Deep model fusion/merging is an emerging technique that integrates parameters or predictions from multiple deep learning (DL) models into a unified framework. It combines the abilities of different models to compensate for the biases and errors of an individual model, improving overall performance. However, deep model fusion, especially on large-scale DL models such as large language models (LLMs) and foundation models, faces several challenges, including high computational cost and interference between different heterogeneous models. In order to understand it better, we present a comprehensive survey to summarize the recent progress. We categorize existing model fusion methods as fourfold: 1) weight average (WA) averages the parameters of multiple models to obtain results closer to the optimal solution; 2) considering that direct averaging of models often yields suboptimal results, “mode connectivity” connects networks via paths of nonincreasing loss in weight spaces before the fusion. Along these paths, initial models are transformed into forms with consistent functions and better fusion effects; 3) similarly, for models with poor direct fusion results, “alignment” matches the corresponding units and merges these models, thus fully exploiting the corresponding relationships between the models; and 4) in addition to the above-mentioned methods of parameter fusion, “ensemble learning” fuses the outputs of multiple models in the inference stage to improve the accuracy and robustness of networks. In addition, we analyze the challenges of deep model fusion and illuminate the possible research directions in the future.

Since the submarine has become the major threat to maritime security, there is an urgent need to find a more efficient method of anti-submarine warfare (ASW). The digital twin theory is one of the most outstanding information technologies, and has been quite popular in recent years. The most influential change produced by digital twin is the ability to enable real-time dynamic interactions between the simulation world and the real world. Digital twin can be regarded as a paradigm by means of which selected online measurements are dynamically assimilated into the simulation world, with the running simulation model guiding the real world adaptively in reverse. By combining digital twin theory and random finite sets (RFSs) closely, a new framework of sensor control in ASW is proposed. Two key algorithms are proposed for supporting the digital twin-based framework. First, the RFS-based data-assimilation algorithm is proposed for online assimilating the sequence of real-time measurements with detection uncertainty, data association uncertainty, noise, and clutters. Second, the computation of the reward function by using the results of the proposed data-assimilation algorithm is introduced to find the optimal control action. The results of three groups of experiments successfully verify the feasibility and effectiveness of the proposed approach.

In the military field, multi-agent simulation (MAS) plays an important role in studying wars statistically. For a military simulation system, which involves large-scale entities and generates a very large number of interactions during the runtime, the issue of how to improve the running efficiency is of great concern for researchers. Current solutions mainly use hybrid simulation to gain fewer updates and synchronizations, where some important continuous models are maintained implicitly to keep the system dynamics, and partial resynchronization (PR) is chosen as the preferable state update mechanism. However, problems, such as resynchronization interval selection and cyclic dependency, remain unsolved in PR, which easily lead to low update efficiency and infinite looping of the state update process. To address these problems, this paper proposes a lookahead behavior model (LBM) to implement a PR-based hybrid simulation. In LBM, a minimal safe time window is used to predict the interactions between implicit models, upon which the resynchronization interval can be efficiently determined. Moreover, the LBM gives an estimated state value in the lookahead process so as to break the state-dependent cycle. The simulation results show that, compared with traditional mechanisms, LBM requires fewer updates and synchronizations.