Abstract Large Language Models (LLMs) exhibit distinct capabilities across different knowledge domains, yet single-model deployments struggle with knowledge-intensive tasks requiring cross-domain reasoning. We present eVoiceClaw Desktop V3, an iteratively optimized multi-model orchestration system that dynamically routes complex queries to specialized models through a novel dimension-direct routing algorithm. Through seven systematic experimental stages spanning from March 24–29, 2026, we evolved from kNN-based routing to a dimension-aware model selection framework that addresses critical challenges in cross-domain task allocation and semantic accumulation bias. Our final system achieves a 98% workflow trigger rate across 53 benchmark questions, leveraging 12 distinct models to generate responses averaging 8,533 characters with balanced diversity (top model ≤24.7% usage share). We clarify that our primary metric is response completeness(measured by character count) rather than quality per se, as length does not guarantee accuracy. Compared to single-model baselines with significantly lower computational cost, eVoiceClaw V3 generates more comprehensive responses at the expense of approximately 9× higher latency and cost. The system demonstrates robust performance on interdisciplinary tasks (e.g., legal-technical cross-domain questions), where earlier kNN versions failed due to anchor sparsity. Quality evaluation via LLM-as-Judge (Opus 4.6) is in progress; this version focuses on system performance and design iteration. eVoiceClaw Desktop V3 is released as a minimal executable open-source implementation at https://github.com/rodneyrui/evoiceclaw-desktop-v3.
Tao Rui (Sun,) studied this question.