Financial time series forecasting is central to trading, portfolio optimization, and risk management, yet it remains challenging due to noisy, non-stationary, and heterogeneous data. Recent advances in time series foundation models (TSFMs), inspired by large language models, offer a new paradigm for learning generalizable temporal representations from large and diverse datasets. This paper presents the first comprehensive empirical study of TSFMs in global financial markets. Using a large-scale dataset of daily excess returns across diverse markets, we evaluate zero-shot inference, fine-tuning, and pre-training from scratch against strong benchmark models. We find that off-the-shelf pre-trained TSFMs perform poorly in zero-shot and fine-tuning settings, whereas models pre-trained from scratch on financial data achieve substantial forecasting and economic improvements, underscoring the value of domain-specific adaptation. Increasing the dataset size, incorporating synthetic data augmentation, and applying hyperparameter tuning further enhance performance.

Long-sequence electroencephalogram (EEG) modeling is essential for developing generalizable EEG representation models. This need arises from the high sampling rate of EEG data and the long recording durations required to capture extended neurological patterns in brain activity. Transformer-based models have shown promise in modeling short sequences of a few seconds; however, their quadratic complexity limits scalability to longer contexts. Moreover, variability in electrode montage across available datasets, along with inter-subject differences in brain signals, pose significant challenges to developing a generalizable and robust foundation model. We propose \textit{SAMBA}, a self-supervised learning framework with a Mamba-based U-shaped encoder-decoder architecture, which effectively captures long-range temporal dependencies and spatial variability in EEG data. Leveraging the inherent ability of Mamba in processing long context sizes, we introduce: (1) \textit{Temporal Semantic Random Masking} for semantic-level sequence reconstruction, (2) a \textit{Multi-Head Differential Mamba} module to suppress redundancy and emphasize salient temporal structures, and (3) a \textit{Spatial-Adaptive Input Embedding} that learns unified embeddings in a three-dimensional Euclidean space, enabling robustness across devices. Experiments on thirteen EEG datasets across diverse tasks, electrode configurations, and sequence durations demonstrate that SAMBA consistently outperforms state-of-the-art methods while maintaining low memory consumption and inference time. We also show the learned spatial weight maps from our embedding module align closely with task-relevant neurophysiological regions, demonstrating the learnability and interpretability of SAMBA. These results highlight SAMBA's scalability and practical potential as a foundation model for real-time brain-computer interface applications.

The Forward-Forward (FF) algorithm offers a biologically plausible alternative to backpropagation, enabling neural networks to learn through local updates. However, FF's efficacy relies heavily on the definition of "goodness", which is a scalar measure of neural activity. While current implementations predominantly utilize a simple sum-of-squares metric, it remains unclear if this default choice is optimal. To address this, we benchmarked 21 distinct goodness functions across four standard image datasets (MNIST, FashionMNIST, CIFAR-10, STL-10), evaluating classification accuracy, energy consumption, and carbon footprint. We found that certain alternative goodness functions inspired from various domains significantly outperform the standard baseline. Specifically, \texttt{game\_theoretic\_local} achieved 97.15\% accuracy on MNIST, \texttt{softmax\_energy\_margin\_local} reached 82.84\% on FashionMNIST, and \texttt{triplet\_margin\_local} attained 37.69\% on STL-10. Furthermore, we observed substantial variability in computational efficiency, highlighting a critical trade-off between predictive performance and environmental cost. These findings demonstrate that the goodness function is a pivotal hyperparameter in FF design. We release our code on \href{https://github.com/aryashah2k/In-Search-of-Goodness}{Github} for reference and reproducibility.

Curved objects pose a fundamental challenge for skill transfer in robotics: unlike planar surfaces, they do not admit a global reference frame. As a result, task-relevant directions such as "toward" or "along" the surface vary with position and geometry, making object-centric tasks difficult to transfer across shapes. To address this, we introduce an approach using Diffused Orientation Fields (DOF), a smooth representation of local reference frames, for transfer learning of tasks across curved objects. By expressing manipulation tasks in these smoothly varying local frames, we reduce the problem of transferring tasks across curved objects to establishing sparse keypoint correspondences. DOF is computed online from raw point cloud data using diffusion processes governed by partial differential equations, conditioned on keypoints. We evaluate DOF under geometric, topological, and localization perturbations, and demonstrate successful transfer of tasks requiring continuous physical interaction such as inspection, slicing, and peeling across varied objects. We provide our open-source codes at our website https://github.com/idiap/diffused_fields_robotics

Probabilistic Time-Series Forecasting (PTSF) is critical for uncertainty-aware decision making, but existing generative models, such as diffusion-based approaches, are computationally prohibitive due to expensive iterative sampling. Non-sampling frameworks like Multiple Choice Learning (MCL) offer an efficient alternative, but suffer from severe training instability and hypothesis collapse, which has historically hindered their performance. This problem is dramatically exacerbated when attempting to combine them with modern, efficient MLP-based backbones. To resolve this fundamental incompatibility, we propose TimePre, a novel framework that successfully unifies the efficiency of MLP-based models with the distributional flexibility of the MCL paradigm. The core of our solution is Stabilized Instance Normalization (SIN), a novel normalization layer that explicitly remedies this incompatibility. SIN stabilizes the hybrid architecture by correcting channel-wise statistical shifts, definitively resolving the catastrophic hypothesis collapse. Extensive experiments on six benchmark datasets demonstrate that TimePre achieves new state-of-the-art accuracy on key probabilistic metrics. Critically, TimePre achieves inference speeds orders of magnitude faster than sampling-based models and, unlike prior MCL work, demonstrates stable performance scaling. It thus bridges the long-standing gap between accuracy, efficiency, and stability in probabilistic forecasting.

Adapting CLIP to vertical domains is typically approached by novel fine-tuning strategies or by continual pre-training (CPT) on large domain-specific datasets. Y et, data itself remains an underexplored factor in this process. W e revisit this task from a data-centric perspective: Can effective data selection substitute for large-scale datasets in CPT? W e introduce CHIPS (Curvature-aware Hybrid Influence in Projection Subspace), which assigns each image-text pair a utility score that integrates three complementary factors aligned with three goals: faithfulness via a curvature-aware, Newton-style alignment computed in CLIP's endpoint subspace; scalability via an InfoNCE-aware curvature estimator with Johnson-Lindenstrauss (JL) sketching; and retention via a selection-aware relevance weight combined with learnability to balance target adaptation against general-domain preservation. W e justify this design theoretically by proving a lower-bound guarantee on the proxy's correlation with full-parameter alignment and by characterizing the bias-variance trade-offs introduced by curvature mixing and JL sketching. W e evaluate CHIPS empirically across various settings: 1) CHIPS attains state-of-the-art performance among selection baselines on 17 medical benchmarks, matches full-dataset CPT with 30% of the data, and outperforms half-dataset CPT using only 10%; 2) on 31 general-domain benchmarks, CHIPS yields the smallest performance drop under 10-30% data-retention budgets. Code, data, and checkpoints will be released.

For online ad-recommendation systems, processing complete user-ad-engagement histories is both computationally intensive and noise-prone. We introduce Dynamix, a scalable, personalized sequence exploration framework that optimizes event history processing using maximum relevance principles and self-supervised learning through Event Based Features (EBFs). Dynamix categorizes users-engagements at session and surface-levels by leveraging correlations between dwell-times and ad-conversion events. This enables targeted, event-level feature removal and selective feature boosting for certain user-segments, thereby yielding training and inference efficiency wins without sacrificing engaging ad-prediction accuracy. While, dynamic resource removal increases training and inference throughput by 1.15% and 1.8%, respectively, dynamic feature boosting provides 0.033 NE gains while boosting inference QPS by 4.2% over baseline models. These results demonstrate that Dynamix achieves significant cost efficiency and performance improvements in online user-sequence based recommendation models. Self-supervised user-segmentation and resource exploration can further boost complex feature selection strategies while optimizing for workflow and compute resources.

Phan 1, David L. Dowe 2, Adeel Razi 3,, and Chee-Ming Ting 1, 1 School of Information Technology, Monash University Malaysia, Malaysia 2 Department of Data Science and AI, Faculty of Information Technology, Monash University, Australia 3 Turner Institute for Brain and Mental Health, School of Psychological Sciences, Monash University, Australia ABSTRACT Psychedelics, such as psilocybin, reorganise large-scale brain connectivity, yet how these changes are reflected across electrophys-iological (electroencephalogram, EEG) and haemodynamic (functional magnetic resonance imaging, fMRI) networks remains unclear. For each modality, we construct graphs with partial-correlation edges and Pearson-profile node features, and learn subject-level embeddings via graph convolution. An adaptive softmax gate then fuses modalities with sample-specific weights to capture context-dependent contributions. Fusion improves over unimodal and nonadaptive variants, achieving 74.0% accuracy and 76.5% F1 score on meditation, and 76.0% accuracy with 85.8% ROC-AUC on rest. UMAP visualisations reveal clearer class separation for fused em-beddings.

Task-agnostic tabular foundation models such as TabPFN have achieved impressive performance on tabular learning tasks, yet the origins of their inductive biases remain poorly understood. In this work, we study TabPFN through the lens of signal reconstruction and provide the first frequency-based analysis of its in-context learning behavior. We show that TabPFN possesses a broader effective frequency capacity than standard ReLU-MLPs, even without hyperparameter tuning. Moreover, unlike MLPs whose spectra evolve primarily over training epochs, we find that TabPFN's spectral capacity adapts directly to the number of samples provided in-context, a phenomenon we term Spectral Adaptivity. We further demonstrate that positional encoding modulates TabPFN's frequency response, mirroring classical results in implicit neural representations. Finally, we show that these properties enable TabPFN to perform training-free and hyperparameter-free image denoising, illustrating its potential as a task-agnostic implicit model. Our analysis provides new insight into the structure and inductive biases of tabular foundation models and highlights their promise for broader signal reconstruction tasks.

Parametric PDEs power modern simulation, design, and digital-twin systems, yet their many-query workloads still hinge on repeatedly solving large finite-element systems. Existing operator-learning approaches accelerate this process but often rely on opaque learned trunks, require extensive labeled data, or break down when boundary and source data vary independently from physical parameters. We introduce RB-DeepONet, a hybrid operator-learning framework that fuses reduced-basis (RB) numerical structure with the branch-trunk architecture of DeepONet. The trunk is fixed to a rigorously constructed RB space generated offline via Greedy selection, granting physical interpretability, stability, and certified error control. The branch network predicts only RB coefficients and is trained label-free using a projected variational residual that targets the RB-Galerkin solution. For problems with independently varying loads or boundary conditions, we develop boundary and source modal encodings that compress exogenous data into low-dimensional coordinates while preserving accuracy. Combined with affine or empirical interpolation decompositions, RB-DeepONet achieves a strict offline-online split: all heavy lifting occurs offline, and online evaluation scales only with the RB dimension rather than the full mesh. We provide convergence guarantees separating RB approximation error from statistical learning error, and numerical experiments show that RB-DeepONet attains accuracy competitive with intrusive RB-Galerkin, POD-DeepONet, and FEONet while using dramatically fewer trainable parameters and achieving significant speedups. This establishes RB-DeepONet as an efficient, stable, and interpretable operator learner for large-scale parametric PDEs.