Do We Need the Largest LLMs for Every Clinical Decision?

Computer Science > Computation and Language arXiv:2610.26937 (cs) [Submitted on 17 Oct 2026] Title: Do We Need the Largest LLMs for Every Clinical Decision? Authors: Ali H. Lazem, William Teahan.

计算机科学 > 计算与语言 arXiv:2610.26937 (cs) [提交于 2026 年 10 月 17 日] 标题：我们是否需要在每一项临床决策中都使用最大的大语言模型？作者：Ali H. Lazem, William Teahan。

Abstract: Large clinical language models often require multi-GPU infrastructure, limiting local deployment. We introduce CERTA, a decision-level statistical certificate that determines when a quantized model may substitute for its native BF16 reference and escalates decisions that cannot be certified to it.

摘要：大型临床语言模型通常需要多 GPU 基础设施，这限制了本地部署。我们引入了 CERTA，这是一种决策级统计证书，用于确定量化模型何时可以替代其原生 BF16 参考模型，并将无法认证的决策升级处理。

Across eleven instruction-tuned models, both general-purpose and domain-adapted, from three families (7-72B), we evaluated 2,500 clinical-verification decisions from de-identified patient contexts (2,484 patients), with patient-disjoint calibration and testing.

在来自三个系列（7-72B）的十一个指令微调模型（包括通用模型和领域适配模型）中，我们评估了来自去标识化患者背景（2,484 名患者）的 2,500 项临床验证决策，并进行了患者不相交的校准和测试。

In the higher-power operational analysis, round-to-nearest 8-bit quantization at BF16 storage served 99.8-100% of held-out decisions across all eleven models; formal finite-sample guarantees were assessed separately using the exact split-sample procedure.

在更高功率的操作分析中，BF16 存储下的四舍五入 8 位量化处理了所有十一个模型中 99.8-100% 的留出决策；正式的有限样本保证是使用精确的拆分样本程序单独评估的。

In the same operational analysis, the class-conditional 4-bit frontier reduced the amortized, BF16-equivalent deployment cost to 25.8% and 27.3% of the BF16 reference for Qwen2.5-72B and HuatuoGPT-o1-72B, respectively, with measured peak memory of 43.0 and 42.7 GB, below the nominal capacity of a single 48 GB accelerator.

在同一操作分析中，类条件 4 位前沿将 Qwen2.5-72B 和 HuatuoGPT-o1-72B 的摊销 BF16 等效部署成本分别降低至 BF16 参考成本的 25.8% 和 27.3%，测得的峰值内存分别为 43.0 GB 和 42.7 GB，低于单个 48 GB 加速器的标称容量。

Deployment on that specific hardware was not directly tested. Under the exact split-sample certificate, which also accounts for configuration selection, these costs were 26.7% and 27.9%.

该特定硬件上的部署未进行直接测试。在同样考虑了配置选择的精确拆分样本证书下，这些成本分别为 26.7% 和 27.9%。

On external disease named-entity benchmarks against human annotations, the frontier-selected NF4 configurations had a mean absolute F1 change of 0.045 on both datasets, with substantial model-specific variation. CERTA converts compression from a fixed precision choice into a risk-controlled deployment policy for this verification task.

在针对人类注释的外部疾病命名实体基准测试中，前沿选择的 NF4 配置在两个数据集上的平均绝对 F1 变化均为 0.045，且存在显著的模型特定差异。CERTA 将压缩从固定的精度选择转换为针对此验证任务的风险可控部署策略。