The emergence of multi-agent systems (MAS) based on large language models (LLMs) has enabled autonomous collaboration on complex, goal-oriented tasks. However, effective interaction is frequently hindered by semantic drift, a phenomenon where heterogeneous agents assign conflicting meanings to shared terminology due to differing internal prompts or domain knowledge. Existing communication paradigms either rely on unconstrained natural language, which suffers from structural vagueness, or rigid symbolic schemas that fail to adapt to emergent concepts. To address this gap, we propose DOA, a novel dynamic ontology alignment framework that serves as a semantic mediation layer for MAS. DOA integrates a proactive semantic prober to detect conceptual mismatches and a neuro-symbolic aligner that reconciles local semantic structures in real time. By grounding fluid natural language dialogues in an evolving shared ontology, our framework ensures deterministic mutual understanding over long-horizon tasks. Empirical evaluations in cross-domain supply chain and healthcare coordination scenarios demonstrate that DOA improves task success rates by an average of 31.5% and reduces communication overhead (token consumption) by 50% compared to state-of-the-art baselines. Our results provide a robust and scalable foundation for semantic consistency in next-generation industrial-grade AI systems.

With the emergence of AI-agent communication networks (ACN) in the 6G era, the efficient training of agents for complex reasoning tasks has become a critical capability for scalable ACN deployment. As a representative complex reasoning task, retrieval-augmented multi-hop question answering (e.g., agentic retrieval-augmented generation) requires agents to perform multi-step reasoning through reflection, planning, and tool-use mechanisms. However, reinforcement learning training still faces reward sparsity and sample efficiency challenges, limiting agents’ rapid evolution and adaptability. We propose dynamic batch-aware advantage weighting (DB-AW), integrating two core components at the batch level: the difficulty-aware weighting component dynamically amplifies positive advantages based on long-term success rates, directing learning toward learnable yet challenging samples; and the batch filtering component removes zero-variance groups, ensuring each update contains non-zero gradient signals. Experiments show that DB-AW achieves 18%, 17%, and 15% relative improvements on Qwen2.5-7B, Qwen2.5-3B, and LLaMA3.2-3B, respectively, while improving the effective update rate from 68% to 100%, significantly reducing agent training costs. As a lightweight and reusable algorithmic module, DB-AW can be readily integrated into methods such as group relative policy optimization (GRPO), providing a practical pathway for efficient training of complex reasoning agents in ACN.