Large language models (LLMs) have revolutionized natural language understanding and generation but face significant memory bottlenecks during training. GaLore, Gradient Low-Rank Projection, addresses this issue by leveraging the inherent low-rank structure of weight gradients, enabling substantial memory savings without sacrificing performance. Recent works further extend GaLore from various aspects, including low-bit quantization and higher-order tensor structures. However, there are several remaining challenges for GaLore, such as the computational overhead of SVD for subspace updates and the integration with state-of-the-art training parallelization strategies (e.g., FSDP). In this paper, we present GaLore 2, an efficient and scalable GaLore framework that addresses these challenges and incorporates recent advancements. In addition, we demonstrate the scalability of GaLore 2 by pre-training Llama 7B from scratch using up to 500 billion training tokens, highlighting its potential impact on real LLM pre-training scenarios.

Cross-lingual transfer in natural language processing (NLP) models enhances multilingual performance by leveraging shared linguistic knowledge. However, traditional methods that process all data simultaneously often fail to mimic real-world scenarios, leading to challenges like catastrophic forgetting, where fine-tuning on new tasks degrades performance on previously learned ones. Our study explores this issue in multilingual contexts, focusing on linguistic differences affecting representational learning rather than just model parameters. We experiment with 52 languages using LoRA adapters of varying ranks to evaluate non-shared, partially shared, and fully shared parameters. Our aim is to see if parameter sharing through adapters can mitigate forgetting while preserving prior knowledge. We find that languages using non-Latin scripts are more susceptible to catastrophic forgetting, whereas those written in Latin script facilitate more effective cross-lingual transfer.

Reflexion is an AI-powered platform designed to enable structured emotional self-reflection at scale. By integrating real-time emotion detection, layered reflective prompting, and metaphorical storytelling generation, Reflexion empowers users to engage in autonomous emotional exploration beyond basic sentiment categorization. Grounded in theories of expressive writing, cognitive restructuring, self-determination, and critical consciousness development, the system scaffolds a progressive journey from surface-level emotional recognition toward value-aligned action planning. Initial pilot studies with diverse participants demonstrate positive outcomes in emotional articulation, cognitive reframing, and perceived psychological resilience. Reflexion represents a promising direction for scalable, theory-informed affective computing interventions aimed at fostering emotional literacy and psychological growth across educational, therapeutic, and public health contexts.

We present a platform for the generation of educational activities oriented to teaching English as a foreign language. The different activities -- games and language practice exercises -- are strongly based on Natural Language Processing techniques. The platform offers the possibility of playing out-of-the-box games, generated from resources created semi-automatically and then manually curated. It can also generate games or exercises of greater complexity from texts entered by teachers, providing a stage of review and edition of the generated content before use. As a way of expanding the variety of activities in the platform, we are currently experimenting with image and text generation. In order to integrate them and improve the performance of other neural tools already integrated, we are working on migrating the platform to a more powerful server. In this paper we describe the development of our platform and its deployment for end users, discussing the challenges faced and how we overcame them, and also detail our future work plans.

Artificial Intelligence has made remarkable advancements in recent years, primarily driven by increasingly large deep learning models. However, achieving true Artificial General Intelligence (AGI) demands fundamentally new architectures rather than merely scaling up existing models. Current approaches largely depend on expanding model parameters, which improves task-specific performance but falls short in enabling continuous, adaptable, and generalized learning. Achieving AGI capable of continuous learning and personalization on resource-constrained edge devices is an even bigger challenge. This paper reviews the state of continual learning and neuroscience-inspired AI, and proposes a novel architecture for Personalized AGI that integrates brain-like learning mechanisms for edge deployment. We review literature on continuous lifelong learning, catastrophic forgetting, and edge AI, and discuss key neuroscience principles of human learning, including Synaptic Pruning, Hebbian plasticity, Sparse Coding, and Dual Memory Systems, as inspirations for AI systems. Building on these insights, we outline an AI architecture that features complementary fast-and-slow learning modules, synaptic self-optimization, and memory-efficient model updates to support on-device lifelong adaptation. Conceptual diagrams of the proposed architecture and learning processes are provided. We address challenges such as catastrophic forgetting, memory efficiency, and system scalability, and present application scenarios for mobile AI assistants and embodied AI systems like humanoid robots. We conclude with key takeaways and future research directions toward truly continual, personalized AGI on the edge. While the architecture is theoretical, it synthesizes diverse findings and offers a roadmap for future implementation.

Personalized food recommendation systems (Food-RecSys) critically underperform due to fragmented component understanding and the failure of conventional machine learning with vast, imbalanced food data. While Large Language Models (LLMs) offer promise, current generic Recommendation as Language Processing (RLP) strategies lack the necessary specialization for the food domain's complexity. This thesis tackles these deficiencies by first identifying and analyzing the essential components for effective Food-RecSys. We introduce two key innovations: a multimedia food logging platform for rich contextual data acquisition and the World Food Atlas, enabling unique geolocation-based food analysis previously unavailable. Building on this foundation, we pioneer the Food Recommendation as Language Processing (F-RLP) framework - a novel, integrated approach specifically architected for the food domain. F-RLP leverages LLMs in a tailored manner, overcoming the limitations of generic models and providing a robust infrastructure for effective, contextual, and truly personalized food recommendations.

To meet the demand for designing efficient neural networks with appropriate trade-offs between model performance (e.g., classification accuracy) and computational complexity, the differentiable neural architecture distillation (DNAD) algorithm is developed based on two cores, namely search by deleting and search by imitating. Primarily, to derive neural architectures in a space where cells of the same type no longer share the same topology, the super-network progressive shrinking (SNPS) algorithm is developed based on the framework of differentiable architecture search (DARTS), i.e., search by deleting. Unlike conventional DARTS-based approaches which yield neural architectures with simple structures and derive only one architecture during the search procedure, SNPS is able to derive a Pareto-optimal set of architectures with flexible structures by forcing the dynamic super-network shrink from a dense structure to a sparse one progressively. Furthermore, since knowledge distillation (KD) has shown great effectiveness to train a compact network with the assistance of an over-parameterized model, we integrate SNPS with KD to formulate the DNAD algorithm, i.e., search by imitating. By minimizing behavioral differences between the super-network and teacher network, the over-fitting of one-level DARTS is avoided and well-performed neural architectures are derived. Experiments on CIFAR-10 and ImageNet classification tasks demonstrate that both SNPS and DNAD are able to derive a set of architectures which achieve similar or lower error rates with fewer parameters and FLOPs. Particularly, DNAD achieves the top-1 error rate of 23.7% on ImageNet classification with a model of 6.0M parameters and 598M FLOPs, which outperforms most DARTS-based methods.

--Strong priors are imposed on the search space of Differentiable Architecture Search (DARTS), such that cells of the same type share the same topological structure and each intermediate node retains two operators from distinct nodes. While these priors reduce optimization difficulties and improve the applicability of searched architectures, they hinder the subsequent development of automated machine learning (Auto-ML) and prevent the optimization algorithm from exploring more powerful neural networks through improved architectural flexibility. This paper aims to reduce these prior constraints by eliminating restrictions on cell topology and modifying the dis-cretization mechanism for super-networks. Specifically, the Flexible DARTS (FX-DARTS) method, which leverages an Entropy-based Super-Network Shrinking (ESS) framework, is presented to address the challenges arising from the elimination of prior constraints. Notably, FX-DARTS enables the derivation of neural architectures without strict prior rules while maintaining the stability in the enlarged search space. Experimental results on image classification benchmarks demonstrate that FX-DARTS is capable of exploring a set of neural architectures with competitive trade-offs between performance and computational complexity within a single search procedure. Derong Liu is with the School of System Design and Intelligent Manufacturing, Southern University of Science and Technology, Shenzhen 518000, China (email: liudr@sustech.edu.cn), and also with the Department of Electrical and Computer Engineering, University of Illinois Chicago, Chicago, IL 60607, USA (e-mail: derong@uic.edu). This manuscript is submitted to IEEE Transaction on Neural Network and Learning Systems and is under reviewed. Personal use of this manuscript is permitted. Over the past decade, the powerful representation ability of deep neural networks (DNNs) has contributed to significant progress in various machine learning tasks, including computer vision [1], [2], natural language processing [3], [4], system identification and control [5], [6], time series prediction [7], and autonomous vehicles [8]-[10], among others. Undoubtedly, the architecture design of neural networks plays a pivotal role in these breakthroughs [11]-[15]. To address the labor-intensive and time-consuming trial-and-error process of DNN architecture design, neural architecture search (NAS) [16] has emerged as a promising approach. NAS automates the exploration of a vast space of potential architectures, traditionally through three key steps: defining a search space, selecting a search algorithm, and identifying an optimal architecture within the search space. The effectiveness of NAS heavily relies on the careful design of both the search space and the search strategy, as a well-constructed search space can significantly enhance the search algorithm's ability to discover optimal neural architectures [17].

Deep Knowledge Tracing (DKT) models student learning behavior by using Recurrent Neural Networks (RNNs) to predict future performance based on historical interaction data. However, the original implementation relied on standard RNNs in the Lua-based Torch framework, which limited extensibility and reproducibility. In this work, we revisit the DKT model from two perspectives: architectural improvements and optimization efficiency. First, we enhance the model using gated recurrent units, specifically Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRU), which better capture long-term dependencies and help mitigate vanishing gradient issues. Second, we re-implement DKT using the PyTorch framework, enabling a modular and accessible infrastructure compatible with modern deep learning workflows. We also benchmark several optimization algorithms SGD, RMSProp, Adagrad, Adam, and AdamW to evaluate their impact on convergence speed and predictive accuracy in educational modeling tasks. Experiments on the Synthetic-5 and Khan Academy datasets show that GRUs and LSTMs achieve higher accuracy and improved training stability compared to basic RNNs, while adaptive optimizers such as Adam and AdamW consistently outperform SGD in both early-stage learning and final model performance. Our open-source PyTorch implementation provides a reproducible and extensible foundation for future research in neural knowledge tracing and personalized learning systems.

Educational question generation (EQG) is a crucial component of intelligent educational systems, significantly aiding self-assessment, active learning, and personalized education. While EQG systems have emerged, existing datasets typically rely on predefined, carefully edited texts, failing to represent real-world classroom content, including lecture speech with a set of complementary slides. To bridge this gap, we collect a dataset of educational questions based on lectures from real-world classrooms. On this realistic dataset, we find that current methods for EQG struggle with accurately generating questions from educational videos, particularly in aligning with specific timestamps and target answers. Common challenges include selecting informative contexts from extensive transcripts and ensuring generated questions meaningfully incorporate the target answer. To address the challenges, we introduce a novel framework utilizing large language models for dynamically selecting and rewriting contexts based on target timestamps and answers. First, our framework selects contexts from both lecture transcripts and video keyframes based on answer relevance and temporal proximity. Then, we integrate the contexts selected from both modalities and rewrite them into answer-containing knowledge statements, to enhance the logical connection between the contexts and the desired answer. This approach significantly improves the quality and relevance of the generated questions. Our dataset and code are released in https://github.com/mengxiayu/COSER.