) algorithm. The resulting policy, DeltaStar, trained on small lattices, demonstrates generalization to higher dimensions and unseen moduli without retraining. AI

A new research paper explores the effectiveness of various Graph Neural Network (GNN) layers for predicting driving trajectories. The study compares 19 different graph layer types, identifying five combinations that consistently outperform others, particularly ARMA, Chebyshev, and topology-aware layers. Key findings suggest that sum-based aggregation, multi-head attention, and weighted hop distances enhance prediction accuracy, offering practical design principles for future autonomous driving systems. AI

A new research paper explores the optimal use of Schatten-p norms in deep learning, particularly in relation to optimizers like Muon. The study demonstrates that the effectiveness of these norms is dependent on the specific regime, with smaller Schatten-p geometries proving optimal in low-dimensional settings, including those relevant to Chinchilla scaling. This analysis also provides insights into why Muon-like methods favor large batches and offers a scaling rule for batch sizes across different values of p. AI

Researchers have introduced Brownian kernel ladders (BKLs), a novel hierarchy of integral reproducing kernel Hilbert spaces designed to capture compositional representations in machine learning. This framework recursively defines layers by integrating Brownian kernels over probability measures, encoding depth directly into the hierarchy. The BKL spaces exhibit desirable analytical and statistical properties, including depth-dependent Hölder regularity and strict monotonicity, and provide a mathematically tractable foundation for studying compositional representations in deep learning. AI

A new research paper published on arXiv explores the concept of "model stealing" attacks, where adversaries create surrogate models that mimic the behavior of proprietary AI systems. The study challenges the assumption that high-fidelity surrogates are equivalent to the original models, demonstrating that multiple near-optimal surrogates can exist with significant differences in deployment-relevant properties. Experiments across various tasks, including tabular data, medical imaging, and natural language processing, reveal that these surrogate models can exhibit considerable variance in critical performance metrics despite similar fidelity to the target model. AI

Researchers have introduced LatentGym, a new testbed designed to study how AI agents learn from sequences of related tasks. This framework provides controllable, ground-truth latent structures that govern task relationships, allowing for the measurement of both exploration and exploitation of learned information. Initial studies using LatentGym explore why current frontier models struggle with cross-task adaptation and how factors like inter-task feedback influence learning dynamics. AI

Researchers have developed an adaptive graph model that enhances the k-nearest neighbors (kNN) algorithm for large-scale AI applications. This new model decouples inference latency from computational complexity by integrating a Hierarchical Navigable Small World (HNSW) graph with a pre-computed voting mechanism. The approach shifts the computational burden of neighbor selection to the training phase, enabling faster navigation through higher graph layers and precise, adaptive neighbor counts in lower layers. Benchmarks across six datasets show this architecture significantly accelerates inference speeds without sacrificing classification accuracy, offering a scalable solution to kNN's inherent inference bottleneck. AI

Researchers have developed a new method called Simulation-Augmented Multi-Step Split Conformal Prediction (SA-MSCP) to improve uncertainty quantification in aggregated forecasting tasks. This technique generates future paths using a block bootstrap from cross-validated residuals and constructs prediction intervals from empirical quantiles. Experiments indicate that SA-MSCP enhances empirical coverage compared to existing baselines, demonstrating its effectiveness for aggregated time-series forecasting. AI

Researchers have developed ProGenMech, a new framework for understanding the internal workings of autoregressive protein language models. This method extends cross-layer transcoders to models like ProGen3, enabling a more faithful recovery of generative computations across layers. A zero-shot circuit discovery framework within ProGenMech identifies specific latent circuits responsible for protein generation and fitness prediction, revealing biologically meaningful motifs and functional regions. AI

Researchers have introduced BRICKS-WM, a novel framework designed to enhance the reusability of structured world models in model-based reinforcement learning. This framework addresses the limitation of monolithic latent dynamics by proposing a modular assembly approach where global dynamics are modeled as a composition of independent dynamical modules. Specifically, BRICKS-WM factors the latent state space into an Agent module and a Background module, connected by a learned latent interface, ensuring functional separation of dynamics. AI

Researchers have introduced Trust-region Diffusion Policies (TruDi), a novel framework designed to enable the effective training of diffusion policies within massively parallel, on-policy reinforcement learning (RL) settings. This approach addresses the challenges of rapidly changing data distributions in on-policy RL by incorporating a trust-region optimization rule to maintain stability with complex policies. Empirical evaluations across four benchmarks and 73 tasks demonstrate that TruDi matches or surpasses existing baselines, showing particular strength in complex humanoid control tasks. AI

Researchers have introduced Phys-JEPA, a novel physics-informed latent world model designed for multivariate time-series forecasting. This model imposes physical consistency directly onto latent states and transitions, rather than solely on decoded outputs. Phys-JEPA aims to create statistically useful yet physically structured predictive states by decomposing them into physical and residual components. Initial experiments on datasets like Jena Climate, Traffic, and Electricity show improvements in mean squared error, particularly at longer forecasting horizons, suggesting this approach enhances interpretable temporal world models. AI

Researchers have developed a unified formulation for one-shot expert pruning in Mixture-of-Experts (MoE) language models. This new approach organizes pruning criteria around routing frequency, gate weighting, and activation strength. The formulation leads to a principle for selecting pruning criteria based on whether the task is task-agnostic or task-specific. Two new task-agnostic criteria, Mean Activation Norm (MAN) and Mean Squared Activation Norm (MSAN), were introduced and demonstrated strong performance across various MoE models and benchmarks. AI

Researchers have developed WavSLM, a novel speech language model that simplifies the generation of coherent speech by distilling self-supervised WavLM representations into a single codebook. This approach allows WavSLM to jointly model semantic and acoustic information within a single token stream, bypassing the need for text supervision or pretraining. Despite its streamlined architecture, WavSLM demonstrates competitive performance on speech generation and consistency benchmarks, utilizing fewer parameters and less training data while enabling streaming inference. AI

Researchers have developed a new geometric framework to understand the fragility of alignment in language models during fine-tuning. Their analysis reveals that even seemingly benign tasks can systematically break safety guardrails, a phenomenon they term "alignment collapse." The framework identifies specific geometric properties, formalized as the Alignment Instability Condition (AIC), that are sufficient to guarantee degradation of safety features. This work provides a theoretical basis for predicting and preventing such alignment degradation, showing that alignment can degrade rapidly even when initial updates appear safe. AI

Researchers have introduced SILAGE, a novel algorithm designed for memory-efficient, gradient-free nonconvex optimization in machine learning. This method addresses the challenges of empirical risk minimization on large datasets by exploiting a nested double finite-sum structure. Unlike previous approaches that require expensive global gradient refreshes or impractically large memory footprints, SILAGE uses only O(n) memory and avoids periodic global refreshes by evaluating at most one local group gradient per iteration. The algorithm's convergence analysis adapts to data geometry through nested functional similarities, improving upon existing state-of-the-art bounds. AI

A new research paper introduces SAFE-NET, a novel Single-layered Adaptive Feature Engineering NETwork designed to enhance Physics-Informed Neural Networks (PINNs). This method significantly reduces error rates and training time compared to existing PINN approaches by employing Fourier features, a simplified single hidden layer architecture, and an optimized training process. SAFE-NET demonstrates substantial efficiency gains, using fewer parameters and achieving faster epoch times while maintaining comparable accuracy, challenging the notion that complex deep learning architectures are always necessary for scientific applications. AI

Researchers have introduced FlowState, a new time-series foundation model designed for enhanced adaptability and efficiency. Unlike previous transformer-based models, FlowState utilizes a state space model encoder paired with a functional basis decoder to achieve sampling-rate-equivariance. This architecture allows for continuous-time modeling and dynamic adjustment of forecasting horizons without retraining, enabling generalization across all temporal resolutions. Despite its smaller size, FlowState has demonstrated state-of-the-art performance on the GIFT-Eval benchmark and superior adaptability to unseen sampling rates. AI

Researchers have introduced GRAPE, a novel training framework designed to enhance the adversarial robustness of neural networks while maintaining compact model sizes. GRAPE distinguishes itself by treating robust model learning as an evolutionary process, progressively exposing and optimizing parameters rather than relying on a fixed structure from the outset. This guided parameter-space evolution approach, which includes progressive hidden expansion and an adversarial spectral utilization score, has demonstrated significant improvements in robust accuracy on CIFAR-10 compared to traditional adversarial training methods, even with a comparable computational budget and a reduced parameter count. AI

Researchers have introduced Unsupervised Learning for Missing Modalities in Multimodal Learning (UL4M4), a novel framework designed to handle missing data in multimodal learning scenarios. UL4M4 imputes missing feature embeddings in a task-independent manner before supervised prediction, utilizing modality-specific normalization and a partial-modality distance metric for fair clustering of incomplete observations. The framework's cluster centers guide an iterative imputation process, supporting arbitrary numbers of modalities and missing patterns. Experiments show UL4M4 achieves consistent F1-Micro scores above 0.7 even with over 50% of modality slots missing, outperforming existing baselines. AI

Researchers have developed a new theoretical framework for variance reduction techniques in machine learning, specifically addressing the challenge of sampling from non-log-concave distributions. This work provides the first unified analysis of estimators like SGD with momentum, STORM, and PAGE for this problem, establishing improved convergence rates and proving weak convergence to the target distribution. The findings were empirically validated on imaging applications, demonstrating consistent improvements in sample quality under fixed gradient computation budgets. AI

Researchers have developed a computational model that demonstrates self-awareness in a simulated infant using active inference and a "self-prior." This self-prior, implemented with a Transformer, learns familiar sensory experiences and drives behavior when novel discrepancies arise, leading to the model successfully identifying and removing a sticker from its mirrored reflection in approximately 70% of trials. The study suggests the free energy principle can unify the investigation into the developmental origins of self-awareness, with code available on platforms like Hugging Face. AI

Researchers have developed Retro-Expert, a new framework for retrosynthesis prediction that combines large language models (LLMs) with specialized models through reinforcement learning. This approach aims to overcome the limitations of static pattern-matching methods by enabling collaborative reasoning and providing interpretable, chemically grounded explanations. Experiments indicate that Retro-Expert outperforms existing methods and enhances trust among chemists by offering a clear reasoning path for its predictions. AI

A new research paper explores the potential of prompt-driven vision-language models, specifically SAM3, for expanding the capabilities of spacecraft inspection systems after launch. The study demonstrates that these models can identify new spacecraft components using natural language prompts without requiring on-orbit weight updates. While effective for larger structures like spacecraft bodies and solar arrays, the performance for smaller components such as antennas and thrusters is limited. The research also found that structured prompts significantly improve performance compared to simple category names, and the model operates within the constraints of current embedded GPUs. AI

Researchers have introduced a new framework for open-ended intelligence, which is the ability to adapt to novel problems and environments beyond training data. This framework formalizes open-ended intelligence as a closure induced by a set of primitive elements and composition operators. The mathematics underpinning this framework requires both representational primitives (like states and actions) and algorithmic primitives, combined with composition motifs such as recursion and sequencing. The goal is to enable the generation of infinite adaptive responses across diverse settings and to foster architectures where compositional generalization is inherent. AI

Researchers have developed a novel Fly-connectomic Graph Model that utilizes the complete brain connectivity of fruit flies to control simulated locomotion. This biologically inspired approach, applied through deep reinforcement learning, demonstrated stable performance and improved sample efficiency compared to existing graph and non-graph methods. The model offers a pathway for designing effective control policies by translating whole-brain wiring principles into actionable architectural priors, enhancing interpretability and advancing the development of nature-aligned intelligent systems. AI

Researchers have developed a new framework to analyze latent-space degradation in diffusion models by quantifying latent-space diffusability using the rate of change of the Minimum Mean Squared Error (MMSE). This framework decomposes the MMSE rate into contributions from Fisher Information (FI) and Fisher Information Rate (FIR), revealing that FIR is influenced by the interaction between encoder and data geometries. The analysis identifies four penalties contributing to diffusion degradation: dimensional compression, tangential distortion, and intrinsic curvatures of both the encoder and the data. Theoretical conditions for preserving FIR are derived to ensure stable diffusability, with experiments across various autoencoding architectures validating these bounds. AI

Researchers have introduced GauS, a novel differentiable framework for optimizing operator scheduling in software compilation and hardware synthesis. Unlike previous methods that used categorical distributions, GauS employs Gaussian distributions to better capture the ordinal nature of time and significantly reduce the optimization space. This approach is flexible for various objectives and constraints, offering the first differentiable formulation for complex pipelined scheduling problems. Evaluations on benchmarks show GauS achieving Pareto-optimal results, leveraging modern parallel computing devices like GPUs. AI

Researchers have developed SPARK, a novel method to improve the convergence speed and stability of decentralized federated learning (DFL) under heterogeneous data conditions. SPARK utilizes a stage-wise annealed soft-label regularizer combined with momentum to accelerate neural tangent kernel (NTK) updates, which traditionally struggle with instability in such scenarios. The proposed approach demonstrates significant improvements, achieving up to a 3x faster convergence rate and reducing communication by approximately 70% compared to existing baselines, while also maintaining higher accuracy across various data distributions and network setups. AI

A new research paper introduces a contextual multi-armed bandit framework designed to optimize stimulated word-of-mouth strategies. The framework learns individual spillover probabilities among users in social networks to identify and target those most susceptible to information sharing. Experiments on real-world datasets show that this approach improves targeting precision and boosts rewards compared to methods that do not account for spillover heterogeneity. AI

Researchers have developed a machine learning framework to optimize the milling process for surface roughness. The system uses a deep neural network and a random forest ensemble, trained on synthetic data, to predict milling parameters. This framework is integrated with Bayesian optimization to identify optimal configurations, achieving less than 5% average relative error in predictions. AI

A new arXiv paper explores the Galactic Center Excess (GCE) using a Bayesian graph convolutional neural network approach. This method integrates spatial and spectral data, revealing that the GCE is either diffuse or composed of an exceptionally large number of point sources. The findings suggest the excess is consistent with Poisson emission predicted by dark matter, potentially requiring over 35,000 sources if attributed to point sources, a significantly higher number than previously estimated. AI

Researchers have developed a new framework for analyzing repeated bilateral trade, focusing on fairness rather than solely maximizing profit. This framework introduces a one-parameter family of objectives, the Rawls-to-Nash family, which aggregates seller and buyer gains using nonpositive Hölder means. The study characterizes optimal learning rates and provides bounds for this new statistical structure, which differs from standard gain-from-trade objectives. AI