Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

[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. The release centers on the 3D-Clarify benchmark and a reproducible code-level evaluation pipeline for the paper's single-step and multi-step experiments.

CLARE framework and self-evolution pipeline

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.

CLARE Assets v1

This repository hosts the source media required for physical 3D tool execution on 3D-Clarify. The benchmark JSON and deterministic CR/SR evaluation code are distributed separately in xyzhu1225/3D-Clarify and the CLARE code repository.

CLARE-assets-v1.zip contains 427 image, mesh, and point-cloud files and occupies approximately 443 MiB compressed. The assets support five operation families:

  1. Text-to-3D generation
  2. Single-view reconstruction
  3. Multi-view reconstruction
  4. Point-cloud editing
  5. Post-processing

They correspond to the benchmark's 500 single-step tasks and 120 multi-step workflows. Each task is evaluated under ambiguity, missing-information, and mistaken-detail conditions.

Source media is not required for the paper's code-level evaluation. Download it only when enabling the underlying 3D tools.

Download And Setup

Download CLARE-assets-v1.zip, then extract it at the root of a CLARE checkout:

unzip CLARE-assets-v1.zip -d /path/to/CLARE/

This creates data/source/, matching the input paths referenced by the released benchmark JSON.

Verify the archive before extraction:

echo "77b5b068f5105e0c95f14d94cd458da52e2187042c8c301e6e7f9076a147fcf9  CLARE-assets-v1.zip" | sha256sum -c -

For benchmark structure, task counts, ground-truth fields, and evaluation commands, see the 3D-Clarify dataset and the dataset guide.

Downloads last month
38

Paper for xyzhu1225/CLARE-assets-v1