Large Language Models (LLMs) have shown remarkable capabilities, yet ensuring their outputs conform to strict structural or grammatical constraints remains challenging, which is critical in function calls and domain-specific language (DSL) generation. Constrained decoding with context-free grammar is a flexible approach to guarantee LLMs’ adherence to a specific format by dynamically building a token logits mask. However, creating this mask requires checking the validity of all tokens in the LLM vocabulary at every decoding step, which often incurs significant overheads in existing constrained decoding engines. To address this challenge, we propose $\textbf{ZapFormat}$, a novel $\textbf{dynamic pruning}$ strategy based on the Earley algorithm that identifies and eliminates invalid or redundant Earley states in real-time, significantly reducing memory occupation of the Earley algorithm’s states. This further enables us to use a state cache to speed up structured generations on a large number of queries. We implemented ZapFormat in a new constrained decoding engine called Formatron which also incorporates existing optimizations. Through comprehensive experiments on structured generation tasks, including JSON generation, JSON Schema validation, and semantic parsing, we demonstrate that Formatron not only $\textbf{consistently maintains}$ high-precision compliant outputs but also achieves $\textbf{significant improvements}$ in inference speed up to 2x compared to state-of-the-art implementations. More importantly, Formatron is generally applicable across various LLM architectures. We release Formatron as open source at https://github.com/Dan-wanna-M/formatron.

Unmanned aerial vehicles (UAV) swarm network (USNET) is a promising solution for diverse applications and usually works in harsh environments. However, it is challenging to rebuild the communication connectivity in USNETs with low time complexity under unpredictable UAV damages. In this paper, we present a graph attention network (GAT)-based deep reinforcement learning topology self-healing (GDR-TS) framework to minimize the topology self-healing (TSH) time of remaining UAVs. Specifically, we decompose the TSH problem into a neighbor selection problem and a trajectory planning problem. For the former, we present an edge update GAT-based actor-critic structure to find an optimal adjacency matrix; for the latter, we present a node update GAT-based position fine-tuning module to compute an optimal position matrix. Simulation results show that under various disruption setups, our GDR-TS framework outperforms the other four baselines in terms of the average TSH time and the response time.

Pre-trained Large Language Models (LLMs) have demonstrated significant potential in the Natural Language to Code (NL2Code) task. However, user-provided natural language descriptions are often ambiguous or misleading, resulting in poor code quality. Additionally, LLMs have limited ability to solve complex programming tasks with multiple requirements or higher difficulty levels . To address this, we developed a collaborative code generation framework integrating intelligent agents with LLMs. This framework optimizes code generation by dividing the NL2Code task into four stages: role definition, demand optimization, code writing, and code review. Based on the task-instruction-prompts and role-definition-prompts dataset we created, we fine-tuned a BART model to develop an intelligent agent enriched with programming knowledge. This agent generates more detailed prompts to guide LLMs at each stage, enhancing code generation capabilities. We evaluated our method on multiple datasets, including HumanEval, MBPP and LLMSecEval, using various metrics such as pass@k, logical error rate, code quality score, vulnerable@k, and secure@k to assess coding ability. To further strengthen our evaluation, we compared our approach against several specialized code-focused LLMs, including CodeGeeX, CodeLlama, and DeepSeek-Coder-V2. Our method improved the pass@1 metric by 18.9% and 23.2%, compared to the GPT-3.5 baseline and reduced the logical error rate from 38.2% and 29.1% to 19.3% and 13.6%. Moreover, the integration of the intelligent agent with GPT-4 led to notable performance improvements , with a 16.9% increase in pass@1, a 8.14% increase in code quality score, and a reduction in the logical error rate from 12.5% to 8.1%. Additionally, the Secure@1 metric for GPT-4 improved from 48.6% to 55.3%, reflecting enhanced code security. These experimental results demonstrate the efficacy of our method in advancing the code generation capabilities of LLMs.

Anomaly detection of Key Performance Indicators (KPIs) e.g. response latency, network throughput, etc., is one of the key techniques to ensure the quality and security of network services. However, state-of-the-art unsupervised deep learning algorithms present limitations: they are sensitive to noise and demand extensive KPI data for training, complicating the detection process. This paper proposes an Unsupervised Deep Transfer Learning (UDTL) approach. First, UDTL uses self-training preprocessing that generates reliable, high-quality samples for model training. Then, UDTL calculates the correlation based on the shape of the KPIs and the deviation scores of the KPIs, and clusters these KPIs into different clusters via correlation. At last, UDTL selects the KPI closest to each cluster’s centroid to train a base anomaly detection model for this cluster. The anomaly detection model of other KPIs adaptively choose the parameters to transfer based on correlation. Our experiments conducted on several public datasets highlight UDTL’s effectiveness. UDTL enhances the F1 score by 15.24% and reduces the training time by 22.17 times compared to baselines.

Due to the highly dynamic network topology in Flying Ad hoc Networks (FANETs), topology-based routing protocols are impractical, while known geographic routing protocols suffer from routing holes. In this paper, a Q-learning-based adaptive geographic routing protocol is presented, in which the impact of delay, mobility, and energy consumption on routing is comprehensively considered. To overcome the dynamic changes in routing caused by the high mobility of Unmanned Aerial Vehicles (UAVs), reinforcement learning parameters and HELLO message interval are adaptively adjusted by sensing local topology changes. Meanwhile, a new routing hole avoidance mechanism is proposed, which involves a scheme of broadcasting routing hole information and an approach of data forwarding route selection based on the node degree of UAVs and the distance to the destination node. To enable the routing protocol to adapt to highly dynamic changes in FANETs topology, link pre-learning and multi-Q learning methods are used to speed up the learning process. We use NS-3 to evaluate the proposed routing protocol. The results show that our protocol improves throughput by 6% and reduces delay by 20% compared to Q-learning-based Multi-objective Optimization Routing (QMR). It is also far superior to Q-learning-based Geographic Routing (QGeo) and Greedy Perimeter Stateless Routing (GPSR).