Careful Queries, Credible Results¹: Teaching RAG Models Advanced Web Search Tools² with Reinforcement Learning³

谨慎查询，可信结果¹：通过强化学习³教会 RAG 模型使用高级网页搜索工具²。

Document-based dialogue requires answers that respect both source evidence and evolving user constraints¹. Passage-only generation supplies relevant text without explicitly recording which dialogue constraints the source quotes support². For this missing evidence-to-constraint link, we introduce structured evidence notes that pair source quotes with manual locations and active dialogue constraints.³ These notes let the generator cite sources and explain how the evidence relates to each constraint. In our comparison on static English manuals, notes increase citation support relative to direct generation but add latency⁴. The comparison evaluates source support; the support scores do not establish that every user constraint is satisfied⁵.

基于文档的对话要求答案同时符合来源证据与不断变化的用户约束¹。只使用片段的生成方式提供相关文本，却不明确记录来源摘句支持哪些对话约束²。针对证据与约束之间缺少对应记录的问题，我们提出结构化证据笔记，将来源摘句与手册位置及当前有效的对话约束配对。³这些笔记帮助生成器引用来源，并说明证据与各项约束的关系。在针对静态英语手册的比较中，相比直接生成，笔记提高了引用支持率，但增加了延迟⁴。这个比较评价的是来源支持，不能据此认定每项用户约束都得到满足⁵。

However, answering this question faces two obstacles.¹ First, data scarcity limits access to real commercial conversations, and identifying complex or ambiguous needs requires costly expert annotation². Second, case-dependent standards complicate evaluation: clothing recommendations require color and outfit coordination, whereas phone recommendations require different suitability checks. Even within one category, requirements and risks vary across cases.³ A fixed checklist can miss these differences. Together, these obstacles hinder accurate evaluation of LLMs on complex, ambiguous, and time-sensitive shopping requests.⁴

然而，回答这个问题有两个困难。¹首先，数据稀缺使真实商业对话难以获取，识别复杂或模糊的需求还需要成本较高的专家标注²。其次，随案例变化的标准增加了评价难度：服装推荐要考虑颜色与穿搭搭配，手机推荐则要检查其他适配条件。即使是同类商品，不同案例的需求与风险也会变化。³一张固定清单可能漏掉这些差异。这两类困难共同妨碍了对 LLM 处理复杂、模糊且有时效性的购物请求进行准确评价。⁴

Therefore, to address these obstacles, we introduce [Benchmark], an evaluation resource that preserves changing needs and unresolved conditions. Business experts screen cases in de-identified shopping conversations. Each case retains timestamped dialogue, including changed needs and unanswered questions. Experts distill Ideal Response Principles¹ from excellent replies manually selected from completed transactions². These principles guide case-specific criteria across Core, Excellence, and Safety³ and inform a tailored reasoning prompt. The tiers assess required completion, added value, and risk avoidance, respectively. Our example (Figure [X]) illustrates how elapsed time and unresolved information affect the advice a case requires.⁴

因此，为处理这些困难，我们提出【Benchmark】，一个保留变化需求与未决条件的评价资源。业务专家从去标识化的购物对话中筛选案例。每个案例保留带时间戳的对话，包括发生变化的需求和未得到回答的问题。专家提炼出理想回复原则¹，依据的是从成功交易中手工筛选出的优秀回复²。这些原则指导针对每个案例制定 Core、Excellence 和 Safety 三层标准³，并用于设计专门的推理提示词。三层分别评价必要要求的完成、额外价值和风险规避。我们的例子（图【X】）说明时间推移和未决信息如何影响一个案例所需的建议。⁴

To answer this question¹, we evaluate [commercial models] and [open-weight models]. Across these models, we observe Product Evidence Overreach, where replies add unsupported product claims, and Shopping Uncertainty Collapse, where unresolved needs are treated as settled.² These failures may make shopping advice less reliable. Further analysis links these failures to changing requirements and unresolved information.³

为回答这个问题¹，我们评价【商业模型】与【开放权重模型】。在这些模型的回复中，我们观察到两类失败：Product Evidence Overreach，即回复添加了缺少证据支持的商品声明；Shopping Uncertainty Collapse，即将尚未确定的需求视为已经明确。²这些失败可能降低购物建议的可靠性。进一步分析将这些失败与需求变化和未决信息联系起来。³

Humans spend nearly 90% of their time indoors [12], and the built environment contributes a substantial share of global carbon emissions [34, 65, 66]. Thus, building design sits at the nexus of daily life and sustainability¹. Traditional design relies heavily on expert knowledge and extensive manual iteration² [2, 14, 20], often taking days to months to reach viable solutions. Recent advances in generative models on controllable [5, 32, 47] and multimodal generation [7, 24, 50] pave the way for automated building design generation³.

人们近 90% 的时间在室内度过 [12]，建筑环境也占全球碳排放的相当大一部分 [34, 65, 66]。因此，建筑设计连接着日常生活与可持续性¹。传统设计高度依赖专家知识和大量人工迭代² [2, 14, 20]，往往需要数天至数月才能得到可行方案。近期，可控生成 [5, 32, 47] 和多模态生成 [7, 24, 50] 的进展，为建筑设计的自动生成铺平了道路³。

To enable automation, prior works have employed conditioned generators [22, 31, 37, 38, 44, 48, 49, 59, 60, 63, 64] that translate coarse user requirements into concrete floorplan layouts¹. However, existing methods rely solely on user-specified conditions and thus often violate unspecified real-world design constraints². This neglect results in over 60% of samples in existing datasets [48] being practically unusable due to excessive energy use and flawed spatial-functional design. Some methods [38, 44] even fail to generate essential structural components such as windows, leading to unrealistic energy evaluations. Therefore, enforcing design constraints in generative models³ remains a major bottleneck for practical deployment.

为实现自动化，已有工作采用条件生成器 [22, 31, 37, 38, 44, 48, 49, 59, 60, 63, 64]，将用户的粗略需求转为具体平面布局¹。然而，现有方法仅依赖用户明确给出的条件，因此往往违反未被明确提出的现实设计约束²。这种忽视使现有数据集 [48] 中超过 60% 的样本因能耗过高、空间与功能设计不合理而无法实际使用。一些方法 [38, 44] 甚至无法生成窗户等必要结构，导致能耗评估不符合现实。因此，在生成模型中落实设计约束³，仍是实际部署的主要瓶颈。

Evidence-grounded dialogue systems select supporting text before generating an answer. [A] retrieves passages for the current question, whereas [B] links retrieved passages to answer claims¹ [citations]. These claim links make support easier to inspect², but [B] is evaluated only on isolated questions. However, these comparisons do not establish whether earlier constraints remain satisfied across dialogue turns.³ To address this gap, we construct constraint-linked evidence notes based on a constraint-to-evidence linking rule.⁴ This enables evaluation of constraint retention across turns, addressing the lack of evidence about whether earlier requirements remain satisfied.⁵

基于证据的对话系统先筛选支持文本，再生成答案。[A] 为当前问题检索片段，[B] 则把检索片段与答案主张关联¹[引用]。这些主张关联便于检查依据²，但 [B] 的评价仅覆盖独立问题。然而，这些比较尚未说明前文约束能否在多轮对话中持续得到满足。³为应对这个缺口，我们基于约束与证据的关联规则构建绑定约束的证据笔记。⁴这支持跨轮约束保留评价，应对前文需求能否持续得到满足缺少检验证据的问题。⁵

Multi-dancer generation is an emerging area currently in its nascent stage. To the best of our knowledge, only a handful of studies (Yao et al. 2023; Le et al. 2023b,a; Yang et al. 2024) have focused on generating scenarios for more than two dancers. Among these, GDanceR (Le et al. 2023b) and GCD (Le et al. 2023a) utilize no special structures to avoid the imbalance issue in motion representation, resulting in a tendency for dancer ambiguity¹. CoDancers (Yang et al. 2024) splits group motions into single-person motions², preventing the occurrence of dancer ambiguity. However, completely isolating individual features results in incomplete group information³, which leads to disharmony generation results. Therefore, a more optimal feature separation strategy is necessary⁴. In this work, we introduce TCDiff, a method that first generates dancers’ trajectories and then produces logical movements⁵. Together with our proposed effective plugin, Fusion Projection, this approach significantly reduces dancer ambiguity.

多舞者生成是一个仍处于早期的新兴方向。据我们所知，只有少量研究（Yao et al. 2023; Le et al. 2023b,a; Yang et al. 2024）关注超过两名舞者的场景。其中，GDanceR（Le et al. 2023b）和 GCD（Le et al. 2023a）没有采用专门结构来避免运动表示的不平衡，因而容易出现舞者混淆¹。CoDancers（Yang et al. 2024）将群体动作拆成单人动作²，避免了舞者混淆。然而，完全隔离个体特征会造成群体信息不完整³，进而导致生成结果不协调。因此，需要更合适的特征分离策略⁴。本文提出 TCDiff，先生成舞者轨迹，再生成合乎逻辑的动作⁵。该做法结合我们提出的有效插件 Fusion Projection，显著减少了舞者混淆。

Given a question, the dialogue history, and a collection of English manuals¹, the task is to generate an answer that respects earlier constraints and cites supporting passages². At inference time, the system receives these inputs but no reference answer³. We use reference answers only for evaluation. A prior requirement such as offline operation remains part of the task unless the user explicitly changes it⁴.

给定一个问题、它的对话历史和一组英语手册¹，任务是生成遵守前文约束、附有支持片段引用²的回答。推理时，系统接收这些输入，但不接收参考答案³。参考答案仅用于评价。离线操作等前文要求仍属于当前任务，除非用户明确更改⁴。

To retain earlier constraints during answering¹, we connect quoted evidence to the dialogue history before generating an answer. In History-aware Retrieval (Sec. [retrieval]), we select² manual passages relevant to the current question and earlier requirements. Using these passages³, in Constraint-linked Notes (Sec. [notes]), we attach the requirements to source quotes and retain their manual locations. Finally, in Grounded Answering (Sec. [answering]), the generator uses the resulting notes⁴ to produce an answer with source citations. Figure [framework] shows these dependencies.⁵

为在回答中保留前文约束¹，我们先把摘录证据与对话历史关联，再生成答案。在历史感知检索（第[retrieval]节）中，我们筛选²与当前问题及前文需求相关的手册片段。使用这些片段³，我们在约束关联笔记（第[notes]节）中将需求绑定到来源摘句，并保留手册位置。最后，在基于证据的回答（第[answering]节）中，生成器使用所得笔记⁴生成带来源引用的答案。图[framework]展示这些依赖关系。⁵

To keep earlier requirements available during answering¹, the note builder links each retrieved quote to the relevant dialogue constraint. A retrieved passage may describe a feature correctly yet be unsuitable for an offline-only request². For each quote, we record the manual location and the constraint supported or contradicted by the quote³; an unverified relation is left unresolved⁴. The generator then receives these notes together with the current question⁵, so the answer can cite the source and explain whether the requirement is met.

为让回答阶段仍能使用前文要求¹，笔记构建器将每条检索摘句与相关对话约束关联。一个片段可能准确描述功能，却不适用于仅限离线的请求²。对每条摘句，我们记录它在手册中的位置，以及它支持或违背的约束³；尚未核验的关系保留为未确定⁴。生成器随后接收这些笔记和当前问题⁵，以便回答引用来源，并解释要求是否满足。

We propose GreenPlanner, an energy- and functionality-aware generative framework for residential design¹ (Fig. 2). It integrates four components: (1) Design Feasibility Dataset (DesignFD), a dataset built from RPLAN [48] via expert-verified energy and functional evaluations²; (2) Practical Design Evaluator (PDE), a predictor trained on DesignFD for fast simulation of energy and functional feasibility³; (3) Green Plan Dataset (GreenPD), a PDE-guided dataset pairing user requirements with regulation-compliant layouts⁴; and (4) GreenFlow, a generator trained on GreenPD with PDE feedback⁵ for controllable, constraint-grounded layout generation. We next describe each component of the framework in detail.

我们提出 GreenPlanner，一套同时考虑能耗与功能的住宅设计生成框架¹（图 2）。该框架集成四个组件：（1）设计可行性数据集（DesignFD），由 RPLAN [48] 经专家核验的能耗与功能评估²构建；（2）实用设计评估器（PDE），在 DesignFD 上训练，用于快速模拟能耗与功能可行性³；（3）绿色规划数据集（GreenPD），由 PDE 引导，将用户需求与符合规范的布局配对⁴；（4）GreenFlow，在 GreenPD 上训练并使用 PDE 反馈⁵，生成可控且以约束为依据的布局。下面将详细介绍框架的各个组件。

To test whether evidence notes improve citation support¹, we evaluate multi-turn questions about static English manuals. Human raters check each factual answer sentence² against its cited passage using a shared rule. A sentence without a citation counts as unsupported.³ For each question, citation support is the fraction of factual sentences supported by their cited evidence; higher is better. We average these per-question fractions rather than pooling sentences across questions.⁴ Answers with no factual sentences are reported separately as not applicable⁵, together with their count.

为检验证据笔记能否提高引用支持率¹，我们评价针对静态英语手册的多轮问题。指标。人工按同一规则检查每句事实²能否由所引片段支持。没有引用的句子按不支持计。³对每个问题，引用支持率是获所引证据支持的事实句比例，越高越好。我们对各问题的比例取平均，不把所有问题的事实句混在一起计算。⁴没有事实句的回答单列为不适用⁵，并报告数量。

Evidence notes improve citation support under a fixed evidence-context budget.¹ In the comparison in Table [main], the note-based system has higher support than direct passage retrieval when the language model, decoding settings, and evidence-context token budget are held fixed². The contrast supports note-guided generation as a whole³; separate ablations are needed to distinguish quote organization from constraint retention. However, the note builder adds processing time⁴, so the quality gain must be considered together with end-to-end latency.

在证据上下文预算固定时，证据笔记提高了引用支持率。¹在表[主结果]所示的对照中，固定语言模型、解码设置和证据上下文词元预算后²，笔记系统的支持率高于直接片段检索。这一对照支持该设置下的笔记引导生成整体³；区分摘句组织与约束保留的作用，还需要分别消融。不过，笔记构建增加了处理时间⁴，因此质量收益需要与端到端延迟一起考虑。

To examine the role of retained dialogue constraints¹, we remove those constraints from the notes while keeping source quotes unchanged². Both variants use the same language model, test questions, decoding settings, and maximum evidence-context budget³. If constraint satisfaction decreases after this single change⁴, the comparison supports retaining constraints in the tested setting. The comparison does not isolate the benefit of the entire note-building pipeline⁵, because source selection and quotation remain active in both variants.

为检验保留对话约束的作用¹，我们从笔记中移除这些约束，保留来源摘句²。两种设置使用相同的语言模型、测试问题、解码设置和证据上下文预算上限³。若这项单一改动后约束满足率下降⁴，该比较支持在实测设置中保留约束有用。它不能单独确定整套笔记构建流程的收益⁵，因为两种设置都仍在进行来源选择和摘句。

We highlight the importance of natural footwork in enhancing visual aesthetics¹ (Tseng, Castellon, and Liu 2023). As shown in Table 3², our Footwork Adaptor markedly enhances the model’s PFC value³. Moreover, when utilized independently⁴, it boosts the model’s GMC, underscoring the FA’s dual benefits: preventing foot slipping and augmenting dance coherence⁵.

我们强调自然脚步对提升视觉美感¹的重要性（Tseng、Castellon 与 Liu，2023）。如表 3 所示²，Footwork Adaptor 明显改善了模型的 PFC 指标³。此外，单独使用该模块时⁴，它也提高了模型的 GMC，体现出 FA 的双重作用：减少脚部滑动，并增强舞蹈协调性⁵。

Higher citation support is consistent with more explicit use of document evidence¹. However, the current comparison changes both quote organization and the amount of constraint information provided to the generator². The comparison therefore cannot identify which change accounts for the gain.³ Our evaluation covers static English manuals; cross-language questions and frequently updated documents remain untested⁴. Separate evaluations are needed because language and document updates can change which quoted evidence supports an answer⁵.

引用支持率提高，与更明确地使用文档证据这一解释相符¹。然而，当前比较同时改变了摘句的组织方式和提供给生成器的约束信息量²。因此，这个比较不能判断哪项改变带来了收益。³当前评价覆盖静态英语手册；跨语言问题与频繁更新的文档尚未检验⁴。这些设置需要另行评价，因为语言与文档更新可能改变哪些摘句能够支持答案⁵。

First, although we have [verified work addressing the main difficulty]¹, [specific special condition] still warrants further study.² Second, although we have [another verified contribution]³, [specific longer-term condition] still warrants further study.⁴

首先，尽管我们已完成[针对主要难点的有据工作]¹，[具体特殊情况]仍值得进一步研究。²其次，尽管我们已完成[另一项有据工作]³，[具体长期情况]仍值得进一步研究。⁴

We present EviNote-RAG, a framework that introduces a note-taking step to extract answer-supportive evidence before answering¹. By training LLMs to produce Supportive-Evidence Notes (SENs) and guiding them via a tailored entailment-based reward², our approach improves answer accuracy. Extensive experiments demonstrate that EviNote-RAG achieves state-of-the-art performance while enhancing training stability³.

我们提出 EviNote-RAG，在回答前加入笔记步骤，提取能够支持答案的证据¹。通过训练 LLM 生成能够支持答案的证据笔记（SENs），并用专门设计的蕴含奖励引导模型²，方法提高了答案准确率。大量实验表明，EviNote-RAG 达到当前最优表现，同时改善了训练稳定性³。

Retrieval-augmented generation combines retrieved passages with answer generation [primary method paper]¹. A study evaluated only on single-turn questions cannot establish performance on changing multi-turn constraints.² We use [retriever and version] [official documentation for that version]³ to retrieve candidate passages. The method citation supports the approach; the version-specific documentation identifies the tool actually used.⁴

检索增强生成将检索到的段落与答案生成结合起来[原始方法论文]¹。只在单轮问题上评价的研究，不能证明方法能够处理多轮对话中变化的约束。²我们使用[检索器与版本][该版本的官方文档]³检索候选段落。方法引用说明采用哪条路线；对应版本的文档说明实际使用什么工具。⁴

To reproduce the citation-support comparison in Section [X]¹, this appendix records the prompts, document versions, retrieval settings, and scoring rules for both systems². Each run is linked to the corresponding configuration and retains failed outputs. We select the maximum note length using development questions and freeze it before evaluating the test questions³. Per-run scores and uncertainty estimates are reported in Table [Y], so the aggregate in the main text can be traced to the same runs⁴.

为复现第[X]节的引用支持率比较¹，本附录记录两个系统的²提示词、文档版本、检索设置与评分规则。每次运行都关联对应配置，并保留失败输出。我们用开发集问题选择笔记最大长度，在评价测试问题前冻结这个选择³。表[Y]报告各次运行得分与不确定性估计，使正文的汇总结果能追溯到同一批运行⁴。