Abstract
Large Language Models (LLMs) have delivered impressive results in language understanding, generation, reasoning, and pushes the ability boundary of multimodal models. Transformer models, as the foundation of modern LLMs, offer a strong baseline with excellent scaling properties. However, the traditional transformer architecture requires substantial computations and poses significant obstacles for large-scale training and practical deployment. In this survey, we offer a systematic examination of innovative LLM architectures that address the inherent limitations of transformers and boost the efficiency. Starting from language modeling, this survey covers the background and technical details of linear and sparse sequence modeling methods, efficient full attention variants, sparse mixture-of-experts, hybrid model architectures incorporating the above techniques, and emerging diffusion LLMs. Additionally, we discuss applications of these techniques to other modalities and consider their wider implications for developing scalable, resource-aware foundation models. By grouping recent studies into the above category, this survey presents a blueprint of modern efficient LLM architectures, and we hope this could help motivate future research toward more efficient, versatile AI systems.

Abstract
Deriving governing equations from observational data, known as Symbolic Regression (SR), is a cornerstone of scientific discovery. Large Language Models (LLMs) have shown promise in this task by leveraging their vast cross-disciplinary scientific knowledge. However, existing LLM-based methods primarily rely on direct inference or prompt engineering, often requiring excessive inference iterations to converge on correct formulas or failing to treating complex equation targets. These limitations in effectiveness and generalization stem from an inherent tension between pre-trained LLMs' proficiency in approximate reasoning and the high-precision demands of SR tasks. To bridge this gap, we propose to fine-tune LLMs for enhanced SR capability. Yet, the absence of dedicated datasets for SR-oriented fine-tuning remains a critical barrier. We thus introduce SymbArena, specifically engineered to optimize LLMs for SR. This benchmark comprises 148,102 diverse equations formulated as corpora of 1.83 billion tokens for LLM utilization, enabling effective training and inference. Further, SymbArena proposes a heuristics metric to precisely quantify form-level consistency, going beyond existing SR numerical-oriented evaluation strategies. With this benchmark, we explore mainstream LLM fine-tuning techniques for SR tasks and establish SymbolicChat, a simple yet effective LLM-based SR strong baseline. Experimental results validate SymbolicChat as the first LLM to exceed traditional numerical methods in both numerical precision and symbolic form accuracy, outperforming the second-best LLM baseline with improvements of 2-fold gains in R2 score and 8.37% in form-level consistency score.

Abstract
Deep learning models often struggle to maintain generalizability in medical imaging, particularly under domain-fracture scenarios where distribution shifts arise from varying imaging techniques, acquisition protocols, patient populations, demographics, and equipment. In practice, each hospital may need to train distinct models - differing in learning task, width, and depth - to match local data. For example, one hospital may use Euclidean architectures such as MLPs and CNNs for tabular or grid-like image data, while another may require non-Euclidean architectures such as graph neural networks (GNNs) for irregular data like brain connectomes. How to train such heterogeneous models coherently across datasets, while enhancing each model's generalizability, remains an open problem. We propose unified learning, a new paradigm that encodes each model into a graph representation, enabling unification in a shared graph learning space. A GNN then guides optimization of these unified models. By decoupling parameters of individual models and controlling them through a unified GNN (uGNN), our method supports parameter sharing and knowledge transfer across varying architectures (MLPs, CNNs, GNNs) and distributions, improving generalizability. Evaluations on MorphoMNIST and two MedMNIST benchmarks - PneumoniaMNIST and BreastMNIST - show that unified learning boosts performance when models are trained on unique distributions and tested on mixed ones, demonstrating strong robustness to unseen data with large distribution shifts. Code and benchmarks: https://github.com/basiralab/uGNN

Abstract
Scientific progress is tightly coupled to the emergence of new research tools. Today, machine learning (ML)-especially deep learning (DL)-has become a transformative instrument for quantum science and technology. Owing to the intrinsic complexity of quantum systems, DL enables efficient exploration of large parameter spaces, extraction of patterns from experimental data, and data-driven guidance for research directions. These capabilities already support tasks such as refining quantum control protocols and accelerating the discovery of materials with targeted quantum properties, making ML/DL literacy an essential skill for the next generation of quantum scientists. At the same time, DL's power brings risks: models can overfit noisy data, obscure causal structure, and yield results with limited physical interpretability. Recognizing these limitations and deploying mitigation strategies is crucial for scientific rigor. These lecture notes provide a comprehensive, graduate-level introduction to DL for quantum applications, combining conceptual exposition with hands-on examples. Organized as a progressive sequence, they aim to equip readers to decide when and how to apply DL effectively, to understand its practical constraints, and to adapt AI methods responsibly to problems across quantum physics, chemistry, and engineering.

Abstract
Despite significant progress, existing research on Multimodal Large Language Models (MLLMs) mainly focuses on general visual understanding, overlooking the ability to integrate textual context associated with objects for a more context-aware multimodal understanding -- an ability we refer to as Region-level Context-aware Multimodal Understanding (RCMU). To address this limitation, we first formulate the RCMU task, which requires models to respond to user instructions by integrating both image content and textual information of regions or objects. To equip MLLMs with RCMU capabilities, we propose Region-level Context-aware Visual Instruction Tuning (RCVIT), which incorporates object information into the model input and enables the model to utilize bounding box coordinates to effectively associate objects' visual content with their textual information. To address the lack of datasets, we introduce the RCMU dataset, a large-scale visual instruction tuning dataset that covers multiple RCMU tasks. We also propose RC\&P-Bench, a comprehensive benchmark that can evaluate the performance of MLLMs in RCMU and multimodal personalized understanding tasks. Additionally, we propose a reference-free evaluation metric to perform a comprehensive and fine-grained evaluation of the region-level context-aware image descriptions. By performing RCVIT on Qwen2-VL models with the RCMU dataset, we developed RC-Qwen2-VL models. Experimental results indicate that RC-Qwen2-VL models not only achieve outstanding performance on multiple RCMU tasks but also demonstrate successful applications in multimodal RAG and personalized conversation. Our data, model and benchmark are available at https://github.com/hongliang-wei/RC-MLLM

Abstract
As AI-generated content becomes widespread, so does the risk of misinformation. While prior research has primarily focused on identifying whether content is authentic, much less is known about how such content influences human perception and behavior. In domains like trading or the stock market, predicting how people react (e.g., whether a news post will go viral), can be more critical than verifying its factual accuracy. To address this, we take a human-centered approach and introduce the MhAIM Dataset, which contains 154,552 online posts (111,153 of them AI-generated), enabling large-scale analysis of how people respond to AI-generated content. Our human study reveals that people are better at identifying AI content when posts include both text and visuals, particularly when inconsistencies exist between the two. We propose three new metrics: trustworthiness, impact, and openness, to quantify how users judge and engage with online content. We present T-Lens, an LLM-based agent system designed to answer user queries by incorporating predicted human responses to multimodal information. At its core is HR-MCP (Human Response Model Context Protocol), built on the standardized Model Context Protocol (MCP), enabling seamless integration with any LLM. This integration allows T-Lens to better align with human reactions, enhancing both interpretability and interaction capabilities. Our work provides empirical insights and practical tools to equip LLMs with human-awareness capabilities. By highlighting the complex interplay among AI, human cognition, and information reception, our findings suggest actionable strategies for mitigating the risks of AI-driven misinformation.

Abstract
Diffusion models have achieved remarkable success in generative modeling. However, this study confirms the existence of overfitting in diffusion model training, particularly in data-limited regimes. To address this challenge, we propose Score Augmentation (ScoreAug), a novel data augmentation framework specifically designed for diffusion models. Unlike conventional augmentation approaches that operate on clean data, ScoreAug applies transformations to noisy data, aligning with the inherent denoising mechanism of diffusion. Crucially, ScoreAug further requires the denoiser to predict the augmentation of the original target. This design establishes an equivariant learning objective, enabling the denoiser to learn scores across varied denoising spaces, thereby realizing what we term score augmentation. We also theoretically analyze the relationship between scores in different spaces under general transformations. In experiments, we extensively validate ScoreAug on multiple benchmarks including CIFAR-10, FFHQ, AFHQv2, and ImageNet, with results demonstrating significant performance improvements over baselines. Notably, ScoreAug effectively mitigates overfitting across diverse scenarios, such as varying data scales and model capacities, while exhibiting stable convergence properties. Another advantage of ScoreAug over standard data augmentation lies in its ability to circumvent data leakage issues under certain conditions. Furthermore, we show that ScoreAug can be synergistically combined with traditional data augmentation techniques to achieve additional performance gains.

Abstract
We introduce measurement-based quantum diffusion models that bridge classical and quantum diffusion theory through randomized weak measurements. The measurement-based approach naturally generates stochastic quantum trajectories while preserving purity at the trajectory level and inducing depolarization at the ensemble level. We address two quantum state generation problems: trajectory-level recovery of pure state ensembles and ensemble-average recovery of mixed states. For trajectory-level recovery, we establish that quantum score matching is mathematically equivalent to learning unitary generators for the reverse process. For ensemble-average recovery, we introduce local Petz recovery maps for states with finite correlation length and classical shadow reconstruction for general states, both with rigorous error bounds. Our framework establishes Petz recovery maps as quantum generalizations of reverse Fokker-Planck equations, providing a rigorous bridge between quantum recovery channels and classical stochastic reversals. This work enables new approaches to quantum state generation with potential applications in quantum information science.

Abstract
We introduce an output layer for neural networks that ensures satisfaction of convex constraints. Our approach, $\Pi$net, leverages operator splitting for rapid and reliable projections in the forward pass, and the implicit function theorem for backpropagation. We deploy $\Pi$net as a feasible-by-design optimization proxy for parametric constrained optimization problems and obtain modest-accuracy solutions faster than traditional solvers when solving a single problem, and significantly faster for a batch of problems. We surpass state-of-the-art learning approaches in terms of training time, solution quality, and robustness to hyperparameter tuning, while maintaining similar inference times. Finally, we tackle multi-vehicle motion planning with non-convex trajectory preferences and provide $\Pi$net as a GPU-ready package implemented in JAX with effective tuning heuristics.

Abstract
Recent years have seen a growing interest and adoption of LLMs, with $\mu$Transfer becoming a key technique for tuning hyperparameters in large-scale training. Meanwhile, Mixture-of-Experts (MoE) has emerged as a leading architecture in extremely large models. However, the intersection of these two advancements has remained unexplored. In this work, we derive a $\mu$-Parameterization ($\mu$P) for MoE, providing theoretical guarantees for feature learning across model widths in both the router and experts. We empirically validate our parameterization and further investigate how scaling the number of experts and granularity affects the optimal learning rate.

Abstract
Graph incremental learning is a learning paradigm that aims to adapt trained models to continuously incremented graphs and data over time without the need for retraining on the full dataset. However, regular graph machine learning methods suffer from catastrophic forgetting when applied to incremental learning settings, where previously learned knowledge is overridden by new knowledge. Previous approaches have tried to address this by treating the previously trained model as an inseparable unit and using techniques to maintain old behaviors while learning new knowledge. These approaches, however, do not account for the fact that previously acquired knowledge at different timestamps contributes differently to learning new tasks. Some prior patterns can be transferred to help learn new data, while others may deviate from the new data distribution and be detrimental. To address this, we propose a dynamic mixture-of-experts (DyMoE) approach for incremental learning. Specifically, a DyMoE GNN layer adds new expert networks specialized in modeling the incoming data blocks. We design a customized regularization loss that utilizes data sequence information so existing experts can maintain their ability to solve old tasks while helping the new expert learn the new data effectively. As the number of data blocks grows over time, the computational cost of the full mixture-of-experts (MoE) model increases. To address this, we introduce a sparse MoE approach, where only the top-$k$ most relevant experts make predictions, significantly reducing the computation time. Our model achieved 4.92\% relative accuracy increase compared to the best baselines on class incremental learning, showing the model's exceptional power.