Pretrained large language models (LLMs) are typically reused as indivisible artifacts, adapted, merged, or ensembled as a whole.In this study, we show that LLMs can instead be structurally recomposed as modular building blocks to create new architectures without access to original training data.We introduce architecturelevel reassembly as a new reuse paradigm, in which Transformer blocks from heterogeneous models are treated as reusable components.This idea is formalized through a twodimensional reassembly space that supports both vertical recombination across depth and horizontal composition within layers.To make this space tractable, we propose a chromosomebased architectural encoding and perform a bilevel multi-objective evolutionary optimization over vertical structure and horizontal composition.To resolve representation incompatibility across heterogeneous blocks, we introduce lightweight glue layers trained via data-free knowledge distillation, enabling valid information flow without modifying pretrained parameters.Our results demonstrate that architecturelevel reassembly unlocks a new dimension of flexibility in model reuse, pointing toward a modular and evolutionary view of LLM design.

Algorithm selection aims to identify the optimal performing algorithm before execution. Existing techniques typically focus on the observed correlations between algorithm performance and meta-features. However, little research has explored the underlying mechanisms of algorithm selection, specifically what characteristics an algorithm must possess to effectively tackle problems with certain feature values. This gap not only limits the explainability but also makes existing models vulnerable to data bias and distribution shift. This paper introduces directed acyclic graph (DAG) to describe this mechanism, proposing a novel modeling paradigm that aligns more closely with the fundamental logic of algorithm selection. By leveraging DAG to characterize the algorithm feature distribution conditioned on problem features, our approach enhances robustness against marginal distribution changes and allows for finer-grained predictions through the reconstruction of optimal algorithm features, with the final decision relying on differences between reconstructed and rejected algorithm features. Furthermore, we demonstrate that, the learned DAG and the proposed counterfactual calculations offer our approach with both model-level and instance-level explainability.

Model merging is a technique that combines multiple large pretrained models into a single model, enhancing performance and broadening task adaptability without original data or additional training. However, most existing model merging methods focus primarily on exploring the parameter space, merging models with identical architectures. Despite its potential, merging in the architecture space remains in its early stages due to the vast search space and challenges related to layer compatibility. This paper designs a hierarchical model merging framework named HM3, formulating a bilevel multi-objective model merging problem across both parameter and architecture spaces. At the parameter level, HM3 integrates existing merging methods to quickly identify optimal parameters. Based on these, an actor-critic strategy with efficient policy discretization is employed at the architecture level to explore inference paths with Markov property in the layer-granularity search space for reconstructing these optimal models. By training reusable policy and value networks, HM3 learns Pareto optimal models to provide customized solutions for various tasks. Experimental results on language and vision tasks demonstrate that HM3 outperforms methods focusing solely on the parameter or architecture space.

Wireless-based federated learning (FL), as an emerging distributed learning approach, has been widely studied for 6G systems. When the paradigm shifts from terrestrial to non-terrestrial networks (NTN), FL may need to address several open challengings, e.g., limited service time of low earth orbit (LEO) satellites and time-efficient uploading and aggregation for massive devices. In this work, we exploit the synergy of LEO and FL for future integrated 6G-satellite systems by taking advantage of ubiquitous wireless access provided by LEO and appealing characteristics of collaborative training and data privacy preservation in FL. The studied LEO-FL framework may need to improve multi-metric performance in practice. Different from most FL works, we simultaneously improve the communication-training efficiency and local training accuracy from a multi-objective optimization (MOO) perspective. To solve the problem, we propose a decomposition, deep reinforcement learning and transfer learning based MOO algorithm for FL (DRT-FL), aiming at adapting to the dynamic satellite-terrestrial environments, achieving efficient uploading and aggregation, and approaching Pareto optimal sets. Compared to the state-of-the-art MOO algorithms, the effectiveness of the proposed LEO-FL framework and DRT-FL algorithm are assessed on MNIST and CIFAR-IO datasets.

Wireless-based federated learning (FL), as an emerging distributed learning approach, has been widely studied for 6G systems. When the paradigm shifts from terrestrial to non-terrestrial networks (NTN), FL may need to address several open challenges, e.g., the limited service time of low earth orbit (LEO) satellites, the straggler issue in synchronous FL, and time-efficient uploading and aggregation for massive devices. In this work, we exploit the synergy of LEO and FL for future integrated 6G-satellite systems by taking advantage of ubiquitous wireless access provided by LEO and appealing characteristics of collaborative training and data privacy preservation in FL. The studied LEO-FL framework may need to improve multi-metric performance in practice. Different from most FL works, we simultaneously improve the communication-training efficiency and local training accuracy from a multi-objective optimization (MOO) perspective. To solve the problem, we propose a decomposition and meta-deep reinforcement learning based MOO algorithm for FL (DMMA-FL), aiming at adapting to the dynamic satellite-terrestrial environments, achieving efficient uploading and aggregation, and approaching Pareto optimal sets. Compared to single-objective optimization, heuristics-based, and learning-based MOO algorithms, the effectiveness and advantages of the proposed LEO-FL framework and DMMA-FL algorithm are assessed on MNIST and CIFAR-10 datasets.