conf-preference-elicitation
Conference Preference Elicitation
The instinct when personalizing a schedule is to interview the attendee — ask them to rate every theme, list their goals, fill out a form. That is the wrong shape. With a good latent structure underneath (the theme map), a couple of well-chosen questions capture most of the signal; the literature shows large personalization gains after only two. The agent's intelligence belongs in which question to ask next, not in how many to ask. A person will answer three sharp, concrete questions gladly and resent ten vague ones — and the ten vague ones model them worse.
So this skill does the opposite of a survey. It holds the person's preferences as an uncertainty region over the themes, and at each step asks the one question that would shrink that region the most. It asks by showing concrete contrasting talks ("would you go to A or B?") and reading the latent dimension off the pick, because what people choose is truer than what they say they like. It deliberately spends some questions exploring themes it cannot yet place the person in, rather than only confirming the obvious. And it treats fighting selection bias as a hard requirement, not a nicety: if it only ever shows popular, central talks, it learns a narrowed caricature of the person and never surfaces the niche session they would have loved — so it deliberately offers outliers.
It is grounded: before it asks anything, it reads the cluster map and representative talks, so every question is anchored in real sessions this conference is actually running. And it is interactive: it must run where it can put the question to the person and read the answer — the main conversation, not a fire-and-forget background task.