[ACM MM 2026] Clarify Before Executing: A Self-Evolving Agent for Resolving Intent Asymmetry in 3D Tool Orchestration

Xiaoye Zhu, Weixin Li, Junan Huo, Bozhong Wang, Jia Zeng, Yi Yang, Cen Chen, Qi Liu^†

South China University of Technology
^†Corresponding author

Paper | Project Page | Code | Benchmark | Source Assets

CLARE places a clarification gate before 3D tool execution, resolving ambiguity, missing information, and mistaken constraints through targeted dialogue. This release centers on the 3D-Clarify benchmark and a reproducible code-level evaluation pipeline for the paper's single-step and multi-step experiments.

Highlights

• Clarify before executing: four specialized components gate execution, maintain clarified state, synthesize 3D tool code, and communicate targeted feedback.
• 3D-Clarify benchmark: 620 tasks across five 3D domains, each evaluated under ambiguity, missing-information, and mistaken-detail conditions.
• Self-evolving clarification: simulated multi-turn interaction and Multi-turn Reward optimization raise task success to 60.40% on Single and 43.34% on Multi.

3D-Clarify Benchmark

3D-Clarify evaluates whether an interactive agent resolves intent asymmetry before producing 3D tool calls. It contains 620 interaction scenarios across text-to-3D generation, single-view reconstruction, multi-view reconstruction, point-cloud editing, and post-processing.

Each task provides a fully specified hidden goal and three controlled initial-query variants: ambiguity, missing information, and mistaken details. The paper evaluates these three variants, giving 1,500 single-step and 360 multi-step task-condition instances.

Dataset Splits

from datasets import load_dataset

dataset = load_dataset("xyzhu1225/3D-Clarify")
single_tasks = dataset["single"]
multi_tasks = dataset["multi"]

Record Schema

The paper reports Goal Completion Rate (CR) and Task Success Rate (SR) over ambiguous, missing_info, and mistake. Multi-step records preserve tool-call order and intermediate dependencies.

Resources

• Paper: https://arxiv.org/abs/2607.16352
• Code and evaluation
• Source assets

The source media is not required for the paper's deterministic code-level evaluation. It is distributed separately for users who enable physical 3D tool execution. The few-shot prompting examples used by baselines are part of the code repository and are intentionally not included in this benchmark dataset.