Research map / Multi-agent systems and games / LLM agents
Agentic Model Predictive Questioning Control in Visual Design
Kuang-Da Wang, Zhao Wang, Wei-Yao Wang, Yotaro Shimose, Jaechang Kim, Shingo Takamatsu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Agentic Model Predictive Questioning Control (A-MPQC), a test-time framework for LLM-based visual design clarification that accounts for the user's cognitive burden when choosing questions.
Problem
LLM-based approaches to clarifying visual design intent focus on selecting informative questions but overlook the cognitive burden of interpreting and answering them, which matters for human-agent interaction.
Approach
A test-time, model-predictive control style questioning policy (A-MPQC) that plans which clarifying questions to ask while explicitly modeling the user's cognitive effort in answering them, rather than only maximizing intent uncovering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove question plans (lookahead) from A-MPQC (ablation) | A-MPQC · MIMO | ΔS: 0.594 -> 0.391 when removed | Improved |
| add respond-or-reject surrogate reward instead of always-answer (ablation) | A-MPQC · MIMO | ΔS: 0.204 -> 0.594 | Improved |
| use rejection feedback in replanning rather than ignoring it (ablation) | A-MPQC · MIMO | ΔS/C efficiency: 0.361 -> 0.480 | Improved |
| add adaptive question format control instead of direct generation (ablation) | A-MPQC · MIMO | ΔS/C efficiency: 0.155 -> 0.480 | Improved |
| apply MPC-style control over within-round question dimension instead of rounds (ablation) | A-MPQC · MIMO | interaction cost: 1.238 -> 6.696 | Hurt |
| add format flexibility (binary/MC/open) to direct generation | DG · MIMO | ΔS/C efficiency: 0.038 -> 0.197 | Improved |
| add retrieval augmentation to DG + Flexible | DG + Flexible · WebGen-V and MIMO | ΔS: WebGen-V 0.418 -> 0.383; MIMO 0.457 -> 0.292 | Mixed |
| split budget B=12 into intermediate n x m allocation (3x4) rather than extremes (ablation) | A-MPQC · budget allocation | ΔS | Improved |
| use fixed binary-only question format | DG · MIMO | ΔS: 0.019 vs 0.256 for DG | Hurt |
Experimental setup
Builds on: model predictive control
Closest papers by meaning
- Strategic Navigation or Stochastic Search? How Agents and Humans Reason Over Document Collections (ICML 2026 · similarity 0.85)
- OpenEQA: Embodied Question Answering in the Era of Foundation Models (CVPR 2024 · similarity 0.85)
- Mind2Web 2: Evaluating Agentic Search with Agent-as-a-Judge (NeurIPS 2025 · similarity 0.85)
- Diversity Over Frequency: Rethinking Tool Use in Visual Chain-of-Thought Agents (ICML 2026 · similarity 0.85)
- Visual Graph Arena: Evaluating Visual Conceptualization of Vision and Multimodal Large Language Models (ICML 2025 · similarity 0.85)
- Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
