Large Language Models (LLMs) have demonstrated strong performance across various reasoning tasks, yet building a single model that consistently excels across all domains remains challenging. This paper addresses this problem by exploring strategies to integrate multiple domain-specialized models into an efficient pivot model.We propose two fusion strategies to combine the strengths of multiple LLMs: (1) a pairwise, multi-step fusion approach that sequentially distills each source model into the pivot model, followed by a weight merging step to integrate the distilled models into the final model. This method achieves strong performance but requires substantial training effort; and (2) a unified fusion approach that aggregates all source models' outputs simultaneously.To improve the fusion process, we introduce a novel Rate-Skewness Adaptive Fusion (RSAF) technique, which dynamically adjusts top-K ratios during parameter merging for enhanced flexibility and stability.Furthermore, we propose an uncertainty-based weighting method for the unified approach, which dynamically balances the contributions of source models and outperforms other logits/distribution ensemble methods.We achieved accuracy improvements of 9.27%, 8.80%, and 8.89% on the GSM8K, MATH, and HumanEval tasks, respectively.

Recent advancements in building domain-specific large language models (LLMs) have shown remarkable success, especially in tasks requiring reasoning abilities like logical inference over complex relationships and multi-step problem solving. However, creating a powerful all-in-one LLM remains challenging due to the need for proprietary data and vast computational resources. As a resource-friendly alternative, we explore the potential of merging multiple expert models into a single LLM. Existing studies on model merging mainly focus on generalist LLMs instead of domain experts, or the LLMs under the same architecture and size. In this work, we propose an unconstrained model merging framework that accommodates both homogeneous and heterogeneous model architectures with a focus on reasoning tasks. A fine-grained layer-wise weight merging strategy is designed for homogeneous models merging, while heterogeneous model merging is built upon the probabilistic distribution knowledge derived from instruction-response fine-tuning data. Across 7 benchmarks and 9 reasoning-optimized LLMs, we reveal key findings that combinatorial reasoning emerges from merging which surpasses simple additive effects. We propose that unconstrained model merging could serve as a foundation for decentralized LLMs, marking a notable progression from the existing centralized LLM framework. This evolution could enhance wider participation and stimulate additional advancement in the field of artificial intelligence, effectively addressing the constraints posed by centralized models.

Simulation data can be accurately labeled and have been expected to improve the performance of data-driven algorithms, including object detection. However, due to the various domain inconsistencies from simulation to reality (sim-to-real), cross-domain object detection algorithms usually suffer from dramatic performance drops. While numerous unsupervised domain adaptation (UDA) methods have been developed to address cross-domain tasks between real-world datasets, progress in sim-to-real remains limited. This paper presents a novel Complex-to-Simple (CTS) framework to transfer models from labeled simulation (source) to unlabeled reality (target) domains. Based on a two-stage detector, the novelty of this work is threefold: 1) developing fixed-size anchor heads and RoI augmentation to address size bias and feature diversity between two domains, thereby improving the quality of pseudo-label; 2) developing a novel corner-format representation of aleatoric uncertainty (AU) for the bounding box, to uniformly quantify pseudo-label quality; 3) developing a noise-aware mean teacher domain adaptation method based on AU, as well as object-level and frame-level sampling strategies, to migrate the impact of noisy labels. Experimental results demonstrate that our proposed approach significantly enhances the sim-to-real domain adaptation capability of 3D object detection models, outperforming state-of-the-art cross-domain algorithms, which are usually developed for real-to-real UDA tasks.