Large language models deliver strong reasoning and tool-use skills, yet their computational demands make them impractical for edge or cost-sensitive deployments. We present \textbf{Xmodel-2.5}, a 1.3-billion-parameter small language model designed as a \emph{drop-in agent core}. Training with maximal-update parameterization ($μ$P) allows hyper-parameters tuned on a 20M-parameter proxy to transfer directly to the full model, even under the parameter-tied \emph{tie-word-embedding} architecture. A 1.4T-token Warmup--Stable--Decay curriculum is used, and we further show that \textbf{switching from AdamW to Muon during the decay phase} improves the 13-task reasoning average by 4.58\,\% while keeping every other hyper-parameter fixed, verifying that early AdamW stability can be paired with late Muon sharpening for better downstream performance. FP8-mixed-precision training balances accuracy and throughput. All checkpoints, recipes, and evaluation code are released under the Apache-2.0 license.\footnote{https://huggingface.co/XiaoduoAILab/Xmodel-2.5 and https://huggingface.co/XiaoduoAILab/Xmodel-2.5-history (training checkpoints).} Training code and evaluation harness: https://github.com/XiaoduoAILab/Xmodel-2.5.

Large Language Models (LLMs) require substantial computational resources, making model compression essential for efficient deployment in constrained environments. Among the dominant compression techniques: knowledge distillation, structured pruning, and low-bit quantization, their individual effects are well studied, but their interactions and optimal sequencing remain unclear. This work systematically examines how these techniques perform both independently and in combination when applied to the Qwen2.5 3B model. We evaluate multiple compression pipelines, including single, and proposed three-technique sequences, using perplexity, G-Eval, clarity, prompt alignment, and compression ratio as metrics. Our experiments show that quantization provides the greatest standalone compression, while pruning introduces moderate quality degradation. Critically, the ordering of techniques significantly affects the final model quality: the sequence Pruning, Knowledge Distillation, Quantization (P-KD-Q) yields the best balance, achieving a 3.68x compression ratio while preserving strong instruction-following and language understanding capabilities. Conversely, pipelines applying quantization early suffer severe performance degradation due to irreversible information loss that impairs subsequent training. Overall, this study offers practical insight into designing effective, ordering-aware compression pipelines for deploying LLMs in resource-limited settings.

METR's time horizon metric has grown exponentially since 2019, along with compute. However, it is unclear whether compute scaling will persist at current rates through 2030, raising the question of how possible compute slowdowns might impact AI agent capability forecasts. Given a model of time horizon as a function of training compute and algorithms, along with a model of how compute investment spills into algorithmic progress (which, notably, precludes the possibility of a software-only singularity), and the empirical fact that both time horizon and compute have grown at constant rates over 2019--2025, we derive that time horizon growth must be proportional to compute growth. We provide additional, albeit limited, experimental evidence consistent with this theory. We use our model to project time horizon growth under OpenAI's compute projection, finding substantial projected delays in some cases. For example, 1-month time horizons at $80\%$ reliability occur $7$ years later than simple trend extrapolation suggests.

Open-world machine learning is an emerging technique in artificial intelligence, where conventional machine learning models often follow closed-world assumptions, which can hinder their ability to retain previously learned knowledge for future tasks. However, automated intelligence systems must learn about novel classes and previously known tasks. The proposed model offers novel learning classes in an open and continuous learning environment. It consists of two different but connected tasks. First, it discovers unknown classes in the data and creates novel classes; next, it learns how to perform class incrementally for each new class. Together, they enable continual learning, allowing the system to expand its understanding of the data and improve over time. The proposed model also outperformed existing approaches in open-world learning. Furthermore, it demonstrated strong performance in continuous learning, achieving a highest average accuracy of 82.54% over four iterations and a minimum accuracy of 65.87%.

Deep autoencoder (DAE) frameworks have demonstrated their effectiveness in reducing channel state information (CSI) feedback overhead in massive multiple-input multiple-output (mMIMO) orthogonal frequency division multiplexing (OFDM) systems. However, existing CSI feedback models struggle to adapt to dynamic environments caused by user mobility, requiring retraining when encountering new CSI distributions. Moreover, returning to previously encountered environments often leads to performance degradation due to catastrophic forgetting. Continual learning involves enabling models to incorporate new information while maintaining performance on previously learned tasks. To address these challenges, we propose a generative adversarial network (GAN)-based learning approach for CSI feedback. By using a GAN generator as a memory unit, our method preserves knowledge from past environments and ensures consistently high performance across diverse scenarios without forgetting. Simulation results show that the proposed approach enhances the generalization capability of the DAE framework while maintaining low memory overhead. Furthermore, it can be seamlessly integrated with other advanced CSI feedback models, highlighting its robustness and adaptability.

Organization's communication is essential for public trust, but the rise of generative AI models has introduced significant challenges by generating persuasive content that can form competing narratives with official messages from government and commercial organizations at speed and scale. This has left agencies in a reactive position, often unaware of how these models construct their persuasive strategies, making it more difficult to sustain communication effectiveness. In this paper, we introduce a large LLM-generated persuasion attack dataset, which includes 134,136 attacks generated by GPT-4, Gemma 2, and Llama 3.1 on agency news. These attacks span 23 persuasive techniques from SemEval 2023 Task 3, directed toward 972 press releases from ten agencies. The generated attacks come in two mediums, press release statements and social media posts, covering both long-form and short-form communication strategies. We analyzed the moral resonance of these persuasion attacks to understand their attack vectors. GPT-4's attacks mainly focus on Care, with Authority and Loyalty also playing a role. Gemma 2 emphasizes Care and Authority, while Llama 3.1 centers on Loyalty and Care. Analyzing LLM-generated persuasive attacks across models will enable proactive defense, allow to create the reputation armor for organizations, and propel the development of both effective and resilient communications in the information ecosystem.

Accurate multivariate time-series prediction of vital signs and laboratory results is crucial for early intervention and precision medicine in intensive care units (ICUs). However, vital signs are often noisy and exhibit rapid fluctuations, while laboratory tests suffer from missing values, measurement lags, and device-specific bias, making integrative forecasting highly challenging. To address these issues, we propose OmniTFT, a deep learning framework that jointly learns and forecasts high-frequency vital signs and sparsely sampled laboratory results based on the Temporal Fusion Transformer (TFT). Specifically, OmniTFT implements four novel strategies to enhance performance: sliding window equalized sampling to balance physiological states, frequency-aware embedding shrinkage to stabilize rare-class representations, hierarchical variable selection to guide model attention toward informative feature clusters, and influence-aligned attention calibration to enhance robustness during abrupt physiological changes. By reducing the reliance on target-specific architectures and extensive feature engineering, OmniTFT enables unified modeling of multiple heterogeneous clinical targets while preserving cross-institutional generalizability. Across forecasting tasks, OmniTFT achieves substantial performance improvement for both vital signs and laboratory results on the MIMIC-III, MIMIC-IV, and eICU datasets. Its attention patterns are interpretable and consistent with known pathophysiology, underscoring its potential utility for quantitative decision support in clinical care.

Foundation models and self-supervised learning (SSL) have become central to modern AI, yet research in this area remains hindered by complex codebases, redundant re-implementations, and the heavy engineering burden of scaling experiments. We present stable-pretraining, a modular, extensible, and performance-optimized library built on top of PyTorch, Lightning, Hugging Face, and TorchMetrics. Unlike prior toolkits focused narrowly on reproducing state-of-the-art results, stable-pretraining is designed for flexibility and iteration speed: it unifies essential SSL utilities--including probes, collapse detection metrics, augmentation pipelines, and extensible evaluation routines--within a coherent and reliable framework. A central design principle is logging everything, enabling fine-grained visibility into training dynamics that makes debugging, monitoring, and reproducibility seamless. We validate the library by demonstrating its ability to generate new research insights with minimal overhead, including depthwise representation probing and the analysis of CLIP degradation under synthetic data finetuning. By lowering barriers to entry while remaining scalable to large experiments, stable-pretraining aims to accelerate discovery and expand the possibilities of foundation model research.

With the rapid evolution of large language models, retrieval enhanced generation technology has been widely used due to its ability to integrate external knowledge to improve output accuracy. However, the performance of the system is highly dependent on the quality of the retrieval module. If the retrieval results have low relevance to user needs or contain noisy information, it will directly lead to distortion of the generated content. In response to the performance bottleneck of existing models in processing tabular features, this paper proposes an XGBoost machine learning regression model based on feature engineering and particle swarm optimization. Correlation analysis shows that answer_quality is positively correlated with doc_delevance by 0.66, indicating that document relevance has a significant positive effect on answer quality, and improving document relevance may enhance answer quality; The strong negative correlations between semantic similarity, redundancy, and diversity were -0.89 and -0.88, respectively, indicating a tradeoff between semantic similarity, redundancy, and diversity. In other words, as the former two increased, diversity significantly decreased. The experimental results comparing decision trees, AdaBoost, etc. show that the VMD PSO BiLSTM model is superior in all evaluation indicators, with significantly lower MSE, RMSE, MAE, and MAPE compared to the comparison model. The R2 value is higher, indicating that its prediction accuracy, stability, and data interpretation ability are more outstanding. This achievement provides an effective path for optimizing the retrieval quality and improving the generation effect of RAG system, and has important value in promoting the implementation and application of related technologies.

Mixture-of-Experts (MoE) architectures are increasingly adopted in large language models (LLMs) for their scalability and efficiency. However, their modular structure introduces a unique vulnerability: adversaries can attempt to compress or repurpose models by pruning experts and cheaply fine-tuning the remainder, effectively bypassing licensing and security constraints. In this paper, we systematically study the prunability of MoE-LLMs under task-specific usage. We first develop an expert attribution framework that identifies the subset of experts most responsible for a given task, then evaluate the performance trade-offs of pruning and re-aligning these experts using active learning-driven fine-tuning. Our findings reveal a critical knowledge loss--recovery trade-off: while certain experts can be isolated to retain task accuracy, significant degradation occurs without targeted re-alignment. Based on this analysis, we propose defense strategies that aim to make MoE models harder to compress and fine-tune without authorization, including entangled expert training and selective fine-tuning protocols that resist unauthorized adaptation. By positioning expert pruning as both a threat vector and a defense target, this work highlights the dual-use nature of MoE modularity and provides the first systematic evaluation framework for secure specialization of MoE-LLMs.