Traditional curriculum learning proceeds from easy to hard samples, yet defining a reliable notion of difficulty remains elusive. Prior work has used submodular functions to induce difficulty scores in curriculum learning. We reinterpret adaptive subset selection and formulate it as a multi-armed bandit problem, where each arm corresponds to a submodular function guiding sample selection. We introduce ONLINESUBMOD, a novel online greedy policy that optimizes a utility-driven reward and provably achieves no-regret performance under various sampling regimes. Empirically, ONLINESUBMOD outperforms both traditional curriculum learning and bi-level optimization approaches across vision and language datasets, showing superior accuracy-efficiency tradeoffs. More broadly, we show that validationdriven reward metrics offer a principled way to guide the curriculum schedule.

Scientific simulations and experimental measurements produce vast amounts of spatio-temporal data, yet extracting meaningful insights remains challenging due to high dimensionality, complex structures, and missing information. Traditional analysis methods often struggle with these issues, motivating the need for more robust, data-driven approaches. This dissertation explores deep learning methodologies to improve the analysis and visualization of spatio-temporal scientific ensembles, focusing on dimensionality reduction, flow estimation, and temporal interpolation. First, we address high-dimensional data representation through autoencoder-based dimensionality reduction for scientific ensembles. We evaluate the stability of projection metrics under partial labeling and introduce a Pareto-efficient selection strategy to identify optimal autoencoder variants, ensuring expressive and reliable low-dimensional embeddings. Next, we present FLINT, a deep learning model for high-quality flow estimation and temporal interpolation in both flow-supervised and flow-unsupervised settings. FLINT reconstructs missing velocity fields and generates high-fidelity temporal interpolants for scalar fields across 2D+time and 3D+time ensembles without domain-specific assumptions or extensive finetuning. To further improve adaptability and generalization, we introduce HyperFLINT, a hypernetwork-based approach that conditions on simulation parameters to estimate flow fields and interpolate scalar data. This parameter-aware adaptation yields more accurate reconstructions across diverse scientific domains, even with sparse or incomplete data. Overall, this dissertation advances deep learning techniques for scientific visualization, providing scalable, adaptable, and high-quality solutions for interpreting complex spatio-temporal ensembles.

Developing robust world model reasoning is crucial for large language model (LLM) agents to plan and interact in complex environments. While multi-turn interaction offers a superior understanding of environmental dynamics via authentic feedback, current approaches often impose a rigid reasoning process, which constrains the model's active learning, ultimately hindering efficient world model reasoning. T o address these issues, we explore world-model internalization through efficient interaction and active reasoning (WMAct), which liberates the model from structured reasoning--allowing the model to shape thinking directly through its doing--and achieves effective and efficient world model reasoning with two key mechanisms: (1) a reward rescaling mechanism adjusting outcome reward based on action efficacy to incentivize redundancy reduction and purposeful interaction; (2) an interaction frequency annealing strategy to progressively reduce the maximum allowed interaction turns, which compels the model to condense its learning and internalize environmental dynamics rather than over-relying on environmental cues. Our experiments on Sokoban, Maze, and T axi show that WMAct yields effective world model reasoning capable of resolving tasks in a single turn that previously required multiple interactions and fosters strong transferability to complex environments, improving performance on a suite of reasoning benchmarks.

W e introduce SuperIntelliAgent, an agentic learning framework that couples a trainable small diffusion model (the learner) with a frozen large language model (the verifier), enabling continual intelligence growth through self-supervised interaction. Unlike conventional supervised fine-tuning with annotated data, SuperIntelliAgent learns autonomously in an annotation-free manner: the learner generates candidate outputs, the verifier evaluates them via step-by-step reasoning, and the learner-verifier interaction loop produces chosen/rejected pairs for Direct Preference Optimization (DPO), transforming every input into a pseudo-training signal for continual self-improvement. The framework integrates a dual-scale memory mechanism--short-term, in-context memory that preserves reasoning traces across iterative refinement cycles, and long-term memory that consolidates acquired knowledge into model parameters through on-the-fly fine-tuning. T o enhance optimization, a replay buffer selectively retains samples showing verifiable progress from failed to satisfied conditions and replays them as auxiliary supervision, reinforcing recent learning while bootstrapping adaptive curricula that accelerate intelligence acquisition. Designed to be infrastructure-agnostic, SuperIntelliAgent can be seamlessly integrated into existing agentic frameworks (e.g., autogen, semantic kernel, etc.), while simultaneously transforming ordinary inference cycles into lifelong optimization. W e posit that agentic pairing constitutes the minimal reliable unit of growing intelligence, as paired feedback, augmented with partial-history replay, yields richer learning curricula, tighter preference alignment, and stronger generalization. With extremely few DPO pairs generated automatically by SuperIntelliAgent and used for lightweight fine-tuning, the learner performance improves across all benchmarks.

We present LFM2, a family of Liquid Foundation Models designed for efficient on-device deployment and strong task capabilities. Using hardware-in-the-loop architecture search under edge latency and memory constraints, we obtain a compact hybrid backbone that combines gated short convolutions with a small number of grouped query attention blocks, delivering up to 2x faster prefill and decode on CPUs compared to similarly sized models. The LFM2 family covers 350M-8.3B parameters, including dense models (350M, 700M, 1.2B, 2.6B) and a mixture-of-experts variant (8.3B total, 1.5B active), all with 32K context length. LFM2's training pipeline includes a tempered, decoupled Top-K knowledge distillation objective that avoids support mismatch; curriculum learning with difficulty-ordered data; and a three-stage post-training recipe of supervised fine-tuning, length-normalized preference optimization, and model merging. Pre-trained on 10-12T tokens, LFM2 models achieve strong results across diverse benchmarks; for example, LFM2-2.6B reaches 79.56% on IFEval and 82.41% on GSM8K. We further build multimodal and retrieval variants: LFM2-VL for vision-language tasks, LFM2-Audio for speech, and LFM2-ColBERT for retrieval. LFM2-VL supports tunable accuracy-latency tradeoffs via token-efficient visual processing, while LFM2-Audio separates audio input and output pathways to enable real-time speech-to-speech interaction competitive with models 3x larger. LFM2-ColBERT provides a low-latency encoder for queries and documents, enabling high-performance retrieval across multiple languages. All models are released with open weights and deployment packages for ExecuTorch, llama.cpp, and vLLM, making LFM2 a practical base for edge applications that need fast, memory-efficient inference and strong task capabilities.

This work explores the challenge of building "Machines that Can Remember", framing long-term memory as the problem of efficient ultra-long context modeling. We argue that this requires three key properties: sparsity, random-access flexibility, and length generalization. To address ultra-long-context modeling, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into Transformers to build HSA-UltraLong, which is an 8B-parameter MoE model trained on over 8 trillion tokens and is rigorously evaluated on different tasks with in-domain and out-of-domain context lengths to demonstrate its capability in handling ultra-long contexts. Results show that our model performs comparably to full-attention baselines on in-domain lengths while achieving over 90% accuracy on most in-context retrieval tasks with contexts up to 16M. This report outlines our experimental insights and open problems, contributing a foundation for future research in ultra-long context modeling. Figure 1: Despite being pre-trained with an 8K context window and mid-trained up to 32K, HSA-UltraLong achieves near-perfect accuracy on S-NIAH even at a 16M-token context length. The red dashed line at 32K marks the boundary between in-domain (left) and out-of-domain (right).

Federated edge learning (FEEL) provides a promising foundation for edge artificial intelligence (AI) by enabling collaborative model training while preserving data privacy. However, limited and heterogeneous local datasets, as well as resource-constrained deployment, severely degrade both model generalization and resource utilization, leading to a compromised learning performance. Therefore, we propose a parameter-efficient FEEL framework that jointly leverages model pruning and client selection to tackle such challenges. First, we derive an information-theoretic generalization statement that characterizes the discrepancy between training and testing function losses and embed it into the convergence analysis. It reveals that a larger local generalization statement can undermine the global convergence. Then, we formulate a generalization-aware average squared gradient norm bound minimization problem, by jointly optimizing the pruning ratios, client selection, and communication-computation resources under energy and delay constraints. Despite its non-convexity, the resulting mixed-integer problem is efficiently solved via an alternating optimization algorithm. Extensive experiments demonstrate that the proposed design achieves superior learning performance than state-of-the-art baselines, validating the effectiveness of coupling generalization-aware analysis with system-level optimization for efficient FEEL.

Recent advances in Vision-Language Models (VLMs) have achieved impressive progress in multimodal mathematical reasoning. Yet, how much visual information truly contributes to reasoning remains unclear. Existing benchmarks report strong overall performance but seldom isolate the role of the image modality, leaving open whether VLMs genuinely leverage visual understanding or merely depend on linguistic priors. To address this, we present MathSight, a university-level multimodal mathematical reasoning benchmark designed to disentangle and quantify the effect of visual input. Each problem includes multiple visual variants -- original, hand-drawn, photo-captured -- and a text-only condition for controlled comparison. Experiments on state-of-the-art VLMs reveal a consistent trend: the contribution of visual information diminishes with increasing problem difficulty. Remarkably, Qwen3-VL without any image input surpasses both its multimodal variants and GPT-5, underscoring the need for benchmarks like MathSight to advance genuine vision-grounded reasoning in future models.

Learning transferable latent actions from large-scale object manipulation videos can significantly enhance generalization in downstream robotics tasks, as such representations are agnostic to different robot embodiments. Existing approaches primarily rely on visual reconstruction objectives while neglecting physical priors, leading to sub-optimal performance in learning universal representations. T o address these challenges, we propose a Universal Latent Action Learning framework that takes task instructions and multiple frames as inputs, and optimizes both future frame reconstruction and action sequence prediction. Unlike prior works, incorporating action predictions (e.g., gripper or hand trajectories and orientations) allows the model to capture richer physical priors such as real-world distances and orientations, thereby enabling seamless transferability to downstream tasks. W e further decompose the latent actions into learnable motion and scene tokens to distinguish the robot's active movements from environmental changes, thus filtering out irrelevant dynamics. By distilling the learned latent actions into the latest VLA models, we achieve strong performance across both simulated (SIMPLER and LIBERO) and real-world robot settings. Notably, with only 10 real-world trajectories per task collected on a Franka robot, our approach successfully completes all five challenging tasks, demonstrating strong few-shot transferability in robotic manipulation.

To interact effectively with humans in the real world, it is important for agents to understand language that describes the dynamics of the environment--that is, how the environment behaves--rather than just task instructions specifying "what to do". Understanding this dynamics-descriptive language is important for human-agent interaction and agent behavior. Recent work address this problem using a model-based approach: language is incorporated into a world model, which is then used to learn a behavior policy. However, these existing methods either do not demonstrate policy generalization to unseen games or rely on limiting assumptions. For instance, assuming that the latency induced by inference-time planning is tolerable for the target task or expert demonstrations are available. Expanding on this line of research, we focus on improving policy generalization from a language-conditioned world model while dropping these assumptions. We propose a model-based reinforcement learning approach, where a language-conditioned world model is trained through interaction with the environment, and a policy is learned from this model--without planning or expert demonstrations. Our method proposes Language-aware Encoder for Dreamer World Model (LED-WM) built on top of DreamerV3. LED-WM features an observation encoder that uses an attention mechanism to explicitly ground language descriptions to entities in the observation. We show that policies trained with LED-WM generalize more effectively to unseen games described by novel dynamics and language compared to other baselines in several settings in two environments: MESSENGER and MESSENGER-WM.To highlight how the policy can leverage the trained world model before real-world deployment, we demonstrate the policy can be improved through fine-tuning on synthetic test trajectories generated by the world model.