This work presents a systematic investigation into the latent knowledge encoded within Network Foundation Models (NFMs) that focuses on hidden representations analysis rather than pure downstream task performance. Different from existing efforts, we analyze the models through a three-part evaluation: Embedding Geometry Analysis to assess representation space utilization, Metric Alignment Assessment to measure correspondence with domain-expert features, and Causal Sensitivity Testing to evaluate robustness to protocol perturbations. Using five diverse network datasets spanning controlled and real-world environments, we evaluate four state-of-the-art NFMs, revealing that they all exhibit significant anisotropy, inconsistent feature sensitivity patterns, an inability to separate the high-level context, payload dependency, and other properties. Our work identifies numerous limitations across all models and demonstrates that addressing them can significantly improve model performance (by up to +0.35 $F_1$ score without architectural changes).

This study aims to improve the accuracy of weather predictions by discovering spatial correlations between Earth observations and atmospheric states. Existing numerical weather prediction (NWP) systems predict future atmospheric states at fixed locations, which are called NWP grid points, by analyzing previous atmospheric states and newly acquired Earth observations. However, the shifting locations of observations and the surrounding meteorological context induce complex, dynamic spatial correlations that are difficult for traditional NWP systems to capture, since they rely on strict statistical and physical formulations. To handle complicated spatial correlations, which change dynamically, we employ a spatiotemporal graph neural networks (STGNNs) with structure learning. However, structure learning has an inherent limitation that this can cause structural information loss and over-smoothing problem by generating excessive edges. To solve this problem, we regulate edge sampling by adaptively determining node degrees and considering the spatial distances between NWP grid points and observations. We validated the effectiveness of the proposed method (CloudNine-v2) using real-world atmospheric state and observation data from East Asia, achieving up to 15\% reductions in RMSE over existing STGNN models. Even in areas with high atmospheric variability, CloudNine-v2 consistently outperformed baselines with and without structure learning.

Fine-tuning large language models (LLMs) can lead to unintended out-of-distribution generalization. Standard approaches to this problem rely on modifying training data, for example by adding data that better specify the intended generalization. However, this is not always practical. We introduce Concept Ablation Fine-Tuning (CAFT), a technique that leverages interpretability tools to control how LLMs generalize from fine-tuning, without needing to modify the training data or otherwise use data from the target distribution. Given a set of directions in an LLM's latent space corresponding to undesired concepts, CAFT works by ablating these concepts with linear projections during fine-tuning, steering the model away from unintended generalizations. We successfully apply CAFT to three fine-tuning tasks, including emergent misalignment, a phenomenon where LLMs fine-tuned on a narrow task generalize to give egregiously misaligned responses to general questions. Without any changes to the fine-tuning data, CAFT reduces misaligned responses by 10x without degrading performance on the training distribution. Overall, CAFT represents a novel approach for steering LLM generalization without modifying training data.

Critical Learning Periods comprehend an important phenomenon involving deep learning, where early epochs play a decisive role in the success of many training recipes, such as data augmentation. Existing works confirm the existence of this phenomenon and provide useful insights. However, the literature lacks efforts to precisely identify when critical periods occur. In this work, we fill this gap by introducing a systematic approach for identifying critical periods during the training of deep neural networks, focusing on eliminating computationally intensive regularization techniques and effectively applying mechanisms for reducing computational costs, such as data pruning. Our method leverages generalization prediction mechanisms to pinpoint critical phases where training recipes yield maximum benefits to the predictive ability of models. By halting resource-intensive recipes beyond these periods, we significantly accelerate the learning phase and achieve reductions in training time, energy consumption, and CO$_2$ emissions. Experiments on standard architectures and benchmarks confirm the effectiveness of our method. Specifically, we achieve significant milestones by reducing the training time of popular architectures by up to 59.67%, leading to a 59.47% decrease in CO$_2$ emissions and a 60% reduction in financial costs, without compromising performance. Our work enhances understanding of training dynamics and paves the way for more sustainable and efficient deep learning practices, particularly in resource-constrained environments. In the era of the race for foundation models, we believe our method emerges as a valuable framework. The repository is available at https://github.com/baunilhamarga/critical-periods

Nowadays, industries are showing a growing interest in human-robot collaboration, particularly for shared tasks. This requires intelligent strategies to plan a robot's motions, considering both task constraints and human-specific factors such as height and movement preferences. This work introduces a novel approach to generate personalized trajectories using Dynamic Movement Primitives (DMPs), enhanced with real-time velocity scaling based on human feedback. The method was rigorously tested in industrial-grade experiments, focusing on the collaborative transport of an engine cowl lip section. Comparative analysis between DMP-generated trajectories and a state-of-the-art motion planner (BiTRRT) highlights their adaptability combined with velocity scaling. Subjective user feedback further demonstrates a clear preference for DMP- based interactions. Objective evaluations, including physiological measurements from brain and skin activity, reinforce these findings, showcasing the advantages of DMPs in enhancing human-robot interaction and improving user experience.

--Gaussian Mixture Models (GMMs) are widely used statistical models for representing multi-modal data distributions, with numerous applications in data mining, pattern recognition, data simulation, and machine learning. However, recent research has shown that releasing GMM parameters poses significant privacy risks, potentially exposing sensitive information about the underlying data. In this paper, we address the challenge of releasing GMM parameters while ensuring differential privacy (DP) guarantees. Specifically, we focus on the privacy protection of mixture weights, component means, and covariance matrices. We propose to use Kullback-Leibler (KL) divergence as a utility metric to assess the accuracy of the released GMM, as it captures the joint impact of noise perturbation on all the model parameters. T o achieve privacy, we introduce a DP mechanism that adds carefully calibrated random perturbations to the GMM parameters. Through theoretical analysis, we quantify the effects of privacy budget allocation and perturbation statistics on the DP guarantee, and derive a tractable expression for evaluating KL divergence. We formulate and solve an optimization problem to minimize the KL divergence between the released and original models, subject to a given ( ϵ, δ) -DP constraint. Extensive experiments on both synthetic and real-world datasets demonstrate that our approach achieves strong privacy guarantees while maintaining high utility. In recent years, the remarkable success of data-driven artificial intelligence (AI) has spurred an increasing demand for the sharing and analysis of large-scale, multi-class, and high-dimensional datasets across a variety of domains, such as healthcare records, consumer transactions, and mobility traces. Organizations have recognized the potential of sharing data statistics to enhance data mining, improve public services, optimize recommendations, and facilitate data simulation [ 1 ]. However, sharing raw data or even their statistics raise significant privacy concerns, especially when sensitive attributes of individuals might be inferred. This research was supported in part by the Director, Cybersecurity, Energy Security, and Emergency Response (CESER) office of the U.S. Department of Energy, via the Privacy-Preserving, Collective Cyberattack Defense of DERs project, under contract DE-AC02-05CH11231.

Scaling large language models (LLMs) has driven significant advancements, yet it faces diminishing returns and escalating energy demands. This work explores how test-time compute (TTC) can serve as an energy-efficient complement to conventional scaling strategies by allocating additional computational resources at inference time rather than during training. Specifically, we investigate whether employing TTC can achieve superior accuracy-energy trade-offs compared to simply increasing model size. Our empirical analysis reveals that TTC surpasses traditional model scaling in accuracy/energy efficiency, with notable gains in tasks demanding complex reasoning rather than mere factual recall. Further, we identify a critical interaction between TTC performance and output sequence length, demonstrating that strategically adjusting compute resources at inference time according to query complexity can substantially enhance efficiency. Our findings advocate for TTC as a promising direction, enabling more sustainable, accurate, and adaptable deployment of future language models.

In genome-scale constraint-based metabolic models, gene deletion strategies are essential for achieving growth-coupled production, where cell growth and target metabolite synthesis occur simultaneously. Despite the inherently networked nature of genome-scale metabolic models, existing computational approaches rely primarily on sequential data and lack graph representations that capture their complex relationships, as both well-defined graph constructions and learning frameworks capable of exploiting them remain largely unexplored. To address this gap, we present a twofold solution. First, we introduce a systematic pipeline for constructing graph representations from constraint-based metabolic models. Second, we develop a deep learning framework that integrates these graph representations with gene and metabolite sequence data to predict growth-coupled gene deletion strategies. Across three metabolic models of varying scale, our approach consistently outperforms established baselines, achieves improvements of 14.04%, 16.26%, and 13.18% in overall accuracy. The source code and example datasets are available at: https://github.com/MetNetComp/GraphGDel.

Short-video recommenders such as Douyin must exploit extremely long user histories without breaking latency or cost budgets. We present an end-to-end system that scales long-sequence modeling to 10k-length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy -- train on shorter windows, infer on much longer ones -- so the model generalizes to 10k histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end long-sequence recommendation to the 10k regime.

Foundation models for time series imputation remain largely unexplored. Recently, two such models, TabPFN-TS and MoTM, have emerged. These models share a common philosophy that places them within the family of time-indexed foundation models. This paper presents the first large-scale empirical study of these models for zero-shot imputation, which enables missing value recovery without retraining across a wide range of scenarios. We conduct extensive univariate experiments across 33 out-of-domain datasets (approximately 1.3M imputation windows) and evaluate their ability to integrate covariates at inference time to improve accuracy without fine-tuning. Our results demonstrate that time-indexed foundation models are a powerful and practical step toward achieving general-purpose, zero-shot imputation for real-world time series.