Large Language Models offer impressive language capabilities but suffer from well-known limitations, including hallucinations, biases, privacy concerns, and high computational costs. These issues are largely driven by the combination of linguistic competence and factual memorization within a single monolithic model. This paper introduces and empirically supports the Fundamental Language Model (FLM) paradigm, which advocates for smaller, linguistically competent models that offload factual retrieval to external tools. We evaluate models ranging from 135M to 32B parameters across three dimensions: linguistic competence, external factual knowledge, and internal factual knowledge. Our findings reveal that while both linguistic competence and factual knowledge improve with scale, internal factual knowledge grows significantly faster, suggesting that model size is more closely tied to memorization than to core language ability. These results support a modular approach to language modeling, where compact, linguistically proficient models serve as the foundation for tool-augmented systems. The FLM paradigm offers a path toward more efficient, interpretable, and sustainable NLP solutions.

Heuristic and scaffolded teacher-student dialogues are widely regarded as critical for fostering students' higher-order thinking and deep learning. However, large language models (LLMs) currently face challenges in generating pedagogically rich interactions. This study systematically investigates the structural and behavioral differences between AI-simulated and authentic human tutoring dialogues. We conducted a quantitative comparison using an Initiation-Response-Feedback (IRF) coding scheme and Epistemic Network Analysis (ENA). The results show that human dialogues are significantly superior to their AI counterparts in utterance length, as well as in questioning (I-Q) and general feedback (F-F) behaviors. More importantly, ENA results reveal a fundamental divergence in interactional patterns: human dialogues are more cognitively guided and diverse, centered around a "question-factual response-feedback" teaching loop that clearly reflects pedagogical guidance and student-driven thinking; in contrast, simulated dialogues exhibit a pattern of structural simplification and behavioral convergence, revolving around an "explanation-simplistic response" loop that is essentially a simple information transfer between the teacher and student. These findings illuminate key limitations in current AI-generated tutoring and provide empirical guidance for designing and evaluating more pedagogically effective generative educational dialogue systems.

Marine ecosystems face increasing pressure due to climate change, driving the need for scalable, AI-powered monitoring solutions. This paper examines the rapid emergence of underwater AI as a major research frontier and analyzes the factors that have transformed marine perception from a niche application into a catalyst for AI innovation. We identify three convergent drivers: environmental necessity for ecosystem-scale monitoring, democratization of underwater datasets through citizen science platforms, and researcher migration from saturated terrestrial computer vision domains. Our analysis reveals how unique underwater challenges - turbidity, cryptic species detection, expert annotation bottlenecks, and cross-ecosystem generalization - are driving fundamental advances in weakly supervised learning, open-set recognition, and robust perception under degraded conditions. We survey emerging trends in datasets, scene understanding and 3D reconstruction, highlighting the paradigm shift from passive observation toward AI-driven, targeted intervention capabilities. The paper demonstrates how underwater constraints are pushing the boundaries of foundation models, self-supervised learning, and perception, with methodological innovations that extend far beyond marine applications to benefit general computer vision, robotics, and environmental monitoring.

Modern deep learning developments create new opportunities for 3D mapping technology, scene reconstruction pipelines, and virtual reality development. Despite advances in 3D deep learning technology, direct training of deep learning models on 3D data faces challenges due to the high dimensionality inherent in 3D data and the scarcity of labeled datasets. Structure-from-motion (SfM) and Simultaneous Localization and Mapping (SLAM) exhibit robust performance when applied to structured indoor environments but often struggle with ambiguous features in unstructured environments. These techniques often struggle to generate detailed geometric representations effective for downstream tasks such as rendering and semantic analysis. Current limitations require the development of 3D representation methods that combine traditional geometric techniques with deep learning capabilities to generate robust geometry-aware deep learning models. The dissertation provides solutions to the fundamental challenges in 3D vision by developing geometric deep learning methods tailored for essential tasks such as camera pose estimation, point cloud registration, depth prediction, and 3D reconstruction. The integration of geometric priors or constraints, such as including depth information, surface normals, and equivariance into deep learning models, enhances both the accuracy and robustness of geometric representations. This study systematically investigates key components of 3D vision, including camera pose estimation, point cloud registration, depth estimation, and high-fidelity 3D reconstruction, demonstrating their effectiveness across real-world applications such as digital cultural heritage preservation and immersive VR/AR environments.

Prompt sensitivity, referring to the phenomenon where paraphrasing (i.e., repeating something written or spoken using different words) leads to significant changes in large language model (LLM) performance, has been widely accepted as a core limitation of LLMs. In this work, we revisit this issue and ask: Is the widely reported high prompt sensitivity truly an inherent weakness of LLMs, or is it largely an artifact of evaluation processes? To answer this question, we systematically evaluate 7 LLMs (e.g., GPT and Gemini family) across 6 benchmarks, including both multiple-choice and open-ended tasks on 12 diverse prompt templates. We find that much of the prompt sensitivity stems from heuristic evaluation methods, including log-likelihood scoring and rigid answer matching, which often overlook semantically correct responses expressed through alternative phrasings, such as synonyms or paraphrases. When we adopt LLM-as-a-Judge evaluations, we observe a substantial reduction in performance variance and a consistently higher correlation in model rankings across prompts. Our findings suggest that modern LLMs are more robust to prompt templates than previously believed, and that prompt sensitivity may be more an artifact of evaluation than a flaw in the models.

Efficient robot control often requires balancing task performance with energy expenditure. A common approach in reinforcement learning (RL) is to penalize energy use directly as part of the reward function. This requires carefully tuning weight terms to avoid undesirable trade-offs where energy minimization harms task success. In this work, we propose a hyperparameter-free gradient optimization method to minimize energy expenditure without conflicting with task performance. Inspired by recent works in multitask learning, our method applies policy gradient projection between task and energy objectives to derive policy updates that minimize energy expenditure in ways that do not impact task performance. We evaluate this technique on standard locomotion benchmarks of DM-Control and HumanoidBench and demonstrate a reduction of 64% energy usage while maintaining comparable task performance. Further, we conduct experiments on a Unitree GO2 quadruped showcasing Sim2Real transfer of energy efficient policies. Our method is easy to implement in standard RL pipelines with minimal code changes, is applicable to any policy gradient method, and offers a principled alternative to reward shaping for energy efficient control policies.

Skill effect models for long-horizon manipulation tasks are prone to failures in conditions not covered by training data distributions. Therefore, enabling robots to reason about and learn from failures is necessary. We investigate the problem of efficiently generating a dataset targeted to observed failures. After fine-tuning a skill effect model on this dataset, we evaluate the extent to which the model can recover from failures and minimize future failures. We propose Fail2Progress, an approach that leverages Stein variational inference to generate multiple simulation environments in parallel, enabling efficient data sample generation similar to observed failures. Our method is capable of handling several challenging mobile manipulation tasks, including transporting multiple objects, organizing a constrained shelf, and tabletop organization. Through large-scale simulation and real-world experiments, we demonstrate that our approach excels at learning from failures across different numbers of objects. Furthermore, we show that Fail2Progress outperforms several baselines.

Biases in machine learning pose significant challenges, particularly when models amplify disparities that affect disadvantaged groups. Traditional bias mitigation techniques often lead to a {\itshape leveling-down effect}, whereby improving outcomes of disadvantaged groups comes at the expense of reduced performance for advantaged groups. This study introduces Bias Mitigation through Continual Learning (BM-CL), a novel framework that leverages the principles of continual learning to address this trade-off. We postulate that mitigating bias is conceptually similar to domain-incremental continual learning, where the model must adjust to changing fairness conditions, improving outcomes for disadvantaged groups without forgetting the knowledge that benefits advantaged groups. Drawing inspiration from techniques such as Learning without Forgetting and Elastic Weight Consolidation, we reinterpret bias mitigation as a continual learning problem. This perspective allows models to incrementally balance fairness objectives, enhancing outcomes for disadvantaged groups while preserving performance for advantaged groups. Experiments on synthetic and real-world image datasets, characterized by diverse sources of bias, demonstrate that the proposed framework mitigates biases while minimizing the loss of original knowledge. Our approach bridges the fields of fairness and continual learning, offering a promising pathway for developing machine learning systems that are both equitable and effective.

In the past year, distillation has seen a renewed prominence in large language model (LLM) pretraining, exemplified by the Llama-3.2 and Gemma model families. While distillation has historically been shown to improve statistical modeling, its effects on new paradigms that are key to modern LLMs, such as test-time scaling and in-context learning, remain underexplored. In this work, we make three main contributions. First, we show that pretraining with distillation yields models that exhibit remarkably better test-time scaling. Second, we observe that this benefit comes with a trade-off: distillation impairs in-context learning capabilities, particularly the one modeled via induction heads. Third, to demystify these findings, we study distilled pretraining in a sandbox of a bigram model, which helps us isolate the common principal factor behind our observations. Finally, using these insights, we shed light on various design choices for pretraining that should help practitioners going forward.

Task load detection is essential for optimizing human performance across diverse applications, yet current models often lack generalizability beyond narrow experimental domains. While prior research has focused on individual tasks and limited modalities, there remains a gap in evaluating model robustness and transferability in real-world scenarios. This paper addresses these limitations by introducing a new multimodal dataset that extends established cognitive load detection benchmarks with a real-world gaming application, using the $n$-back test as a scientific foundation. Task load annotations are derived from objective performance, subjective NASA-TLX ratings, and task-level design, enabling a comprehensive evaluation framework. State-of-the-art end-to-end model, including xLSTM, ConvNeXt, and Transformer architectures are systematically trained and evaluated on multiple modalities and application domains to assess their predictive performance and cross-domain generalization. Results demonstrate that multimodal approaches consistently outperform unimodal baselines, with specific modalities and model architectures showing varying impact depending on the application subset. Importantly, models trained on one domain exhibit reduced performance when transferred to novel applications, underscoring remaining challenges for universal cognitive load estimation. These findings provide robust baselines and actionable insights for developing more generalizable cognitive load detection systems, advancing both research and practical implementation in human-computer interaction and adaptive systems.