小语言模型驱动数据库代理的失败原因
原标题：What Stops a Small Language Model From Driving a Database Agent
AI 摘要
该研究在生产级开源 SQL 客户端 LibreDB Studio 的 agent 模式上，用 39 个本地开源模型和 1 个托管对照模型进行了 11 天、8199 次运行、110711 条账本事件的测试。结果发现，2100 次归因于模型的 agent 模式失败中，75.7% 来自至少调用过一次工具的运行，说明失败主因并非模型能力不足，而是模型与服务器之间的传输/接口问题。研究还指出无上下文上限和交换压力两个测量混淆因素，并释放了语料、评分器和验证器。
正文节选
What Stops a Small Language Model From Driving a Database Agent Abstract Small open-weight language models are assumed to fail at agentic database work because they lack the reasoning capacity for it. We test that against a production system. Over eleven days we drove the agent mode of an open-source SQL client with 39 open-weight models served locally and one hosted control, across six task surfaces: 8,199 runs, 110,711 ledger events, 14,008 refused tool calls. Of the 2,100 model-attributed age