NeurIPS 2026 · Oral

Agentic robotics · Proactive agents

ProCompNav

Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries

Junhyuk Kwon1, Seungjoon Lee2, Hyejin Park1, Kyle Min3, Jungseul Ok1,2†

1GSAI, POSTECH 2CSE, POSTECH 3Oracle

†Corresponding author

NeurIPS 2026 Poster Session 5 · Thu Dec 10, 10:00 AM – 1:00 PM AEDT · Hall 1-4, ICC Sydney

TL;DR A proactive robot should take the initiative without becoming a burden. ProCompNav explores without interrupting you, then asks a few yes/no questions that split the candidates: higher success on ambiguous requests, with far less user effort.

Can you beat the robot?

Fewer yes/no questions wins.

Play: You vs ProCompNav
You vs ProCompNav game preview

Why proactive?

Proactive, with minimal user effort

Digital agentsdaily software tasks
Agentic robotsdaily physical tasks

Human requests are rarely complete, so agents must ask clarifying questions.

How should a robot ask, so the user does the least work?

  1. Dark wood one, by the mirror. OK.

    Follows commands

    You spell out everything

  2. What color is it? Where is it? What's it made of? What's next to it? What's on top?

    Proactive, but burdensome

    Question after question

  3. Wood-tone? Yes TV on top? No

    ProCompNavProactive, with minimal user effort

    Collect first, then yes/no

Interrupt Often vs. Ask Once

Open full screen ↗
Asking each time pulls the user away. ProCompNav asks once, after collecting. 14 of 20 people preferred it.

Describe It vs. Just Say Yes

Open full screen ↗
Describing takes effort; yes/no takes one tap. All 20 people preferred yes/no.

Overview

Abstract

In human-robot interaction, agentic robots should resolve ambiguous user requests through proactive clarification while keeping user burden low. We study this challenge in interactive instance navigation, where an agent must identify and reach the user-intended target among similar objects. Existing approaches ask open-ended questions and select a candidate that strongly matches the accumulated user-provided information. However, this process can require lengthy user responses and lead the agent to make premature decisions without distinguishing the target from distractors.

To address these limitations, we propose Proactive Instance Navigation with Comparative Judgment (ProCompNav), a two-stage framework that first constructs a candidate pool and then identifies the target through Recursive Comparative Judgment (RCJ). RCJ iteratively narrows the pool by selecting an attribute-value pair that divides the candidates, asking the user a binary question, and removing inconsistent candidates, without requiring an attribute unique to the target.

ProCompNav substantially improves Success Rate and reduces Response Length on an interactive navigation benchmark, and achieves the highest Success Rate among evaluated methods on a non-interactive benchmark. In two human studies, participants preferred ProCompNav’s binary questions to open-ended questions and its batched delivery to interleaved interaction. Our results highlight the need to jointly consider task performance and user-friendly interaction strategies when designing agentic robots.

ProCompNav teaser figure
Figure 1. Three ways to handle an ambiguous request. (a) Independent matching checks each candidate alone. (b) Pooled independent matching collects candidates first, then still checks each alone. (c) Comparative judgment (ours) collects candidates, then asks yes/no questions that split them.

Method

Collect first, then compare

  1. STAGE 1

    Pool construction

    Explore before deciding. Every object of the requested category becomes a candidate.

  2. STAGE 2

    Recursive Comparative Judgment

    Ask one yes/no question about an attribute that splits the candidates, and keep the side that fits. Repeat until one is left.

  • Easier for the robot

    Find one difference among the candidates, instead of matching each one to a vague target.

  • Easier for the user

    Answer yes or no, asked once after collecting.

How ProCompNav Works

Open full screen ↗
Collect candidates, ask about the attribute that best splits them, and drop the group that doesn’t fit.
ProCompNav method overview
Figure 2. Recursive Comparative Judgment. Split the pool into a core set and the rest, ask whether the target has an attribute common to the core set, and keep the side that matches.

Comparison

Three strategies for ambiguous instance navigation

  • PRIOR WORK

    Independent matching

    Asks open-ended questions and checks each candidate alone. Long answers, and it can stop at a look-alike.

  • NATURAL FIX

    Pooled independent matching

    Collects candidates first, but still checks each one alone. Answers get even longer.

  • OURS

    Comparative judgment

    Collects first, then asks yes/no questions that split the candidates.

Same Request, Three Robots

Open full screen ↗
One request, “Find the cabinet,” handled three ways.

Same candidate pool, different questions: comparing raises success and cuts the user’s answers from hundreds of tokens to about four.

Decision Strategy Stage Val Seen Val Seen Synonyms Val Unseen
PoolCompare SR↑RL↓NQ↓ SR↑RL↓NQ↓ SR↑RL↓NQ↓
(a) Independent Matching [Taioli et al., 2025] ✗✗ 10.5109.51.2 15.3129.21.2 8.9122.81.3
(b) Pooled Independent Matching ✓✗ 17.5460.24.1 22.0519.24.3 13.3467.74.0
(c) Comparative Judgment (Ours) ✓✓ 23.74.22.2 28.14.32.2 17.04.22.3

Table 1. Three strategies on CoIN-Bench. SR: success rate (%). RL: user response length per episode (tokens). NQ: questions per episode with dialogue.

Qualitative comparison

Asking for a description gets a long answer and can still pick a look-alike. Asking what splits the candidates gets short answers and the right dresser.

Qualitative comparison: independent matching vs ProCompNav
Figure 4. “Find the dresser.” Left: independent matching. Right: ProCompNav.

Results

Higher success, far less user effort

23.7% SR

Success Rate on CoIN-Bench Val Seen

vs 17.5 for Pooled Independent Matching and 10.5 for AIUTA*

4.2tokens

User response length per episode

vs 460.2 for Pooled Independent Matching (Val Seen)

20/20

participants preferred yes/no questions

over open-ended questions

Main results on CoIN-Bench

Highest success rate on every CoIN-Bench split, with only the object category as input.

Method Condition Val Seen Val Seen Synonyms Val Unseen
JudgeInteract SR↑SPL↑ SR↑SPL↑ SR↑SPL↑
Training-based
Monolithic-GOAT [Khanna et al., 2024]† –✗ 6.63.1 13.16.5 0.20.1
PSL [Sun et al., 2024]† –✗ 8.83.3 8.92.8 4.61.4
Training-free (description input)
3D-Mem* [Yang et al., 2025] Comp✗ 15.911.0 18.913.6 5.73.9
Context-Nav [Jang and Kim, 2026]‡ Indep✗ 13.56.7 20.310.9 11.35.2
Training-free (category input)
VLFM [Yokoyama et al., 2024]† Indep✗ 0.40.3 0.00.0 0.00.0
AIUTA [Taioli et al., 2025]† Indep✓ 7.42.9 14.48.0 6.72.3
AIUTA* Indep✓ 10.54.5 15.38.4 8.94.0
Pooled Independent Matching Indep✓ 17.55.1 22.08.1 13.35.1
ProCompNav (Ours) Comp✓ 23.77.0 28.18.5 17.06.2

Table 2. Results on CoIN-Bench. Indep/Comp: independent/comparative judgment. *Our reproduction with the same models as ProCompNav. †From [Taioli et al., 2025]. ‡From [Jang and Kim, 2026].

Termination-step analysis

Pooling stops early mistakes; comparing turns the collected candidates into more successes.

Tap a numbered marker to see what it shows. Hover or tap a bin for its numbers.
Figure 3. When episodes end (bars, left axis) and cumulative successes (lines, right axis), CoIN-Bench Val Seen. The last bin is the 500-step limit.

Non-interactive setting: TextNav

With no one to ask, ProCompNav reads attributes from the text description and still compares candidates. It reaches the highest success rate on TextNav.

Method Judgment SR↑ SPL↑
Training-based
Modular-GOAT [Khanna et al., 2024]†–17.08.8
PSL [Sun et al., 2024]†–16.57.5
Training-free
UniGoal [Yin et al., 2025]†Indep20.211.4
AIUTA* [Taioli et al., 2025]Indep10.92.7
3D-Mem* [Yang et al., 2025]Comp14.19.6
Context-Nav [Jang and Kim, 2026]‡Indep26.29.1
ProCompNav (Ours)Comp28.56.9

Table 3. Results on TextNav (no questions). Indep/Comp: independent/comparative judgment. †From [Yin et al., 2025]. ‡From [Jang and Kim, 2026]. *Our reproduction with the same models as ProCompNav.

Human Study

Human evaluation of interaction strategies

Two studies, 20 people each. Everyone preferred yes/no questions, and 14 of 20 preferred being asked once.

MetricStudy 1: response formatStudy 2: question delivery
Open-endedBinaryInterleavedBatched
Response time9.25 s2.69 s6.27 s5.28 s
Preference0/2020/206/2014/20

Table 4. Median response time per question, and how many people preferred each condition.

See both in Why proactive? above.

Cite

BibTeX

@inproceedings{kwon2026proactive,
  title={{ProCompNav}: Proactive Instance Navigation with Comparative Judgment for Ambiguous User Queries},
  author={Kwon, Junhyuk and Lee, Seungjoon and Park, Hyejin and Min, Kyle and Ok, Jungseul},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026},
  url={https://arxiv.org/abs/2605.06223}
}

Acknowledgements

This codebase is built upon VLFM and AIUTA / CoIN. We thank the authors for their excellent work and for releasing their code and benchmarks.