Dual-stream Expert Reasoning for Maintaining Focus in Chinese Counseling Dialogues

Sustaining dialogue focus remains a bottleneck for LLM-based counseling, where conventional linear generation paradigms conflate peripheral narrative noise with essential therapeutic topic, leading to narrative drift. To address this, we propose Dual-stream Expert Reasoning (DER), a framework that decomposes etiology into two orthogonal streams: the 4Ps framework (external) and the Insight Triangle (internal). Unlike serial reasoning, this parallel architecture isolates situational stressors from intrapsychic conflicts, effectively preventing information interference and error propagation. To synthesize these decoupled insights, a cross-stream fusion module cross-validates divergent perspectives to strictly anchor the dialogue response to the primary distress. To enable privacy-preserving local deployment, we further introduce a reasoning-aware distillation strategy that explicitly internalizes the expert's dual-stream reasoning trajectories into a parameter-efficient student model. Empirical results demonstrate that DER establishes a new state-of-the-art in focus maintenance, while the distilled model retains over 90% reasoning fidelity for efficient on-device inference.