Recent advances in large language models (LLMs) have been limited by their processing requirements and vulnerability to adversarial assaults, whilst short language models (SLMs) struggle with performance consistency in complex tasks. This research introduces ResilioMate, a resilient multi-agent framework that enhances SLMs by utilizing distributed cognitive burden distribution, dual-scale memory systems, and collaborative bias prevention strategies. The method employs dynamic task decomposition across specialized agents (e.g., Assistant, Checker) to minimize computational costs and combines short-term trajectory tracking with long-term self-reflective optimization for adaptive execution. At its core, the LeptoConnect model series (1.8B/7B parameters) is created using hybrid attention distillation and dynamic curriculum learning to enable cross-domain competence while retaining SLM efficiency. ResilioMate accomplishes three critical improvements: (1) The 1.8B LeptoConnect model attains 81.6% of GPT-4’s performance in knowledge graph construction through parameter-efficient fine-tuning with structured weight matrices; (2) LeptoConnect-7B achieves a score of 41.3 in database operations, compared to GPT-4’s 32.0, through collaborative cognitive load allocation; (3) A bias-interception network effectively suppresses adversarial propagation while achieving code correction performance of ROUGE-L’s 42.86. The framework’s dual-scale memory architecture reduces computational redundancy by 26.4% through real-time task tracking and multi-agent knowledge refinement. These developments illustrate ResilioMate’s efficacy in bridging the performance gap between SLMs and LLMs, offering a scalable solution for the deployment of efficient language agents in real-time and peripheral computing environments.
High-precision Sea Surface Temperature (SST) prediction is critical for understanding ocean–atmosphere interactions and climate anomaly monitoring. We propose GRU_EKAN, a novel hybrid model where Gated Recurrent Units (GRUs) capture temporal dependencies and the Enhanced Kolmogorov–Arnold Network (EKAN) models complex feature interactions between SST and multivariate ocean predictors. This study integrates GRU with EKAN, using B-spline-parameterized activation functions to model high-dimensional nonlinear relationships between multiple ocean variables (including sea water potential temperature at the sea floor, ocean mixed layer thickness defined by sigma theta, sea water salinity, current velocities, and sea surface height) and SST. L2 regularization addresses multicollinearity among predictors. Experiments were conducted at 25 South China Sea sites using 2011–2021 CMEMS data. The results show that GRU_EKAN achieves a superior mean R2 of 0.85, outperforming LSTM_EKAN, GRU, and LSTM by 5%, 25%, and 23%, respectively. Its average RMSE (0.90 °C), MAE (0.76 °C), and MSE (0.80 °C2) represent reductions of 31.3%, 27.0%, and 53.2% compared to GRU. The model also exhibits exceptional stability and minimal Weighted Quality Evaluation Index (WQE) fluctuation. During the 2019–2020 temperature anomaly events, GRU_EKAN predictions aligned closest with observations and captured abrupt trend shifts earliest. This model provides a robust tool for high-precision SST forecasting in the South China Sea, supporting marine heatwave warnings.
As large language models (LLMs) improve their capabilities in handling complex tasks, the issues of computational cost and efficiency due to long prompts are becoming increasingly prominent. To accelerate model inference and reduce costs, we propose an innovative prompt compression framework called LinguaShrink. Inspired by the observation that LLM performance depends on the density and position of key information in the input prompts, LinguaShrink leverages psycholinguistic principles and the Ebbinghaus memory curve to achieve task-agnostic prompt compression. This effectively reduces prompt length while preserving essential information. We referred to the training method of OpenChat [1]. The framework introduces part-of-speech priority compression and data distillation techniques, using smaller models to learn compression targets and employing a KL-regularized reinforcement learning strategy for training. Additionally, we adopt a chunk-based compression algorithm to achieve adjustable compression rates. We evaluate our method on multiple datasets, including LongBench, ZeroScrolls, Arxiv Articles, and a newly constructed novel test set. Experimental results show that LinguaShrink maintains semantic similarity while achieving up to 26 times compression. Compared to existing prompt compression methods, LinguaShrink improves end-to-end latency by 1.43 times.
To address the issues of insufficient convergence performance and high sensitivity to local optima in the traditional Whale Optimization Algorithm (WOA) when handling 3D path planning tasks for unmanned aerial vehicles (UAVs), this paper proposes an improved UAV path planning algorithm based on the Whale Optimization Algorithm (R*WOA). Firstly, the global search capability and path optimisation mechanism of the Rapidly Expanding Random Tree Star (RRT*) algorithm are utilised to generate a high-quality initial population, thereby enhancing population diversity and the algorithm’s global exploration capability; Secondly, the linear convergence factor of the traditional WOA is adjusted to a non-linear dynamic adjustment strategy based on the cosine function, enhancing global search capability in the early stages of iteration and local search capability in the later stages; simultaneously, a non-linear inertial weight is employed to modulate the position update mechanism of individuals, further enhancing the algorithm’s local optimisation accuracy and convergence stability in the later stages of iteration. Finally, comparative experimental results on a basic test function set and in scenarios constructed using Digital Elevation Models (DEMs) demonstrate that R*WOA exhibits stable optimisation performance, capable of planning safer paths that are shorter in length and smoother in trajectory.