{
  "abstract": "Objective With the increasing adoption of large language models (LLMs) in practical applications, selecting between ‘thinking’ (high-reasoning) and ‘non-thinking’ (fast, low-cost) modes requires balancing performance and operational cost. In medical question answering, task complexity varies widely—from routine medication adjustments to complex rare disease diagnosis—placing different demands on LLM reasoning capabilities. While the thinking mode excels at complex inference, it often incurs higher computational cost, latency and hallucination risk. In contrast, the non-thinking mode provides faster, cheaper responses but struggles with complex reasoning. This study proposes SynapseRoute, an intelligent routing framework that dynamically assigns queries to either mode to optimise accuracy, latency and token cost.Methods and analysis Questions from four medical datasets (USMLE, MedMCQA, PubMedQA and CareQA) were automatically labelled as thinking or non-thinking using the Qwen3-30B-a3b model. Multiple routing algorithms, including traditional machine learning methods and fine-tuned LLMs, were evaluated. Performance was assessed using standard metrics and a novel cross-dimensional metric, the Accuracy–Inference–Token (AIT) index, which integrates accuracy, latency and token cost with adjustable weights.Results Approximately 58% of medical questions were correctly answered by the non-thinking mode alone. Logistic Regression achieved the best routing performance (area under the curve=0.82). Compared with using the thinking mode alone, SynapseRoute improved accuracy (0.8390 vs 0.8272), while reducing inference time by 36.8% and token consumption by 39.66%.Discussion Dynamic mode routing improves the cost-efficiency and practicality of dual-mode LLMs in medical question answering. SynapseRoute and the AIT index provide practical tools for deploying adaptive LLM systems with controllable performance–cost trade-offs.",
  "authors": [
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Wencheng Zhang"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Shiqin Qiao"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Lingjie Luo"
    },
    {
      "affiliations": [
        "Xiamen University, Xiamen, Fujian, China"
      ],
      "name": "Sicheng Li"
    },
    {
      "affiliations": [
        "Xidian University, Xi’an, Shaanxi, China"
      ],
      "name": "Yinfeng Li"
    },
    {
      "affiliations": [
        "The Chinese University of Hong Kong, Hong Kong, China"
      ],
      "name": "Chuanyang Zheng"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Qian Xu"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Meng Li"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Yong Gui"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Yijun He"
    },
    {
      "affiliations": [
        "Mohamed bin Zayed University of Artificial Intelligence, Masdar City, UAE"
      ],
      "name": "Jianing Qiu"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China",
        "Peking University, Beijing, China"
      ],
      "name": "Jindong Hong"
    },
    {
      "affiliations": [
        "Bytedance, Beijing, China"
      ],
      "name": "Jiankai Sun"
    }
  ],
  "title": "SynapseRoute: an auto-route switching framework on dual-state large language model for medical question answering",
  "uid": "8e73ab6c-9451-56c4-93a6-3b845732c2b9"
}