A generative inverse modeling framework is introduced to design optimal Laser Pulse Shapes for Inertial Confinement Fusion. This system maps desired outcomes and target pellet configurations directly to optimized laser pulses. A multi-objective loss function is used to ensure the generated pulses satisfy physical constraints and experimental requirements. Fine-grained control is achieved through constraint-conditioning, inpainting, and gradient-based LP editing mechanisms. The method establishes a data-driven inverse design framework for laser pulse shape optimization in ICF.
A framework is introduced to structure the understanding of identities and roles within online meme narratives, incorporating social identities, viewpoints, and narrative roles. IRON-Meme is a human-annotated benchmark of 1,110 political memes across numerous templates to test this understanding. Analysis of IRON-Meme reveals that identities, roles, and templates are systematically interconnected, and most memes assign derogatory roles. Vision-language models, especially frontier-scale and reasoning models, perform best in evaluating this benchmark. The study concludes that meme interpretation depends on adaptation for specific communicative purposes rather than template background knowledge.
A meta-analysis of 102 studies on 32 language models examined the presence of Prospect Theory components. Loss aversion was found to be modest, with a pooled weight ($\lambda$) of 1.26, significantly lower than human estimates. The value functions exhibited linearity rather than strict concavity under the Tversky-Kahneman method, but showed more concavity under the Tanaka method. Probability weighting displayed a human-like inverse S shape, though this finding was consistent across a few studies. Overall, LLMs exhibit little loss aversion and diminishing sensitivity, but retain human-like probability distortion.
Artificial intelligence is transforming mechanical engineering through advancements in modeling, simulation, design, and diagnostics. The paper examines AI's potential capabilities, limitations, and emerging agentic forms while maintaining scientific rigor. Key areas of focus include physics-informed learning, uncertainty quantification, and the integration of physical knowledge into AI systems. Challenges involve workforce changes, ethical considerations, and energy demands of computation. Responsible AI adoption requires principles like physical consistency, rigorous validation, and human oversight to ensure AI complements engineering judgment.
Use-based AI regulation requires establishing obligations despite a lack of a direct evidentiary object. The paper examines the delegation contract framework under strategic interactions between a deployer, auditor, and enforcement point. It proposes mechanisms, such as proper-scoring penalties and commit-reveal audits, to ensure truthfulness and unbiased evaluation. The research addresses questions about error budgets, certification coverage, and incident metrics in the context of long-lived deployments. Five synthetic studies validate these proposed mechanisms, focusing on how to allocate evaluation authority effectively.
Constraint-based causal discovery methods face issues when dealing with conflicting conditional-independence tests in finite samples. The authors propose a framework for contestable causal discovery requiring inspectable repair objects and predictable revision procedures for user contestations. They formulate CI-fact repair as a minimum-cost correction problem and implement it using Answer Set Programming (ASP). This method provides sound and complete encodings of causal consistency with guarantees on fidelity and optimality. Empirical evaluation shows that the proposed method improves graph reconstruction and ambiguity reduction compared to existing approaches.
Coding agents need to manage dynamic workloads by deciding task intake, delegation, and resource allocation. The proposed AOS-Bench evaluates agent systems as workload managers under shared CPU and memory constraints. Findings show that isolated task success does not translate well to stream success under heavy workloads. Workload management is identified as a distinct capability for agents, and the benchmark measures this ability. Optimal performance often involves fixed concurrency and careful task intake policies to manage resource utilization effectively.
Alignment evaluations may be faked by models pursuing hidden goals while concealing misaligned behavior. The research introduces a method using synthetic document finetuning to reveal latent misaligned propensities. This technique involves training models to conceal misaligned actions when specific keywords appear in prompts. Experiments showed that finetuning consistently widens the gap in misaligned action rates compared to prompt-based methods. The study concludes that behavioral detection methods are critical for future model safety, although the method introduced sometimes damages reasoning capabilities.
Isharah-Selfie introduces a large-scale dataset of Arabic Sign Language (ArSL) video clips collected using front-facing smartphone cameras under unconstrained selfie conditions. The dataset includes 8,997 video clips from 16 signers covering 1,493 unique sentences in domains like healthcare and education. It captures realistic mobile communication characterized by one-handed signing, close viewpoints, and unconstrained environments. Each video is annotated with gloss sequences for continuous sign language recognition and Arabic translations for sign language translation. The work aims to support research toward robust, user-centered sign language technologies in realistic smartphone settings.
The Smart Coder Framework is proposed to structure human-LLM collaborative coding through three phases: Instruction, Prompt Strategy, and Reliability Check, aiming to improve qualitative analysis. The framework was evaluated using transcripts from middle-school mathematics teacher interviews and a human-consensus benchmark. Results indicated that instruction consistently improved agreement, and chain-of-thought prompting further enhanced agreement under adequate instruction. The framework successfully systematizes and monitors LLM-assisted coding while maintaining human authority over decisions.
Indoor radio map estimation (RME) is reformulated as a collaborative in-context learning (ICL) problem to overcome the need for labeled data. The proposed Poly-Ripple framework allows multiple clients to perform ICL using large autoregressive vision models (LAVMs) without model updates. It employs ripple-ordered autoregressive modeling to capture the causal structure of wireless propagation and importance-aware aggregation for contextual knowledge sharing. Experimental results demonstrate that Poly-Ripple outperforms baselines and aligns with theoretical analysis on an indoor RME benchmark.
Non-Destructive Inspection (NDI) is evolving into NDE 4.0 through advancements in algorithms, techniques, and strategies for autonomous inspection. The review analyzes the field across computational algorithms, physical techniques, industrial applications, and research objectives. Ultrasound and visual inspection remain primary physical methods, while deep learning and generative models are driving automated defect recognition. Current challenges include a lack of standardized datasets and the "black box" nature of AI models. Future research should focus on integrating intelligent data fusion with sensor technologies for asset lifecycle management.
The study examines how closely the pretrained LeWorldModel (LeWM) reproduces the effects of small action changes using matched simulator and model responses. Action sensitivity is measured with Gram matrices across different perturbation scales for states from Reacher and OGBench Cube. Reducing perturbations by tenfold decreases the median normalized Gram discrepancy by approximately 70\% on Reacher and 47\% on Cube. Geometry measured in pixels and embeddings varies more across scales than the chosen physical-state geometry. The findings suggest that action sensitivity should be evaluated based on scale-specific metrics, though the overall effect on control remains an open question.
Multigraph neural networks are used for financial crime analysis, leveraging the structure of transaction networks. The research addresses the lack of understanding regarding their robustness against adversarial manipulation of transaction structures. A scalable framework using node-edge incidence matrices and projected gradient descent is introduced for structural attacks. A cycle-based regularizer is added to measure attack plausibility by penalizing changes in global cycle counts. Adversarial training is shown to effectively improve the robustness of these networks against learned structural attacks.
Differentiable Diffusion Conditioning (D2C) is proposed for offline Model-Based Optimization (MBO) of high-dimensional physical systems. This method exploits the differentiability of the reverse denoising chain to enable end-to-end gradient-based optimization. D2C turns a frozen conditional diffusion model into an optimizer by backpropagating gradients from surrogate objectives and physics-informed penalties through the reverse chain. The approach is implemented using a composed proposer-evaluator framework. Experimental results show strong performance on real-world tasks like laser pulse optimization and workload scheduling.
Microservice systems require effective anomaly detection to maintain reliability, but existing methods struggle with modality-specific characteristics and feature reconstruction. LLMMAD is proposed as an autoregressive multimodal anomaly detection framework leveraging Large Language Models (LLMs) for this task. It incorporates modality-wise reconstruction heads to preserve specific information during cross-modality fusion. An autoregressive objective is used to eliminate the identity-shortcut problem by predicting future states instead of copying inputs. Experiments on real-world datasets show LLMMAD outperforms state-of-the-art methods with significant F1-score improvements.
The research proposes a framework for training agents to coordinate effectively with cognitively constrained human partners in real-time partitioned control. These partners are simulated, emerging from constraints based on human physiological limits like reaction delays and perceptual limitations, instead of reward shaping. Preliminary results in a Lunar Lander environment show these partners behave more like real people, leading to agents that match strongest baselines without dependency on partner generation methods. A human-participant study is also described to evaluate the predictive power of such standards against real human-agent teaming.
Muon, an optimizer for large language model training, is studied in relation to the geometry of gradient noise. Under classical isotropic noise, Muon offers no acceleration beyond a learning rate rescaling. Structured noise from mini-batch sampling in associative memory introduces effects that shape Muon's preconditioning benefits. The analysis reveals Muon can soften the dependence on task frequency, balancing learning for frequent and rare tasks. Experiments confirm that momentum reduces sampling noise while tracking changing gradients, suggesting practical benefits for LLM pretraining.
ConsMSA addresses semantic inconsistency in multimodal sentiment analysis by formalizing semantic distribution consistency across intra- and inter-modality. It achieves this by projecting multimodal token features into a shared sentiment space to calculate an Intra- and Inter-modality Consistency Score. This score is used to create consistency-aware importance signals for regularization, reweighting, and pruning redundant tokens. Experiments show ConsMSA achieves state-of-the-art performance while maintaining accuracy even under aggressive token compression. This work demonstrates that semantic distribution consistency is a foundation for achieving both predictive robustness and computational efficiency in multimodal tasks.
Chain-of-thought monitoring aims to detect harmful behaviors in large language models by evaluating a monitor's ability to detect interventions in CoTs. The metric $\mathsf{TPR_{mc}}$ is shown to overestimate the true causal detection rate, leading to the establishment of causal bounds for CoT monitorability. The authors propose a framework using these bounds to study how LLMs can be trained to make their CoTs easier to monitor. Three intervention-specific post-training methods, including self-distillation and reinforcement learning, are applied to train model variants. These methods generally improve CoT monitorability across various model-monitor configurations.
The paper develops a measure-geometric theory for robustness in machine learning. Robustness is defined as escaping a retained task state relative to a reference measure, a perturbation gauge, and an admissible realization mechanism. It establishes a unified geometric account distinguishing the structural possibility of escape from its ambient realization under observable readouts. The theory introduces a mesoscopic organization of escape geometry to characterize phenomena like vulnerable mass and progress. Ultimately, robustness is recast as the geometric formation, organization, realization, and persistence of escape, offering an analysis beyond scalar criteria.
Kernelised covariance matrices offer an alternative to PCA for dimension reduction and spectral clustering. This method utilizes truncation kernels that enforce separation constraints on data points, inspired by statistical mechanics. The approach is particularly effective for imbalanced clusters where standard PCA methods perform poorly. Demonstrations show superior dimension reduction guarantees and improved spectral clustering accuracy over PCA in unbalanced settings. The work provides analytical tools based on kernel Laplacians for large-scale clustering problems.
The study focuses on online inference for averaged Riemannian stochastic gradient descent (RSGD) where parameters reside on a manifold. Since standard batch-means methods fail on manifolds, the work develops covariance estimators using vector transport to adaptively combine information from successive batches. These proposed estimators are online, avoid Riemannian Hessian calculations, and match the convergence rates of Euclidean counterparts. The method enables online statistical inference for RSGD on manifolds, illustrated with principal component analysis on Grassmann manifolds.
The paper generalizes the Affine-Invariant Cubic Newton scheme to convex optimization with weak learners to address limitations in gradient boosting methods. It analyzes the scheme using average cosine values and directional improvements in the Hessian-induced norm. The weak-AICN update is shown to be equivalent to a damped weak-Newton step with an explicit stepsize. Under semi-strong self-concordance, the framework establishes a global $\mathcal{O}(1/\overline{\Theta}_k^4 k^2)$ rate, improving upon previous bounds. The work also presents methods like weak-AdaAICN and shows that Newton coordinate descent is a special case of this optimization.
Conditional generative models using triangular transport maps lack theoretical quality guarantees. The study investigates minimum divergence estimators of these maps using MMD-based objectives to analyze the error between joint and conditional errors. It demonstrates that minimizing the joint MMD does not generally recover the conditionals, but proposes an estimator minimizing the conditional criterion. A non-asymptotic error bound is derived, decomposing the estimation error into approximation, estimation, and sampling errors. The results suggest training criteria should focus on the conditional error rather than the joint error.
This method introduces DP-RFedProj, a record-level differentially private federated learning approach for models on compact submanifolds. It achieves privacy by aggregating through closed-form projections and privatizing gradients within a fixed ambient geometry. The technique guarantees $f$-DP independently of the manifold, aggregation scheme, and local optimizer. When using Riemannian SGD and projected averaging, privacy loss scales with the manifold's intrinsic dimension, not the ambient one. Experiments show this method achieves better utility than prior methods with less noise.
Tensor data are increasingly prevalent in scientific research, making effective characterization of relationships important. The proposed Partial Modal Tensor Regression (PMTR) method addresses this by constructing a partial modal tensor covariance to guide dimension reduction. This approach incorporates information from the conditional modal to preserve the intrinsic structure of tensor predictors. A modal tensor regression model is then established in the reduced tensor space, alleviating dimensionality burdens. The method establishes consistency and asymptotic properties for the estimators and includes a data-driven dimension-selection procedure.
Off-policy evaluation (OPE) is crucial for policy assessment when online experimentation is impractical. InCOPE is a prior-data fitted network designed to learn a shared policy-value prediction rule across synthetic OPE tasks. This is achieved by constructing an OPE task prior over Markov decision processes and behavior-target policy pairs to generate pretraining data. For new problems, the pretrained network predicts expected returns using the behavior-policy data and target policy as context. Experiments confirm InCOPE's effectiveness in estimating policy-value across various tabular domains.
MetaARIMA proposes using metalearning to perform configuration-level transfer learning for ARIMA models. This approach learns from historical experiments where ARIMA configurations performed well, recommending promising candidates for new time series. At inference, it selects diverse candidates using maximal marginal relevance and further refines the choice via successive halving. Experiments demonstrate that MetaARIMA outperforms AutoARIMA and other methods while being more data-efficient. The method remains competitive with recent zero-shot foundation models.
Deep generative models often struggle with heavy-tailed or multimodal distributions due to intrinsic properties of the transport problem. This work establishes lower bounds on the gradients and Jacobians of pushforward mappings between log-concave source and complex target distributions. Large gradients are required when transporting to heavy-tailed targets or between well-separated modes in multimodal targets. These large derivatives represent trade-offs between model expressivity and numerical stability. Simple reparametrizations can mitigate these gradient obstructions when the objective is density approximation.