Proactive Preference Elicitation for Conversational Agents

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 151 papers63 conceptsfrozen 2026-10-04

In classical AI planning and recommendation, we often assume the user arrives with a perfectly formed set of requirements. We treat the user as a static database to be queried. However, as Sheila A. McIlraith and Jorge A. Baier noted in Planning with Preferences, real-world preferences are rarely pre-specified; they are discovered through the process of exploration.

When users interact with Large Language Models (LLMs), they often face a "specification gap." They know what they like when they see it, but they struggle to articulate the precise constraints—the "fuzzy" preferences—required for a perfect result. The emerging research direction in proactive preference elicitation shifts the burden of specification from the human to the AI, transforming agents from passive responders into active interviewers.

The Specification Bottleneck

Traditional systems are reactive. They wait for a prompt, guess the intent, and provide an answer. If the prompt is ambiguous, the system provides a generic response. This is particularly problematic in "cold-start" scenarios where the system has no historical data on the user.

Shuyue Stella Li and Avinandan Bose demonstrate in PrefDisco: Benchmarking Proactive Personalized Reasoning that even frontier LLMs tend to be "under-proactive." They often attempt to execute a task prematurely rather than asking the clarifying questions necessary to resolve ambiguity. This passivity leads to a discovery bottleneck, where users spend more time searching for the right way to ask than the AI spends solving the problem.

Generative Active Task Elicitation

To bridge this gap, researchers like Alex Tamkin, Belinda Z. Li, and Jacob Andreas introduced GATE (Generative Active Task Elicitation). Instead of relying on static instructions, GATE uses the LLM to interactively "interview" the user.

The intuition is simple: it is easier for a human to react to a specific scenario than to write a general rule. By presenting tricky edge cases or open-ended questions, the model surfaces latent preferences that the user might have overlooked. For instance, instead of asking "What kind of movies do you like?", a proactive agent might ask, "Would you prefer a mindless action movie for a stressful day, or something intellectually demanding?" This targeted probing, as explored in Eliciting Human Preferences with Language Models, captures nuanced constraints more efficiently than traditional prompt engineering.

Conversation as an Information-Seeking Game

How does an agent decide what to ask? This is increasingly framed as an optimization problem. In Optimal Incremental Preference Elicitation during Negotiation, Tim Baarslag and Enrico H. Gerding describe the need to balance the "cost of bothering" the user against the utility of the information gained.

Recent frameworks like CPER (Conversation Preference Elicitation and Recommendation), developed by Sarvesh Baskar and Manas Gaur, allow models to quantify their own uncertainty. If the model realizes it is "guessing" a user's persona, it triggers a clarification request.

Similarly, the MACO framework (A Multi-Agent Conversational Bandit Approach) treats conversation as a "bandit" process. Instead of random trial-and-error, the system selects high-level descriptors or "key terms" to query the user. This strategy rapidly shrinks the "preference uncertainty space," ensuring that every turn of the dialogue provides maximum information gain.

Modeling the Hidden Persona

The most sophisticated versions of this idea treat the user's mind as a "partially observable" state. In PAMDP: Interact to Persona Alignment via a Partially Observable Markov Decision Process, ZHE YANG and colleagues argue that agents should treat the initial turns of a conversation as "active probing steps."

By modeling the interaction as a POMDP, the agent learns to navigate the dialogue to reveal the user's hidden persona. This allows the system to adapt to a specific user's expertise and communication style within a single session, without needing a pre-existing profile. As Kirandeep Kaur and Hideo Joho suggest in their work on Personalized Agent Adaptability, this ability to re-contextualize preferences across multiple sessions is the key to building long-term, trustworthy AI partners.

The Power to Steer

While proactive elicitation solves the specification gap, it also introduces a new ethical dimension. Tobias Werner and Iyad Rahwan have shown that conversational AI can subtly steer consumer behavior without detection. Because the agent is the one asking the questions, it can shape the user's decision-making framework in real-time. This highlights the importance of "Digital Humanism" in recommender systems, as advocated by Julia Neidhardt, ensuring that proactive agents remain aligned with human values rather than just maximizing engagement.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_proactive_preference_elicitation_for_con_568716bf,
  title = {Proactive Preference Elicitation for Conversational Agents},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proactive-preference-elicitation-for-conversational-agents/rdr_568716bfcacb42eda3db1c92effefcd0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_proactive_preference_elicitation_for_con_568716bf_v1,
  title = {Proactive Preference Elicitation for Conversational Agents},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/proactive-preference-elicitation-for-conversational-agents/rdr_568716bfcacb42eda3db1c92effefcd0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}