让对话智能体自动识别并优先放松用户不重要的查询约束。
AWARE-US: Preference-Aware Infeasibility Resolution in Tool-Calling Agents
- 基于对话上下文用大模型推断约束重要性顺序
- 局部加权法在用户偏好匹配度上达48%
- 构建120+带人物设定的查询基准测试
工具调用对话智能体在查询结构化数据库时常面临两个关联问题：约束不足（缺少精确查询所需条件）和不可行性（完整查询返回空集）。现有系统通常回应“无结果”或进行随意松弛，可能违背用户意图而丢弃高价值要求。本文将不可行性处理视为偏好感知的查询修复：当查询无法满足时，应放松最重要性最低的约束。提出三种基于大模型的方法推断约束相对重要性：(1) 局部加权，(2) 全局单次加权，(3) 成对排序。在汽车推荐任务中，经监督微调与直接偏好优化训练的局部加权法最符合用户偏好（48%），全局加权法正确松弛准确率达56%；三者均优于先前基线。同时引入AWARE-US基准，包含120+基于人物设定的查询，要求智能体通过对话澄清初始请求，并以符合人物隐含偏好的方式解决不可行性。代码见GitHub：https://github.com/mhtkrmz/Infeasible-task，数据集可在Hugging Face获取。
原文摘要 · Abstract (English)
Tool-calling conversational agents querying structured databases often face two linked failures: underspecification (missing constraints needed for a precise query) andinfeasibility (a fully specified query returns anemptyset). Prior systems often respond with "no results" or apply ad hoc relaxations, which can violate user intent by discarding highly valued requirements. Wecast infeasibility handling as preference-aware query repair: when a query is unsatisfiable, the agent should relax the least important constraints. We propose three LLM-based methods to infer relative constraint importance from dialogue: (1) local weighting, (2) global one-shot weighting, and (3) pairwise ranking. Across extensive experiments in car recommendation, the local-weighting method trained with supervised fine-tuning and direct preference optimization best aligns with user preferences (48%), while global weighting achieves the highest correct-relaxation accuracy (56%); all three outperform prior infeasibility-resolution basel. We also introduce AWARE-US, a benchmark of 120+ persona-grounded queries requiring agents to (i) disambiguate a base request via conversa tion and (ii) resolve infeasibility in a way consistent with persona-implied preferences. For code refer to Github: https://github.com/mhtkrmz/Infeasible-task and the dataset is available on Hugging Face
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。