DailyClue Dataset

Seek-and-Solve: Benchmarking MLLMs for Visual Clue-Driven Reasoning in Daily Scenarios

Dataset Summary

DailyClue is a benchmark for evaluating visual clue-driven reasoning in Multimodal Large Language Models (MLLMs). Unlike benchmarks that test pre-existing knowledge, DailyClue requires models to actively identify decisive visual clues from images before producing answers.

The dataset spans 4 major domains and 16 distinct subtasks, with 666 total questions.

Dataset Structure

DailyClue/
├── daily_life/          # Images for Daily Commonsense Reasoning
├── location/            # Images for Location Identification
├── science/             # Images for Scientific Commonsense
├── spatial/             # Images for Spatial Reasoning
├── daily_life.json
├── location.json
├── science.json
└── spatial.json

Statistics

Category	# Questions	Formats
Daily Commonsense Reasoning	180	Multiple Choice, Yes/No, Open-ended
Location Identification	200	Open-ended
Spatial Reasoning	163	Multiple Choice, Yes/No
Scientific Commonsense	123	Multiple Choice, Yes/No, Open-ended
Total	666

Data Fields

Each JSON entry contains:

Field	Type	Description
image	list[str]	Image filename(s) within the category subfolder
question	str	The question posed to the model
clues	str	Human-annotated ground-truth visual clues (see note below)
ground_truth	str	The correct answer
format	str	"Multiple choice", "Yes or no", or "Open-ended"
category_1	str	Primary domain (one of the four above)
category_2	str	Subtask within the primary domain
language	str	"English"

Note on clues: This field contains human-annotated ground-truth visual clues. It is used in ablation experiments (injecting GT clues to probe the impact on model accuracy) and in the Rigorous Evaluation Protocol (checking whether model-predicted clues semantically align with GT clues). It is not fed to the model during standard inference.

Usage

Download

# via git
git clone https://huggingface.co/datasets/Crysun/DailyClue

# via huggingface_hub
from huggingface_hub import snapshot_download
snapshot_download(repo_id="Crysun/DailyClue", repo_type="dataset", local_dir="./dataset")

Run Inference

After downloading, point the inference script to the local directory:

python infer/inference.py \
  --dataset ./dataset \
  --model_names "gpt-4o" \
  --prompt_mode "b"

See the GitHub repository for the full evaluation pipeline.

Citation

@article{dailyclue2026,
  title={Seek-and-Solve: Benchmarking MLLMs for Visual Clue-Driven Reasoning in Daily Scenarios},
  author={Li, Xiaomin and Wang, Tala and Zhong, Zichen and Zhang, Ying and Zheng, Zirui and Isobe, Takashi and Li, Dezhuang and Lu, Huchuan and He, You and Jia, Xu},
  journal={arXiv preprint arXiv:2604.14041},
  year={2026}
}