Cost-Performance Adaptive Large Language Model Routing for Multi-task Scenarios

This paper proposes a cost-performance adaptive external routing method for large language models (LLMs) in multi-task scenarios with heterogeneous task requirements, changing invocation costs, latency variation, and unstable commercial API availability. The method jointly models task type, sample complexity, task-model profiles, and runtime status. A unified interface organizes six task families: classification, extraction, generation, question answering, mathematical reasoning, and summarization. The router first constructs task-model profiles from quality, success rate, cost, and latency statistics. It then combines profile-aware scoring, rule-enhanced sample biases, a lightweight local reranking head, and robustness checking to select an execution model from a heterogeneous commercial model pool. Experiments on 600 online samples show that the proposed method achieves a 100% effective success rate, an average cost of 0.000357 USD per sample, and a final budget-node accuracy of 0.6441. Comparative analysis indicates that RouteLLM remains stronger at the lowest budget nodes, whereas the proposed method is more suitable for full-budget or quality-oriented settings where sustained accuracy gains are required.