← ProCompNav project page

Describe It vs. Just Say Yes

The same person, the same cabinet, the same robot asking two questions. One robot asks them to describe it; the other asks yes/no. Which one would you rather answer?

0.0 selapsed (real time)
  1. 1 · Two questions
  2. 2 · Still typing…
  3. 3 · What people chose

STUDY 1 · RESPONSE FORMAT · 20 PARTICIPANTS

Every participant preferred yes/no.

Preference

20 of 20 preferred yes/no
yes/nodescribe it (none)

Preferred: yes/no 20 of 20 participants (describe it 0 of 20)

Median response time per question

Median response time per question: 9.25 s describing vs 2.69 s yes/no

Real people answer open questions briefly: 89.1% of the Korean open-ended responses contained at most three words. Simulated users write long answers instead.

On CoIN-Bench Val Seen, simulated users wrote 460.2 response tokens per episode for open-ended questions (Pooled Independent Matching) vs 4.2 with ProCompNav. Yes/no questions keep real and simulated users aligned.

So ProCompNav asks short yes/no questions: one tap per answer.

Effort meters and curves are illustrative, not measured. The open-ended example questions and answers are illustrative; the yes/no questions follow the paper's running example. Answer durations use the Study 1 medians.