CARV: A Diagnostic Benchmark for Compositional Analogical Reasoning in Multimodal LLMs

Authors: Yongkang Du, Xiaohan Zou, Minhao Cheng, Lu Lin · Pennsylvania State University

Dataset Description

CARV evaluates whether multimodal LLMs can compose transformation rules from multiple image pairs via logical set operations. Given n context pairs each depicting an atomic visual change, the model must synthesize a new rule through Union (∪), Intersection (∩), or Difference (\) and apply it to a query image.

The benchmark spans 5,500 samples across three settings:

Task	#Input pairs	#Atomic transforms	Count
Single-step (baseline)	1	2	500
Compositional — Shared Source (SS)	2	2, 3, 4	3,500
Compositional — Different Source (DS)	2	2	1,500
Total			5,500

Shared Source (SS): The two context pairs share the same source image as the query (I_q = I_1 = I_2). Lower contextual complexity.

Different Source (DS): All three images are distinct (I_q ≠ I_1 ≠ I_2). Requires abstracting the rule away from the visual context — strictly harder.

Complexity Scaling: Extends Shared Source by increasing the number of atomic transformations per pair from 2 to 3 and 4.

Images are drawn from a controlled visual domain with five properties: subject, subject_number, object, object_color, spatial_relation.

Dataset Structure

carv-dataset/
├── data/
│   ├── mix/          # JPEG images — mix track (SS + DS, |T|=2)
│   └── large/        # JPEG images — complexity scaling track (SS, |T|=3,4)
├── tasks/
│   ├── mix/
│   │   ├── union/        {easy_500.json, hard_500.json}
│   │   ├── intersection/ {easy_500.json, hard_500.json}
│   │   └── difference/   {easy_500.json, hard_500.json}
│   ├── large/
│   │   ├── union/        {easy_500.json, hard_500.json}
│   │   └── intersection/ {easy_500.json, hard_500.json}
│   └── single/
│       └── union/        {easy_500.json}
└── labels/
    ├── mix/
    │   ├── union/        {easy_500.json, hard_500.json}
    │   ├── intersection/ {easy_500.json, hard_500.json}
    │   └── difference/   {easy_500.json, hard_500.json}
    └── large/
        ├── union/        {easy_500.json, hard_500.json}
        └── intersection/ {easy_500.json, hard_500.json}

Task JSON Fields

Each item in a tasks/ JSON file contains:

Field	Type	Description
context_image	list[str]	Context image filenames (relative to data/{track}/)
options	list[str]	Candidate answer image filenames
answer_index	int	Index of the correct option in options
wrong_position_index	int	Index of the spatially-wrong distractor
wrong_subject_index	int	Index of the subject-wrong distractor

Label JSON Fields

Each item in a labels/ JSON file contains:

Field	Type	Description
changed_properties1	list[str]	Properties that changed in the first context pair
changed_properties2	list[str]	Properties that changed in the second context pair

Usage

Download

pip install huggingface-hub
huggingface-cli download duyongka/CARV --repo-type dataset --local-dir ./carv-data

Run inference with the CARV codebase

git clone https://github.com/YongkDu/CARV
cd CARV
pip install -r requirements.txt

export CARV_CONFIG_DIR=/path/to/your/configs   # see README for config format

python src/inference/analogy_composition.py \
    --task mix \
    --prompt step_by_step \
    --model gemini2.5flash \
    --number 500

See the GitHub repository for full setup instructions, evaluation scripts, and the 4-stage failure diagnosis pipeline.

Results

Performance of frontier MLLMs on CARV compositional tasks (Accuracy %). SS = Shared Source, DS = Different Source.

Model	Single	Union SS	Union DS	Inter. SS	Inter. DS	Diff. SS	Diff. DS
GPT-5.1	81.2	64.6	51.0	75.0	73.0	59.0	62.4
Gemini-2.5 Pro	79.2	51.6	40.4	62.8	67.2	59.4	55.6
Gemini-2.5 Flash	75.0	49.4	37.4	60.2	64.2	41.4	33.2
GPT-4o	44.8	27.2	9.0	34.0	36.0	25.6	21.2
Qwen3VL-8B-Thinking	66.7	33.6	20.8	56.8	59.2	33.3	25.0
Qwen2.5VL-32B	39.0	18.6	6.2	30.4	32.8	25.4	15.0
Human	—	100.0	100.0	100.0	100.0	100.0	100.0

Citation

@article{du2026carv,
  title={CARV: A Diagnostic Benchmark for Compositional Analogical Reasoning in Multimodal LLMs},
  author={Du, Yongkang and Zou, Xiaohan and Cheng, Minhao and Lin, Lu},
  journal={arXiv preprint arXiv:2603.27958},
  year={2026}
}