Large language models (LLMs) have demonstrated remarkable success across diverse artificial intelligence tasks, driven by scaling laws that correlate model size and training data with performance improvements. However, this scaling paradigm incurs substantial memory overhead, creating significant challenges for both training and inference. While existing research has primarily addressed parameter and optimizer state memory reduction, activation memory-particularly from feed-forward networks (FFNs)-has become the critical bottleneck, especially when FlashAttention is implemented. In this work, we conduct a detailed memory profiling of LLMs and identify FFN activations as the predominant source to activation memory overhead. Motivated by this, we introduce Mixture-of-Channels (MoC), a novel FFN architecture that selectively activates only the Top-K most relevant channels per token determined by SwiGLU's native gating mechanism. MoC substantially reduces activation memory during pre-training and improves inference efficiency by reducing memory access through partial weight loading into GPU SRAM. Extensive experiments validate that MoC delivers significant memory savings and throughput gains while maintaining competitive model performance.

It is challenging to reduce the complexity of neural networks while maintaining their generalization ability and robustness, especially for practical applications. Conventional solutions for this problem incorporate quantum-inspired neural networks with Kronecker products and hybrid tensor neural networks with MPO factorization and fully-connected layers. Nonetheless, the generalization power and robustness of the fully-connected layers are not as outstanding as circuit models in quantum computing. In this paper, we propose a novel tensor circuit neural network (TCNN) that takes advantage of the characteristics of tensor neural networks and residual circuit models to achieve generalization ability and robustness with low complexity. The proposed activation operation and parallelism of the circuit in complex number field improves its non-linearity and efficiency for feature learning. Moreover, since the feature information exists in the parameters in both the real and imaginary parts in TCNN, an information fusion layer is proposed for merging features stored in those parameters to enhance the generalization capability. Experimental results confirm that TCNN showcases more outstanding generalization and robustness with its average accuracies on various datasets 2\%-3\% higher than those of the state-of-the-art compared models. More significantly, while other models fail to learn features under noise parameter attacking, TCNN still showcases prominent learning capability owing to its ability to prevent gradient explosion. Furthermore, it is comparable to the compared models on the number of trainable parameters and the CPU running time. An ablation study also indicates the advantage of the activation operation, the parallelism architecture and the information fusion layer.

Khmer polarity classification is a fundamental natural language processing task that assigns a positive, negative, or neutral label to a given Khmer text input. Existing Khmer models typically predict the label without explaining the rationale behind the prediction. This paper proposes an explainable Khmer polarity classifier by fine-tuning an instruction-based reasoning Qwen-3 model. The notion of explainability in this paper is limited to self-explanations, which the model uses to rationalize its predictions. Experimental results show that the fine-tuned model not only predicts labels accurately but also provides reasoning by identifying polarity-related keywords or phrases to support its predictions. In addition, we contribute a new Khmer polarity dataset consisting of short- to medium-length casual, romanized, and mixed-code Khmer expressions. This dataset was constructed using both heuristic rules and human curation and is publicly available through a gated Hugging Face repository (rinabuoy/khmerpolarity_nonreasoning). The fine-tuned Qwen-3 models are also made available in the same Hugging Face account.

People have different creative writing preferences, and large language models (LLMs) for these tasks can benefit from adapting to each user's preferences. However, these models are often trained over a dataset that considers varying personal tastes as a monolith. To facilitate developing personalized creative writing LLMs, we introduce LiteraryTaste, a dataset of reading preferences from 60 people, where each person: 1) self-reported their reading habits and tastes (stated preference), and 2) annotated their preferences over 100 pairs of short creative writing texts (revealed preference). With our dataset, we found that: 1) people diverge on creative writing preferences, 2) finetuning a transformer encoder could achieve 75.8% and 67.7% accuracy when modeling personal and collective revealed preferences, and 3) stated preferences had limited utility in modeling revealed preferences. With an LLM-driven interpretability pipeline, we analyzed how people's preferences vary. We hope our work serves as a cornerstone for personalizing creative writing technologies.

Measuring GUI task difficulty is crucial for user behavior analysis and agent capability evaluation. Yet, existing benchmarks typically quantify difficulty based on motor actions (e.g., step counts), overlooking the cognitive demands underlying task completion. In this work, we propose Cognitive Chain, a novel framework that models task difficulty from a cognitive perspective. A cognitive chain decomposes the cognitive processes preceding a motor action into a sequence of cognitive steps (e.g., finding, deciding, computing), each with a difficulty index grounded in information theories. We develop an LLM-based method to automatically extract cognitive chains from task execution traces. Validation with linear regression shows that our estimated cognitive difficulty correlates well with user completion time (step-level R-square=0.46 after annotation). Assessment of state-of-the-art GUI agents shows reduced success on cognitively demanding tasks, revealing capability gaps and Human-AI consistency patterns. We conclude by discussing potential applications in agent training, capability assessment, and human-agent delegation optimization.

Although neural networks are a powerful tool, their widespread use is hindered by the opacity of their decisions and their black-box nature, which result in a lack of trustworthiness. To alleviate this problem, methods in the field of explainable Artificial Intelligence try to unveil how such automated decisions are made. But explainable AI methods are often plagued by missing faithfulness/correctness, meaning that they sometimes provide explanations that do not align with the neural network's decision and logic. Recently, transformations to decision trees have been proposed to overcome such problems. Unfortunately, they typically lack exactness, scalability, or interpretability as the size of the neural network grows. Thus, we generalize these previous results, especially by considering convolutional neural networks, recurrent neural networks, non-ReLU activation functions, and bias terms. Our findings are accompanied by rigorous proofs and we present a novel algorithm RENTT (Runtime Efficient Network to Tree Transformation) designed to compute an exact equivalent decision tree representation of neural networks in a manner that is both runtime and memory efficient. The resulting decision trees are multivariate and thus, possibly too complex to understand. To alleviate this problem, we also provide a method to calculate the ground truth feature importance for neural networks via the equivalent decision trees - for entire models (global), specific input regions (regional), or single decisions (local). All theoretical results are supported by detailed numerical experiments that emphasize two key aspects: the computational efficiency and scalability of our algorithm, and that only RENTT succeeds in uncovering ground truth explanations compared to conventional approximation methods like LIME and SHAP. All code is available at https://github.com/HelenaM23/RENTT .

Clinical time series derived from electronic health records (EHRs) are inherently irregular, with asynchronous sampling, missing values, and heterogeneous feature dynamics. While numerical laboratory measurements are highly informative, existing embedding strategies usually combine feature identity and value embeddings through additive operations, which constrains their ability to capture value-dependent feature interactions. We propose MedFuse, a framework for irregular clinical time series centered on the MuFuse (Multiplicative Embedding Fusion) module. MuFuse fuses value and feature embeddings through multiplicative modulation, preserving feature-specific information while modeling higher-order dependencies across features. Experiments on three real-world datasets covering both intensive and chronic care show that MedFuse consistently outperforms state-of-the-art baselines on key predictive tasks. Analysis of the learned representations further demonstrates that multiplicative fusion enhances expressiveness and supports cross-dataset pretraining. These results establish MedFuse as a generalizable approach for modeling irregular clinical time series.

Multimodal Large Language Models (MLLMs) often suffer from hallucinations, particularly errors in object existence, attributes, or relations, which undermine their reliability. We introduce TACO (Verified Atomic Confidence Estimation), a simple framework that mitigates hallucinations through self-verification and confidence calibration without relying on external vision experts. TACO decomposes responses into atomic queries, paraphrases them to reduce sensitivity to wording, and estimates confidence using self-consistency (black-box) or self-confidence (gray-box) aggregation, before refining answers with a language model. Experiments on five benchmarks (POPE, MME, HallusionBench, AMBER, and MM-Hal Bench) with two MLLMs (\texttt{LLaVA-1.5-7B} and \texttt{CogVLM2}) show that TACO consistently outperforms direct prompting and Visual Contrastive Decoding, reduces systematic biases, and improves confidence calibration, demonstrating its effectiveness in enhancing the faithfulness of MLLMs.

This paper reports on pretraining ModernBERT encoder models in six different sizes, ranging from 51M to 475M parameters, with a focus on limited multilingualism, emphasizing languages relevant to Finland. Our models are competitive with, or superior to, existing multilingual models. They outperform monolingual models on tasks that require a context longer than 512 tokens. We present empirical results on using different data in the final stage of training. The code and models are publicly released.

Large language models are increasingly used for many applications. To prevent illicit use, it is desirable to be able to detect AI-generated text. Training and evaluation of such detectors critically depend on suitable benchmark datasets. Several groups took on the tedious work of collecting, curating, and publishing large and diverse datasets for this task. However, it remains an open challenge to ensure high quality in all relevant aspects of such a dataset. For example, the DetectRL benchmark exhibits relatively simple patterns of AI-generation in 98.5% of the Claude-LLM data. These patterns may include introductory words such as "Sure! Here is the academic article abstract:", or instances where the LLM rejects the prompted task. In this work, we demonstrate that detectors trained on such data use such patterns as shortcuts, which facilitates spoofing attacks on the trained detectors. We consequently reprocessed the DetectRL dataset with several cleansing operations. Experiments show that such data cleansing makes direct attacks more difficult. The reprocessed dataset is publicly available.