将验证作为LLM智能体的架构层：V模型设计与确定性核心试点研究
原标题：Verification as an Architectural Layer for LLM Agents: A V-Model Design, and a Pilot Study of Its Deterministic Core
AI 摘要
论文提出将验证作为LLM智能体系统的架构层，借鉴软件工程的V模型，让规范层级从需求逐级下降到单步，每级配备专用验证器，由确定性控制器强制执行判定，且只有验证结果才能写入记忆。在MuSiQue四跳数据集上用8B参数骨干进行试点，47次执行中两个未验证配置十题全未答对，而无规划器的验证配置答对八题并对其余弃权；九个纠正中有八个来自零边际成本的确定性门控，且验证存在时规划反而降低性能。作者强调该结果刻画的是终止行为而非规模化准确率，并给出十二个月计划以完成和评估完整架构。
正文节选
Verification as an Architectural Layer for LLM Agents Abstract. Large language model (LLM) agents built on the ReAct pattern concentrate four responsibilities in a single model: selecting a strategy, choosing each action, formatting it, and judging whether the result is adequate. No component outside the generative loop can reject the loop’s output, so an agent that cannot make progress does not report failure; it continues until an externally imposed budget terminates it. We propose treating ve