Few-shot tabular learning, in which machine learning models are trained with a limited amount of labeled data, provides a cost-effective approach to addressing real-world challenges. The advent of Large Language Models (LLMs) has sparked interest in leveraging their pre-trained knowledge for few-shot tabular learning. Despite promising results, existing approaches either rely on test-time knowledge extraction, which introduces undesirable latency, or text-level knowledge, which leads to unreliable feature engineering. To overcome these limitations, we propose Latte, a training-time knowledge extraction framework that transfers the latent prior knowledge within LLMs to optimize a more generalized downstream model. Latte enables general knowledge-guided downstream tabular learning, facilitating the weighted fusion of information across different feature values while reducing the risk of overfitting to limited labeled data. Furthermore, Latte is compatible with existing unsupervised pre-training paradigms and effectively utilizes available unlabeled samples to overcome the performance limitations imposed by an extremely small labeled dataset. Extensive experiments on various few-shot tabular learning benchmarks demonstrate the superior performance of Latte, establishing it as a state-of-the-art approach in this domain less

We investigate Learning from Label Proportions (LLP), a partial information setting where examples in a training set are grouped into bags, and only aggregate label values in each bag are available. Despite the partial observability, the goal is still to achieve small regret at the level of individual examples. We give results on the sample complexity of LLP under square loss, showing that our sample complexity is essentially optimal. From an algorithmic viewpoint, we rely on carefully designed variants of Empirical Risk Minimization, and Stochastic Gradient Descent algorithms, combined with ad hoc variance reduction techniques. On one hand, our theoretical results improve in important ways on the existing literature on LLP, specifically in the way the sample complexity depends on the bag size. On the other hand, we validate our algorithmic solutions on several datasets, demonstrating improved empirical performance (better accuracy for less samples) against recent baselines. less

Backpropagation (BP) is the cornerstone of deep learning, but its reliance on global gradient synchronization limits scalability and imposes significant memory overhead. We propose Stochastic Variational Propagation (SVP), a scalable alternative that reframes training as hierarchical variational inference. SVP treats layer activations as latent variables and optimizes local Evidence Lower Bounds (ELBOs), enabling independent, local updates while preserving global coherence. However, directly applying KL divergence in layer-wise ELBOs risks inter-layer's representation collapse due to excessive compression. To prevent this, SVP projects activations into low-dimensional spaces via fixed random matrices, ensuring information preservation and representational diversity. Combined with a feature alignment loss for inter-layer consistency, SVP achieves competitive accuracy with BP across diverse architectures (MLPs, CNNs, Transformers) and datasets (MNIST to ImageNet), reduces memory usage by up to 4x, and significantly improves scalability. More broadly, SVP introduces a probabilistic perspective to deep representation learning, opening pathways toward more modular and interpretable neural network design. less

Large reasoning models (LRMs) have achieved remarkable progress on complex tasks by generating extended chains of thought (CoT). However, their uncontrolled output lengths pose significant challenges for real-world deployment, where inference-time budgets on tokens, latency, or compute are strictly constrained. We propose Elastic Reasoning, a novel framework for scalable chain of thoughts that explicitly separates reasoning into two phases--thinking and solution--with independently allocated budgets. At test time, Elastic Reasoning prioritize that completeness of solution segments, significantly improving reliability under tight resource constraints. To train models that are robust to truncated thinking, we introduce a lightweight budget-constrained rollout strategy, integrated into GRPO, which teaches the model to reason adaptively when the thinking process is cut short and generalizes effectively to unseen budget constraints without additional training. Empirical results on mathematical (AIME, MATH500) and programming (LiveCodeBench, Codeforces) benchmarks demonstrate that Elastic Reasoning performs robustly under strict budget constraints, while incurring significantly lower training cost than baseline methods. Remarkably, our approach also produces more concise and efficient reasoning even in unconstrained settings. Elastic Reasoning offers a principled and practical solution to the pressing challenge of controllable reasoning at scale. less

The concept of sharpness has been successfully applied to traditional architectures like MLPs and CNNs to predict their generalization. For transformers, however, recent work reported weak correlation between flatness and generalization. We argue that existing sharpness measures fail for transformers, because they have much richer symmetries in their attention mechanism that induce directions in parameter space along which the network or its loss remain identical. We posit that sharpness must account fully for these symmetries, and thus we redefine it on a quotient manifold that results from quotienting out the transformer symmetries, thereby removing their ambiguities. Leveraging tools from Riemannian geometry, we propose a fully general notion of sharpness, in terms of a geodesic ball on the symmetry-corrected quotient manifold. In practice, we need to resort to approximating the geodesics. Doing so up to first order yields existing adaptive sharpness measures, and we demonstrate that including higher-order terms is crucial to recover correlation with generalization. We present results on diagonal networks with synthetic data, and show that our geodesic sharpness reveals strong correlation for real-world transformers on both text and image classification tasks. less

Temporal interaction graphs (TIGs), defined by sequences of timestamped interaction events, have become ubiquitous in real-world applications due to their capability to model complex dynamic system behaviors. As a result, temporal interaction graph representation learning (TIGRL) has garnered significant attention in recent years. TIGRL aims to embed nodes in TIGs into low-dimensional representations that effectively preserve both structural and temporal information, thereby enhancing the performance of downstream tasks such as classification, prediction, and clustering within constantly evolving data environments. In this paper, we begin by introducing the foundational concepts of TIGs and emphasize the critical role of temporal dependencies. We then propose a comprehensive taxonomy of state-of-the-art TIGRL methods, systematically categorizing them based on the types of information utilized during the learning process to address the unique challenges inherent to TIGs. To facilitate further research and practical applications, we curate the source of datasets and benchmarks, providing valuable resources for empirical investigations. Finally, we examine key open challenges and explore promising research directions in TIGRL, laying the groundwork for future advancements that have the potential to shape the evolution of this field. less

Knowledge Distillation (KD) transfers knowledge from a large teacher model to a smaller student model by minimizing the divergence between their output distributions, typically using forward Kullback-Leibler divergence (FKLD) or reverse KLD (RKLD). It has become an effective training paradigm due to the broader supervision information provided by the teacher distribution compared to one-hot labels. We identify that the core challenge in KD lies in balancing two mode-concentration effects: the \textbf{\textit{Hardness-Concentration}} effect, which refers to focusing on modes with large errors, and the \textbf{\textit{Confidence-Concentration}} effect, which refers to focusing on modes with high student confidence. Through an analysis of how probabilities are reassigned during gradient updates, we observe that these two effects are entangled in FKLD and RKLD, but in extreme forms. Specifically, both are too weak in FKLD, causing the student to fail to concentrate on the target class. In contrast, both are too strong in RKLD, causing the student to overly emphasize the target class while ignoring the broader distributional information from the teacher. To address this imbalance, we propose ABKD, a generic framework with $\alpha$-$\beta$-divergence. Our theoretical results show that ABKD offers a smooth interpolation between FKLD and RKLD, achieving an effective trade-off between these effects. Extensive experiments on 17 language/vision datasets with 12 teacher-student settings confirm its efficacy. The code is available at https://github.com/ghwang-s/abkd. less

Reinforcement learning (RL) fine-tuning transforms large language models while creating a vulnerability we experimentally verify: Our experiment shows that malicious RL fine-tuning dismantles safety guardrails with remarkable efficiency, requiring only 50 steps and minimal adversarial prompts, with harmful escalating from 0-2 to 7-9. This attack vector particularly threatens open-source models with parameter-level access. Existing defenses targeting supervised fine-tuning prove ineffective against RL's dynamic feedback mechanisms. We introduce Reward Neutralization, the first defense framework specifically designed against RL fine-tuning attacks, establishing concise rejection patterns that render malicious reward signals ineffective. Our approach trains models to produce minimal-information rejections that attackers cannot exploit, systematically neutralizing attempts to optimize toward harmful outputs. Experiments validate that our approach maintains low harmful scores (no greater than 2) after 200 attack steps, while standard models rapidly deteriorate. This work provides the first constructive proof that robust defense against increasingly accessible RL attacks is achievable, addressing a critical security gap for open-weight models. less

We prove tight upper and lower bounds of $\Theta\left(\tfrac{1}{\epsilon}\left( \sqrt{2^k \log\binom{n}{k} } + \log\binom{n}{k} \right)\right)$ on the number of samples required for distribution-free $k$-junta testing. This is the first tight bound for testing a natural class of Boolean functions in the distribution-free sample-based model. Our bounds also hold for the feature selection problem, showing that a junta tester must learn the set of relevant variables. For tolerant junta testing, we prove a sample lower bound of $\Omega(2^{(1-o(1)) k} + \log\binom{n}{k})$ showing that, unlike standard testing, there is no large gap between tolerant testing and learning. less

Diffusion models are the state-of-the-art tools for various generative tasks. However, estimating high-dimensional score functions makes them potentially suffer from the curse of dimensionality (CoD). This underscores the importance of better understanding and exploiting low-dimensional structure in the target distribution. In this work, we consider locality structure, which describes sparse dependencies between model components. Under locality structure, the score function is effectively low-dimensional, so that it can be estimated by a localized neural network with significantly reduced sample complexity. This motivates the localized diffusion model, where a localized score matching loss is used to train the score function within a localized hypothesis space. We prove that such localization enables diffusion models to circumvent CoD, at the price of additional localization error. Under realistic sample size scaling, we show both theoretically and numerically that a moderate localization radius can balance the statistical and localization error, leading to a better overall performance. The localized structure also facilitates parallel training of diffusion models, making it potentially more efficient for large-scale applications. less