Tokenization disparities pose a significant barrier to achieving equitable access to artificial intelligence across linguistically diverse populations. This study conducts a large-scale cross-linguistic evaluation of tokenization efficiency in over 200 languages to systematically quantify computational inequities in large language models (LLMs). Using a standardized experimental framework, we applied consistent preprocessing and normalization protocols, followed by uniform tokenization through the tiktoken library across all language samples. Comprehensive tokenization statistics were collected using established evaluation metrics, including Tokens Per Sentence (TPS) and Relative Tokenization Cost (RTC), benchmarked against English baselines. Our cross-linguistic analysis reveals substantial and systematic disparities: Latin-script languages consistently exhibit higher tokenization efficiency, while non-Latin and morphologically complex languages incur significantly greater token inflation, often 3-5 times higher RTC ratios. These inefficiencies translate into increased computational costs and reduced effective context utilization for underrepresented languages. Overall, the findings highlight structural inequities in current AI systems, where speakers of low-resource and non-Latin languages face disproportionate computational disadvantages. Future research should prioritize the development of linguistically informed tokenization strategies and adaptive vocabulary construction methods that incorporate typological diversity, ensuring more inclusive and computationally equitable multilingual AI systems.

Although deep neural networks have provided impressive gains in performance, these improvements often come at the cost of increased computational complexity and expense. In many cases, such as 3D volume or video classification tasks, not all slices or frames are necessary due to inherent redundancies. To address this issue, we propose a novel learnable subsampling framework that can be integrated into any neural network architecture. Subsampling, being a nondifferentiable operation, poses significant challenges for direct adaptation into deep learning models. While some works, have proposed solutions using the Gumbel-max trick to overcome the problem of non-differentiability, they fall short in a crucial aspect: they are only task-adaptive and not inputadaptive. Once the sampling mechanism is learned, it remains static and does not adjust to different inputs, making it unsuitable for real-world applications. To this end, we propose an attention-guided sampling module that adapts to inputs even during inference. This dynamic adaptation results in performance gains and reduces complexity in deep neural network models. We demonstrate the effectiveness of our method on 3D medical imaging datasets from MedMNIST3D as well as two ultrasound video datasets for classification tasks, one of them being a challenging in-house dataset collected under real-world clinical conditions.

Shallow Recurrent Decoder networks are a novel data-driven methodology able to provide accurate state estimation in engineering systems, such as nuclear reactors. This deep learning architecture is a robust technique designed to map the temporal trajectories of a few sparse measures to the full state space, including unobservable fields, which is agnostic to sensor positions and able to handle noisy data through an ensemble strategy, leveraging the short training times and without the need for hyperparameter tuning. Following its application to a novel reactor concept, this work investigates the performance of Shallow Recurrent Decoders when applied to a real system. The underlying model is represented by a fluid dynamics model of the TRIGA Mark II research reactor; the architecture will use both synthetic temperature data coming from the numerical model and leveraging experimental temperature data recorded during a previous campaign. The objective of this work is, therefore, two-fold: 1) assessing if the architecture can reconstruct the full state of the system (temperature, velocity, pressure, turbulence quantities) given sparse data located in specific, low-dynamics channels and 2) assessing the correction capabilities of the architecture (that is, given a discrepancy between model and data, assessing if sparse measurements can provide some correction to the architecture output). As will be shown, the accurate reconstruction of every characteristic field, using both synthetic and experimental data, in real-time makes this approach suitable for interpretable monitoring and control purposes in the framework of a reactor digital twin.

Understanding the alignment between large language models (LLMs) and human brain activity can reveal computational principles underlying language processing. We introduce a fine-grained input attribution method to identify the specific words most important for brain-LLM alignment, and leverage it to study a contentious research question about brain-LLM alignment: the relationship between brain alignment (BA) and next-word prediction (NWP). Our findings reveal that BA and NWP rely on largely distinct word subsets: NWP exhibits recency and primacy biases with a focus on syntax, while BA prioritizes semantic and discourse-level information with a more targeted recency effect. This work advances our understanding of how LLMs relate to human language processing and highlights differences in feature reliance between BA and NWP . Beyond this study, our attribution method can be broadly applied to explore the cognitive relevance of model predictions in diverse language processing tasks.

As users increasingly interact with large language models (LLMs) using private information, secure and encrypted communication becomes essential. Homomorphic encryption (HE) provides a principled solution by enabling computation directly on encrypted data. Although prior work has explored aspects of running LLMs under HE, the challenge of text generation, particularly next-token prediction, has received limited attention and remains a key obstacle to practical encrypted interaction. In this work, we propose a TSP-based token reordering strategy to address the difficulties of encrypted text generation, together with a post-processing step that further reduces approximation error. Theoretical analysis and experimental results demonstrate that our method prevents collapse, improves coherence in generated text, and preserves data privacy throughout. Overall, our contributions advance the feasibility of practical and privacy-preserving LLM inference.

Multi-vector dense retrieval methods like ColBERT systematically use a single-layer linear projection to reduce the dimensionality of individual vectors. In this study, we explore the implications of the MaxSim operator on the gradient flows of the training of multi-vector models and show that such a simple linear projection has inherent, if non-critical, limitations in this setting. We then discuss the theoretical improvements that could result from replacing this single-layer projection with well-studied alternative feedforward linear networks (FFN), such as deeper, non-linear FFN blocks, GLU blocks, and skip-connections, could alleviate these limitations. Through the design and systematic evaluation of alternate projection blocks, we show that better-designed final projections positively impact the downstream performance of ColBERT models. We highlight that many projection variants outperform the original linear projections, with the best-performing variants increasing average performance on a range of retrieval benchmarks across domains by over 2 NDCG@10 points. We then conduct further exploration on the individual parameters of these projections block in order to understand what drives this empirical performance, highlighting the particular importance of upscaled intermediate projections and residual connections. As part of these ablation studies, we show that numerous suboptimal projection variants still outperform the traditional single-layer projection across multiple benchmarks, confirming our hypothesis. Finally, we observe that this effect is consistent across random seeds, further confirming that replacing the linear layer of ColBERT models is a robust, drop-in upgrade.

Retrieval-Augmented Generation (RAG) has emerged as a fundamental paradigm for expanding Large Language Models beyond their static training limitations. However, a critical misalignment exists between current RAG capabilities and real-world information environments. Modern knowledge repositories are inherently multimodal, containing rich combinations of textual content, visual elements, structured tables, and mathematical expressions. Yet existing RAG frameworks are limited to textual content, creating fundamental gaps when processing multimodal documents. We present RAG-Anything, a unified framework that enables comprehensive knowledge retrieval across all modalities. Our approach reconceptualizes multimodal content as interconnected knowledge entities rather than isolated data types. The framework introduces dual-graph construction to capture both cross-modal relationships and textual semantics within a unified representation. We develop cross-modal hybrid retrieval that combines structural knowledge navigation with semantic matching. This enables effective reasoning over heterogeneous content where relevant evidence spans multiple modalities. RAG-Anything demonstrates superior performance on challenging multimodal benchmarks, achieving significant improvements over state-of-the-art methods. Performance gains become particularly pronounced on long documents where traditional approaches fail. Our framework establishes a new paradigm for multimodal knowledge access, eliminating the architectural fragmentation that constrains current systems. Our framework is open-sourced at: https://github.com/HKUDS/RAG-Anything.

Counter-speech generation is at the core of many expert activities, such as fact-checking and hate speech, to counter harmful content. Yet, existing work treats counter-speech generation as pure text generation task, mainly based on Large Language Models or NGO experts. These approaches show severe drawbacks due to the limited reliability and coherence in the generated countering text, and in scalability, respectively. To close this gap, we introduce a novel framework to model counter-speech generation as knowledge-wise text generation process. Our framework integrates advanced Retrieval-Augmented Generation (RAG) pipelines to ensure the generation of trustworthy counter-speech for 8 main target groups identified in the hate speech literature, including women, people of colour, persons with disabilities, migrants, Muslims, Jews, LGBT persons, and other. We built a knowledge base over the United Nations Digital Library, EUR-Lex and the EU Agency for Fundamental Rights, comprising a total of 32,792 texts. We use the MultiTarget-CONAN dataset to empirically assess the quality of the generated counter-speech, both through standard metrics (i.e., JudgeLM) and a human evaluation. Results show that our framework outperforms standard LLM baselines and competitive approach, on both assessments. The resulting framework and the knowledge base pave the way for studying trustworthy and sound counter-speech generation, in hate speech and beyond.

As natural language corpora expand at an unprecedented rate, manual annotation remains a significant methodological bottleneck in corpus linguistic work. We address this challenge by presenting a scalable, unsupervised pipeline for automating grammatical annotation in voluminous corpora using large language models (LLMs). Unlike previous supervised and iterative approaches, our method employs a four-phase workflow: prompt engineering, pre-hoc evaluation, automated batch processing, and post-hoc validation. We demonstrate the pipeline's accessibility and effectiveness through a diachronic case study of variation in the English consider construction. Using GPT-5 through the OpenAI API, we annotate 143,933 sentences from the Corpus of Historical American English (COHA) in under 60 hours, achieving 98%+ accuracy on two sophisticated annotation procedures. Our results suggest that LLMs can perform a range of data preparation tasks at scale with minimal human intervention, opening new possibilities for corpus-based research, though implementation requires attention to costs, licensing, and other ethical considerations.

Self-improvement has emerged as a state-of-the-art paradigm in Neural Combinatorial Optimization (NCO), where models iteratively refine their policies by generating and imitating high-quality solutions. Despite strong empirical performance, existing methods face key limitations. Training is computationally expensive, as policy updates require sampling numerous candidate solutions per instance to extract a single expert trajectory. More fundamentally, these approaches fail to exploit the structure of combinatorial problems involving the coordination of multiple agents, such as vehicles in min-max routing or machines in scheduling. By supervising on single-action trajectories, they fail to exploit agent-permutation symmetries, where distinct sequences of actions yield identical solutions, hindering generalization and the ability to learn coordinated behavior. We address these challenges by extending self-improvement to operate over joint multi-agent actions. Our model architecture predicts complete agent-task assignments jointly at each decision step. To explicitly leverage symmetries, we employ a set-prediction loss, which supervises the policy on multiple expert assignments for any given state. This approach enhances sample efficiency and the model's ability to learn coordinated behavior. Furthermore, by generating multi-agent actions in parallel, it drastically accelerates the solution generation phase of the self-improvement loop. Empirically, we validate our method on several combinatorial problems, demonstrating consistent improvements in the quality of the final solution and a reduced generation latency compared to standard self-improvement.