Abstract
6G on-demand services require refined scenario cognition, evidence-grounded reasoning, and trustworthy service recommendation under highly heterogeneous environments. Existing template-based and feature-driven models offer limited semantic abstraction and weak cross-scenario generalization, while autoregressive LLM agents incur significant multi-round latency in retrieval-intensive service reasoning tasks. This paper proposes a retrieval-conditioned diffusion large language model (RC-DLLM) framework for scenario cognition and explainable recommendation in 6G on-demand services. The framework jointly encodes structured service semantics and instantaneous network conditions into a shared latent scenario state, which drives hybrid evidence retrieval over an OSOD-style 6G knowledge base, retrieval-conditioned latent diffusion cognition, and network-aware DRL-based service decision optimization. A P-ReAct-inspired scheduler further coordinates reasoning and evidence acquisition to reduce inference latency. Unlike disconnected modular pipelines, RC-DLLM explicitly couples scenario representation, knowledge retrieval, diffusion cognition, network-feasible recommendation, and explanation generation within a unified framework. Extensive experiments show that RC-DLLM consistently improves cognition accuracy, ranking quality, evidence-grounded explainability, generalization, and latency efficiency over strong baselines.