PapersSelected

CMU 论文:AI 助手展示真实选项比询问偏好更容易获得用户接受

Why we picked it

CMU 用数据说话:加一个隐藏权衡,方案接受率从 12% 干到 55%。Claude Code 这类主流智能体还在傻问不展示。做 Agent 交互的都该看看,一条规则就能让产品可用性翻几倍。

Carnegie Mellon 的研究团队测试了面向旅行、购物等任务的 AI 助手,发现给用户直接展示几个真实选项,比先询问偏好再执行的效果更好。用户口头提出的硬性条件往往隐藏着未说出口的权衡,比如"评分 3.8 以上"会排除一个评分 3.7 但更便宜、用户其实愿意接受的餐厅。

Body

实验数据显示,在请求中加入一个隐藏权衡后,同一规划器产出的方案被用户接受的比例从 12% 升至 55%。而当前主流智能体如 Claude Code、Gemini CLI 的做法以提问为主、很少主动展示方案,最多只能命中 32% 用户会接受的计划;展示方案的智能体明显优于提问式智能体。

这项研究的实用启示是:智能体在把请求固化为计划前,应尽早展示几个略超出用户明示限制的真实选项,让用户在看到具体选项时自行发现权衡。对做 Agent 产品和交互设计的开发者而言,这提供了一条可以用数字验证的设计原则。

Original · rohanpaul_ai

New Carnegie Mellon paper finds that AI assistants that plan trips and shopping find more options people would accept by showing them a few real choices than by asking about their preferences.

A travel request like “restaurants rated 3.8 or higher” throws out a cheaper place rated 3.7 that the user would happily pick. People rarely say they would accept a lower rating to save money, because they only realize it when they see the actual options.

> Adding that 1 hidden trade-off to the request raised the share of plans the user accepted from 12% to 55%, with the same planner.

> Current agents (Claude Code, Gemini CLI and others) mostly asked questions and rarely showed plans, so they caught at most 32% of plans the user would accept.

> Agents that showed plans did clearly better than agents that asked.

When your agent turns a request into a plan, show the user a few options outside their stated limits early, before locking the spec.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Carnegie Mellon, Claude Code, Gemini CLI, agents.

Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).