IslamMesabah/myopia-benchmark

MYOPIA: Understanding the Limits of Tree-Based Reasoning in Vision-Language Models

MYOPIA is a benchmark for testing whether vision-language models (VLMs) classify images better when they follow a human-designed decision tree, one question at a time, than when they are asked for the label directly (zero-shot).

It provides:

• Four fixed evaluation subsets: GTSRB (901 images, 43 classes), MNIST (1,000, 10), CIFAR-10 (1,000, 10), and single-object MSCOCO (850, 90).

• One decision tree per dataset (128 internal nodes and 153 leaves in total). Each internal node asks a visual question with a closed answer set, and each leaf is a class.

• Every prompt used in the paper: tree-node and zero-shot templates, ten zero-shot paraphrases per dataset, and class descriptions.

• Reference predictions from GPT-4o, Qwen-VL-Max, and LLaMA-3.2-11B-Vision: 114,177 per-image records with full tree-traversal paths.

• An evaluation harness that runs any model behind an OpenAI-compatible API.

• Paper: arXiv:2610.21573

• Code: github.com/islammesabah/VLM_Classification

Headline result

Accuracy (%) at temperature 0.7 in the primary condition (no conversation history, no descriptions), computed from the predictions config:

Dataset	Model	Zero-shot	Tree
GTSRB	GPT-4o	71.14	57.71
	Qwen-VL-Max	47.06	24.64
	LLaMA-3.2-V	37.07	33.19
MNIST	GPT-4o	91.50	34.60
	Qwen-VL-Max	94.40	29.30
	LLaMA-3.2-V	59.30	7.70
CIFAR-10	GPT-4o	87.90	76.00
	Qwen-VL-Max	77.60	56.60
	LLaMA-3.2-V	36.70	16.40
MSCOCO	GPT-4o	72.82	46.71
	Qwen-VL-Max	74.82	39.65
	LLaMA-3.2-V	23.76	3.65

Tree-based inference is below zero-shot in every cell.

Quick start

from datasets import load_dataset

gtsrb = load_dataset("IslamMesabah/myopia-benchmark", "gtsrb", split="test")
print(gtsrb[0])
# {'image_id': '00004/00039_00029', 'image': <PIL.Image>, 'image_url': None,
#  'label': 4, 'label_name': '70 Kph Limit', 'dataset': 'GTSRB'}

preds = load_dataset("IslamMesabah/myopia-benchmark", "predictions", split="gtsrb")

Download the trees and prompts with the rest of the repository:

hf download IslamMesabah/myopia-benchmark --repo-type dataset --local-dir myopia-benchmark

Repository layout

data/<dataset>/test.parquet        evaluation subset
data/<dataset>/class_names.json    label id -> class name
trees/<dataset>.json               decision tree, machine-readable
trees/<dataset>.txt                decision tree, human-readable
prompts/templates.json             all prompt templates used by the harness
prompts/<DATASET>/zs_prompts.json  ten zero-shot paraphrases (index 0 is the paper's)
prompts/<DATASET>/classes_description.json   short and long class descriptions
predictions/<dataset>.parquet      reference model outputs
evaluation/evaluate.py             evaluation harness

<dataset> is one of gtsrb, mnist, cifar10, mscoco.

Data fields

gtsrb, mnist, cifar10, mscoco

Field	Type	Description
image_id	string	Identifier in the source dataset (GTSRB class/track_frame, MNIST training index, CIFAR-10 batch_<b>_image_<i>, COCO image id)
image	image	The image. null for MSCOCO (see Licensing)
image_url	string	Official COCO URL for MSCOCO, null otherwise
label	int	Ground-truth class id (COCO category ids are 1-based and non-contiguous)
label_name	string	Class name
dataset	string	Source dataset

Trees (trees/<dataset>.json)

{
  "dataset": "GTSRB", "num_internal_nodes": 22, "num_leaves": 43,
  "node_prompt": "{question} Choose one of these answers: {answers}.",
  "root": {
    "id": "n0", "question": "What's the sign's primary shape?",
    "answers": ["triangle", "circle", "diamond", "inverted-triangle", "octagon"],
    "children": {"triangle": {"id": "n1", "question": "...", "...": "..."}}
  }
}

Internal nodes have question, answers, and children (answer -> node). Leaves have label and label_name. To run tree inference, start at root, fill node_prompt with the node's question and answer list, match the model's reply against answers (case-insensitive, whole word, longest answer first), and move to that child. Stop at a leaf. If no answer matches, the prediction is -1.

predictions

One row per (image, model, condition). Each row holds a tree-based and a zero-shot output for the same image.

Field	Description
dataset, image_id, label	Join keys to the evaluation subsets
model	gpt-4o, qwen-vl-max, or llama3.2-vision:11b-instruct-fp16
temperature	Sampling temperature (0.0 or 0.7)
run_id	Zero-shot prompt paraphrase index (0-9); 0 is the primary prompt
include_memory	Tree questions asked in one conversation, with history (+H)
include_description	0: none; 1: image and class descriptions added to tree inference (Tree+D); 2: added to zero-shot inference (ZS+D)
include_zero_shot_label	Reserved; always false
tree_path	JSON list of {question, answer} steps the model took
tree_prediction	Leaf reached, or -1 if traversal failed
zero_shot_output	Raw zero-shot response; "0" where the zero-shot half was not run for that row
zero_shot_prediction	Parsed class id, or -1

Not every condition was run for every model and dataset. The paper reports a condition only when it covers at least 90% of a dataset, so check row counts before averaging.

Evaluate a model

pip install -r evaluation/requirements.txt
export OPENROUTER_API_KEY=...

# Tree-based inference
python evaluation/evaluate.py run --dataset gtsrb --method tree \
    --model openai/gpt-4o --base-url https://openrouter.ai/api/v1 \
    --api-key-env OPENROUTER_API_KEY --out gtsrb_tree.jsonl

# Zero-shot baseline
python evaluation/evaluate.py run --dataset gtsrb --method zero_shot \
    --model openai/gpt-4o --base-url https://openrouter.ai/api/v1 \
    --api-key-env OPENROUTER_API_KEY --out gtsrb_zs.jsonl

python evaluation/evaluate.py score gtsrb_tree.jsonl gtsrb_zs.jsonl

Any OpenAI-compatible endpoint works, including vLLM, Ollama (http://localhost:11434/v1), and OpenRouter. Runs can be resumed, and --limit N runs a quick subset. The harness reproduces the paper's primary condition: temperature 0.7, 100 max tokens, JPEG-encoded images, and no history or descriptions. The full harness with every ablation is in the GitHub repository.

How the benchmark was built

• Subsets. The images were sampled once with a fixed seed and kept fixed for every model and condition. MSCOCO is restricted to images annotated with exactly one object category, so each image has one correct label.
• Trees. The questions ask about visual attributes (shape, colour, symbol, stroke structure) that should be answerable from the image alone. Every tree was reviewed by a human. The paper (Sec. 3.1) describes the generator-evaluator pipeline used to draft trees.
• Labels are the source datasets' original annotations.

Limitations

• Tree quality reflects one human reviewer's judgment. A different tree for the same label set could give different results, and the paper's audit (Sec. 5.4) studies this.
• Each tree is specific to its dataset's label set and does not transfer to other datasets.
• The subsets are small (850-1,000 images), so differences of 1-2 points between models are within sampling noise.
• MSCOCO images must be downloaded from images.cocodataset.org. If that host changes, the URLs will break.
• The benchmark is for research on structured reasoning in VLMs. It is not a certification suite for deploying a model.

Licensing

Component	License
Decision trees, prompts, predictions, evaluation code	MIT (see LICENSE)
GTSRB images	Free for research use; cite Stallkamp et al. (2012)
MNIST images	CC BY-SA 3.0 (LeCun & Cortes)
CIFAR-10 images	No formal license; the authors ask that Krizhevsky (2009) be cited
MSCOCO	Annotations CC BY 4.0. Images are not redistributed: each stays under its Flickr uploader's license, so only the official URL is included

You are responsible for following the source datasets' terms when you use their images.

Citation

@misc{mesabah2026myopia,
  title         = {MYOPIA: Understanding the Limits of Tree-Based Reasoning in Vision-Language Models},
  author        = {Mesabah, Islam and Elmansoury, Sary and Gro{\ss}mann, Gerrit and Neigel, Peter and Vollmer, Sebastian J.},
  year          = {2026},
  eprint        = {2610.21573},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2610.21573},
  url           = {https://arxiv.org/abs/2610.21573}
}

If you use the images, please also cite the source datasets: GTSRB (Stallkamp et al., 2012), MNIST (LeCun et al., 1998), CIFAR-10 (Krizhevsky, 2009), and MSCOCO (Lin et al., 2014).