ARMBench-VL

ARMBench-VL is the evaluation benchmark introduced in ARM-Thinker. It evaluates multimodal reward models on evidence-grounded judgment tasks that benefit from image manipulation, document-page retrieval, or instruction-verification tools.

This release contains all 1,499 benchmark annotations and the visual resources used as model inputs and tool-retrieval resources.

Benchmark composition

Config	Examples	Evaluation form	Tool family
fine_grained_perception	550	295 pair-RM, 163 two-way pair-RM, 92 four-way pair-RM	image crop and zoom-in
multimodal_long_document_qa	460	173 single-RM, 287 pair-RM	document-page retrieval
multimodal_instruction_following	489	489 single-RM	instruction validators

The counts match Table 6 of the paper.

Repository layout

annotations/
  fine_grained_perception.jsonl
  multimodal_long_document_qa.jsonl
  multimodal_instruction_following.jsonl
resources/
  fine_grained_perception/vstar_bench/   # images referenced by the FG annotations
  multimodal_long_document_qa/           # initial composite images and document pages
  multimodal_long_document_qa_index/     # Chroma/CLIP retrieval index for the 119 documents
  multimodal_instruction_following/      # images referenced by the IF annotations
scripts/
  load_armbench_vl.py
  validate_release.py
dataset_config.json

The images field in each annotation is relative to that task's resource root:

Config	Resource root
fine_grained_perception	resources/fine_grained_perception/
multimodal_long_document_qa	resources/multimodal_long_document_qa/
multimodal_instruction_following	resources/multimodal_instruction_following/

For long-document QA, files such as <doc_id>_concat1_0.jpg are the initial composite inputs. Files named <doc_id>_<page>.jpg are individual pages used by the document-retrieval tools. The doc_id field identifies the corresponding document.

The release also includes a Chroma collection named multi_pdf_clip_collection, filtered to the 119 documents used by ARMBench-VL. To use the repository's document tools with a local download, set:

export RAG_IMAGE_ROOT="$PWD/resources/multimodal_long_document_qa"
export DB_PATH="$PWD/resources/multimodal_long_document_qa_index"
export COLLECTION_NAME="multi_pdf_clip_collection"

Download and load

Download the complete release:

hf download ChrisDing1105/ARMBench-VL \
  --repo-type dataset \
  --local-dir ARMBench-VL

Load annotations with Hugging Face Datasets:

from datasets import load_dataset

fg = load_dataset(
    "ChrisDing1105/ARMBench-VL",
    "fine_grained_perception",
    split="test",
)
doc = load_dataset(
    "ChrisDing1105/ARMBench-VL",
    "multimodal_long_document_qa",
    split="test",
)
instruction = load_dataset(
    "ChrisDing1105/ARMBench-VL",
    "multimodal_instruction_following",
    split="test",
)

For local evaluation, scripts/load_armbench_vl.py resolves image paths and document pages from a downloaded copy. The evaluation implementation and tool configurations are maintained in the ARM-Thinker repository.

Validate a complete local download with:

python scripts/validate_release.py .

Main fields

• idx: task-local example index.
• rm_tag: single_rm, pair_rm, 2-way-pair_rm, or 4-way-pair_rm.
• question: the input question or instruction.
• preds: candidate response or responses to judge.
• gt: reference judgment.
• images: task-relative image paths.
• constraints: instruction-following constraints, when applicable.
• doc_id, evidence_pages, evidence_sources: long-document metadata, when applicable.

Source data and terms

ARMBench-VL annotations are derived from tasks based on V* Bench, MMLongBench-Doc, and MM-IFEval/MM-IFEngine. The annotations created for ARM-Thinker are released for non-commercial research under CC BY-NC 4.0. Source images and documents remain subject to their original datasets' terms; this repository does not relicense those source assets. Please cite the corresponding upstream work when using a task subset.

Citation

@misc{ding2025armthinkerreinforcingmultimodalgenerative,
  title={ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning},
  author={Shengyuan Ding and Xinyu Fang and Ziyu Liu and Yuhang Zang and Yuhang Cao and Xiangyu Zhao and Haodong Duan and Xiaoyi Dong and Jianze Liang and Bin Wang and Conghui He and Dahua Lin and Jiaqi Wang},
  year={2025},
  eprint={2512.05111},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.05111}
}