Beyond expert users: agents should help users construct preferences, not just elicit them

2026-06-29 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

The article introduces CoPref, a model for user preference construction, and CoShop, an interactive benchmark for agentic recommender systems. It argues that agents should help users construct preferences, not just elicit them, challenging the common assumption of expert users with well-formed preferences. The CoPref model formalizes these principles by drawing on the Search-Experience-Credence framework from Information Economics, detailing how users construct preferences based on agent dialog actions. CoShop concretely studies these ideas, evaluating an agent's performance based on its ability to help a CoPref user gain the necessary knowledge to specify a task well. Evaluation of five frontier models on CoShop revealed that no agent exceeded 56% accuracy despite five turns of interaction, with failures primarily stemming from the agents' limited ability to expand what users know about their own desires, rather than their capacity to find items.

Key takeaway

For AI Engineers designing conversational agents, recognize that users often need help forming preferences, not just expressing them. Your agent's success hinges on its ability to educate users with domain knowledge, such as via examples or explanations, rather than merely asking clarifying questions. Prioritize metrics that measure user knowledge expansion and consider using benchmarks like CoShop to rigorously evaluate this critical aspect of agent performance, moving beyond simple item retrieval accuracy.

Key insights

Agents should help users construct preferences by providing domain knowledge, not just elicit pre-formed ones.

Principles

• Users often lack well-formed preferences.
• Agents must help users learn domain knowledge.
• Agent performance depends on user knowledge expansion.

Method

CoPref models user preference construction based on agent dialog actions, drawing on the Search-Experience-Credence framework. CoShop is an interactive benchmark for evaluating this.

In practice

• Design agents to provide examples/explanations.
• Focus agent metrics on user knowledge gain.
• Evaluate agent systems with CoShop benchmark.

Topics

• Conversational AI
• Recommender Systems
• User Preference Construction
• CoPref Model
• CoShop Benchmark
• Agent Evaluation

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• The agent line keeps going up—and you need to identify what's below it. — Developing documentation agents requires curated question datasets and specific model/context configurations.
• KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation — Evaluating personalized mobile agents requires dynamic preference inference and proactive interaction, not just static context lookup.
• Mapping the Design Space of User Experience for Computer Use Agents — Understanding user experience is crucial for designing effective LLM-based computer use agents.
• Knowledge Graph Engineering For Agents: The Multidomain Problem — AI agents require explicit engineering to integrate multi-domain knowledge and accumulate persistent memory, unlike human pattern-matching.
• Better Instructions, Better AI Results — AI's unforgiving nature reveals human communication flaws, demanding clearer instructions.
• Prompt the agent - Perplexity — Distinguish agent "instructions" (persistent rules) from "input" (turn-specific task) for precise control and grounded retrieval.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.