On-Policy Optimization with Group Equivalent Preference for Multi-Programming Language Understanding

Large language models (LLMs) achieve remarkable performance in code generation tasks. However, a significant performance disparity persists between popular programming languages (e.g., Python, C++) and others. To address this capability gap, we leverage the code translation task to train LLMs, thereby facilitating the transfer of coding proficiency across diverse programming languages. Moreover, we introduce OORL for training, a novel reinforcement learning (RL) framework that integrates on-policy and off-policy strategies. Within OORL, on-policy RL is applied during code translation, guided by a rule-based reward signal derived from unit tests. Complementing this coarse-grained rule-based reward, we propose Group Equivalent Preference Optimization (GEPO), a novel preference optimization method. Specifically, GEPO trains the LLM using intermediate representations (IRs) groups. LLMs can be guided to discern IRs equivalent to the source code from inequivalent ones, while also utilizing signals about the mutual equivalence between IRs within the group. This process allows LLMs to capture nuanced aspects of code functionality. By employing OORL for training with code translation tasks, LLMs improve their recognition of code functionality and their understanding of the relationships between code implemented in different languages. Extensive experiments demonstrate that our OORL for LLMs training with code translation tasks achieves significant performance improvements on code benchmarks across multiple programming languages.

TDGCN-Based Mobile Multiuser Physical-Layer Authentication for EI-Enabled IIoT

Physical-Layer Authentication (PLA) offers endogenous security, lightweight implementation, and high reliability, making it a promising complement to upper-layer security methods in Edge Intelligence (EI)-empowered Industrial Internet of Things (IIoT). However, state-of-the-art Channel State Information (CSI)-based PLA schemes face challenges in recognizing mobile multi-users due to the limited reliability of CSI fingerprints in low Signal-to-Noise Ratio (SNR) environments and the constantly shifting CSI distributions with user movements. To address these issues, we propose a Temporal Dynamic Graph Convolutional Network (TDGCN)-based PLA scheme. This scheme harnesses Intelligent Reflecting Surfaces (IRSs) to refine CSI fingerprint precision and employs Graph Neural Networks (GNNs) to capture the spatio-temporal dynamics induced by user movements and IRS deployments. Specifically, we partition hierarchical CSI fingerprints into multivariate time series and utilize dynamic GNNs to capture their associations. Additionally, Temporal Convolutional Networks (TCNs) handle temporal dependencies within each CSI fingerprint dimension. Dynamic Graph Isomorphism Networks (GINs) and cascade node clustering pooling further enable efficient information aggregation and reduced computational complexity. Simulations demonstrate the proposed scheme's superior authentication accuracy compared to seven baseline schemes.