THE CENTRAL IDEA核心观点
Do not make the model a better SQL writer; make it a disciplined evidence acquirer. Interpose a source-aware acquisition layer before SQL generation, schedule every query and question by cost-normalized priority, and let the model commit only when the acquired evidence is sufficient.不要让模型成为更好的 SQL 写手，而要让它成为有纪律的证据获取者：在 SQL 生成之前插入一层感知来源的获取层，按成本归一化优先级调度每一次查询与提问，只有当证据充分时才允许模型作答。
The challenge: the bottleneck has shifted挑战：瓶颈已经转移
In interactive text-to-SQL, the evidence needed to write a correct query is not given upfront. It is distributed across the database schema, external knowledge bases and the user, and must be acquired through environment interactions under a finite budget. The paper's central observation: strong LLMs often fail not from deficient SQL synthesis, but from premature commitment—they start writing SQL before acquiring task-critical evidence.在交互式 Text-to-SQL 中，写出正确查询所需的证据并不会在一开始给全：它分散在数据库模式、外部知识库和用户那里，必须在有限的交互预算下逐步获取。论文的核心观察是：强 LLM 失败的原因往往不是 SQL 合成能力不足，而是过早作答——在获取任务关键证据之前就开始写 SQL。
This reframes where capability work should go. As model synthesis ability keeps improving, the marginal failure moves upstream to acquisition: what to look up, who to ask, and when to stop gathering and commit. Teams that keep investing only in generation quality are optimizing the wrong term.这重新定义了能力建设的落点。随着模型合成能力持续增强，边际失败点已经上移到获取环节：该查什么、该问谁、何时停止收集并作答。只继续投资生成质量的团队，是在优化错误的项。
The Information Sufficiency Hypothesis信息充分性假设
Through controlled trajectory analysis, the authors formalize the observation as the Information Sufficiency Hypothesis: given sufficient task-relevant evidence, current models are largely capable of producing correct SQL; failures concentrate where evidence is missing. They complement the hypothesis with a five-class gap taxonomy that structures the acquisition space—making "what is missing" an explicit, classifiable object rather than a vague intuition.通过受控轨迹分析，作者把这一观察形式化为信息充分性假设：只要任务相关证据充分，现有模型基本能够生成正确的 SQL；失败集中在证据缺失之处。与假设配套的是一个五类缺口分类法，把获取空间结构化——让「缺什么」成为可以明确分类的对象，而不是模糊的直觉。
The taxonomy is the diagnostic deliverable: each failure can be attributed to a gap class, and each gap class implies a different acquisition action. Diagnosis before therapy—this is what turns an agent framework from a bag of prompts into an engineered system.这套分类法本身就是诊断交付物：每个失败都能归因到某类缺口，而每类缺口对应不同的获取动作。先诊断后治疗——这正是把智能体框架从一堆提示词变成工程化系统的关键。
Three mechanisms of PASCALPASCAL 的三个机制
Guided by this diagnosis, the authors build PASCAL, a training-free acquisition-control data agent. First, a source-aware information-acquisition layer interposed before SQL generation: the agent explicitly distinguishes schema lookups, knowledge-base retrieval and user clarification as different evidence sources with different costs. Second, cost-normalized priority scheduling: candidate acquisition actions are ranked by expected evidence value per unit of interaction cost, so the finite budget is spent on the most decisive gaps first. Third, training-free composition: PASCAL wraps existing LLM backbones without fine-tuning, so the framework transfers across models as they improve.在这一诊断指导下，作者构建了 PASCAL——免训练的获取控制数据智能体。第一，感知来源的信息获取层插入 SQL 生成之前：智能体明确区分模式查询、知识库检索与用户澄清，把它们当作成本不同的证据来源。第二，成本归一化的优先级调度：候选获取动作按「单位交互成本的预期证据价值」排序，有限预算优先花在最决定性的缺口上。第三，免训练组合：PASCAL 包裹现有 LLM 基座而无需微调，模型升级时框架随之平移。
The training-free property matters as much as the scheduling. An acquisition layer that requires fine-tuning would have to be retrained for every backbone upgrade; a control layer that lives outside the model inherits each new generation of models for free.免训练这一属性与调度同等重要：需要微调的获取层在每次基座升级时都得重训，而位于模型之外的控制层可以免费继承每一代新模型。
Results: a local 35B beats frontier models结果：本地 35B 超越前沿模型
On BIRD-Interact, PASCAL with a local 35B backbone achieves 51.0% Phase-1 and 37.0% two-phase success rate, outperforming the frontier-model leaderboard results by up to 14.7 percentage points. The comparison carries the paper's thesis in one line: a mid-size local model with disciplined acquisition beats frontier models that commit early.在 BIRD-Interact 上，PASCAL 配合本地 35B 基座取得 51.0% 的一阶段成功率和 37.0% 的两阶段成功率，最高超出前沿模型榜单结果 14.7 个百分点。这组对比用一句话承载了论文主旨：有纪律地获取证据的中等规模本地模型，胜过仓促作答的前沿模型。
For deployment, the 35B result is the headline: acquisition control is exactly the kind of capability that lets an on-premises or onboard model punch above its weight class, without sending data to a frontier API.对部署而言，35B 这个结果才是重点：获取控制正是让本地或星载模型越级发挥的那类能力——无需把数据交给前沿 API。
Where the gains concentrate增益集中在哪里
Gains concentrate on BI tasks requiring cross-source evidence acquisition—tasks where the answer depends on combining schema facts, external knowledge and user intent. Where evidence is fully present in the prompt, acquisition control adds little; where evidence must be gathered, it is decisive. The gains map is thus itself a guide: it tells adopters which workloads justify an acquisition layer and which do not.增益集中在需要跨来源获取证据的 BI 任务上——答案依赖模式事实、外部知识与用户意图的组合。证据已在提示中齐备的任务，获取控制带来的提升有限；证据必须主动收集的任务，它是决定性的。因此这张「增益分布图」本身就是采用指南：告诉工程团队哪些负载值得上获取层，哪些不必。
A useful mental model: PASCAL does not raise the model's ceiling; it raises the floor of the evidence the model stands on. The two are easy to confuse in aggregate benchmarks and must be separated when planning adoption.一个有用的认知模型：PASCAL 并不抬高模型的能力上限，它抬高的是模型脚下证据的地基。在汇总榜单上两者容易混淆，做采用规划时必须分开。
Applicability and boundaries适用条件与边界
The framework presumes an interactive setting with a queryable environment: databases that can be probed, knowledge bases that can be retrieved, users who can clarify. Where the environment cannot answer—or answers unreliably—priority scheduling cannot manufacture evidence. The five-class taxonomy, likewise, is a structuring of observed gaps rather than a completeness proof; new deployment domains should expect to extend it.该框架以「存在可查询的交互环境」为前提：数据库可探查、知识库可检索、用户可澄清。当环境无法回答或回答不可靠时，优先级调度无法凭空制造证据。同样，五类缺口分类是对已观察缺口的结构化，而非完备性证明；进入新的部署领域时应预期对它扩展。
The authors are careful not to claim a universal agent recipe: the evidence supports acquisition control for interactive data tasks on the evaluated benchmark and backbones, and the design principle—diagnose the gap, then schedule acquisition by cost-normalized priority—is the transferable part.作者没有声称通用的智能体配方：证据支持的是在所评基准与基座上、面向交互式数据任务的获取控制；可迁移的是设计原则本身——先诊断缺口，再按成本归一化优先级调度获取。
Engineering mapping and evaluation boundaries工程映射与评估边界
Mapped to an operational data stack, the acquisition layer sits between the user's question and the generation model: it reads the gap taxonomy, prices candidate probes against the remaining interaction budget, and only releases the question to SQL generation when evidence is sufficient. For space-operations data agents—querying telemetry stores, mission databases and procedure manuals—the same discipline applies verbatim: acquire first, commit second, and never let a fluent answer outrun its evidence.映射到业务数据栈：获取层位于用户提问与生成模型之间——读取缺口分类，用剩余交互预算为候选探查定价，只有当证据充分时才把问题放行给 SQL 生成。对航天运控数据智能体——查询遥测库、任务数据库与规程手册——同样的纪律可以逐字照搬：先获取，后作答，绝不让流利的答案跑赢它的证据。
A concrete evaluation plan for adoption: instrument the target workload with the gap taxonomy and measure the current failure mix; deploy the acquisition layer in shadow mode, logging which probes it would have scheduled; then compare success rate and interaction cost per solved task against the incumbent pipeline, reporting per-gap-class outcomes rather than only pooled averages. Only after the acquisition policy survives workloads the taxonomy never saw should it take over the live path.面向采用的可执行评估方案：先用缺口分类法对目标负载做插桩，测出当前的失败构成；以影子模式部署获取层，记录它本会调度哪些探查；随后按「每解决一个任务的成功率与交互成本」与现有流程对比，并按缺口类别分别报告而非只给合并均值。只有当获取策略在分类法从未覆盖的负载上依然成立，它才应接管在线链路。
SOURCE PAPER & REFERENCE MATERIAL原论文与参考资料
When the Bottleneck Shifts: Diagnosing and Closing Information Gaps in Interactive Text-to-SQL
Mengyi Yan, Zhinian Ma, Yue Wang, Yaoshu Wang, Xiaoke Zhu, Zizhe Wang, Haoyi Zhou, Zhumin Chen
Proceedings of EMNLP 2026 (Main Conference) · ACL Anthology 2026.emnlp-main.465