TIC-Bench

Deeply Interleaved Text-Image Contexts for Multimodal LLMs Assessment

Zihao Wang, Xi Xiang, Yuwen Sun, Yingyu Li, Yabo Zhang, Yihan Zeng, Fan Li, and Wangmeng Zuo

TIC-Bench evaluates whether multimodal large language models can bind, integrate, and propagate evidence across long, deeply interleaved sequences of images and text. It contains 2,280 unique questions covering spatial, logical, and temporal association reasoning.

Benchmark composition

The benchmark contains 2,280 questions and 45,776 image instances. Each question contains 20.08 images on average, and each image is referenced 1.86 times on average.

Domain	Reasoning type	Questions
Logical Association	Linear	260
Logical Association	Cyclic	249
Logical Association	Convergent	275
Logical subtotal		784
Temporal Association	Sequential	270
Temporal Association	Retrospective	226
Temporal Association	Parallel	230
Temporal subtotal		726
Spatial Association	Map	385
Spatial Association	Photo	385
Spatial subtotal		770
Total		2,280

Tasks

• Spatial Association combines overlapping local crops and substituted text clues to infer relative directions. It contains aerial maps and natural photographs.
• Logical Association follows shared objects and textual relations through linear, cyclic, or convergent visual structures.
• Temporal Association follows identities and events through sequential, retrospective, or parallel story structures.

Download-oriented release format

The dataset is distributed as 15 uncompressed tar shards rather than more than 23,000 individually downloaded files. Every question_* folder remains intact inside one shard, so its QA file and referenced images are always downloaded together. The whole release is approximately 19 GiB, and each domain can be downloaded independently.

data/
├── Spatial Association/
│   └── ticbench-spatial-00000-of-00001.tar
├── Logical Association/
│   └── ticbench-logical-00000-of-00010.tar ... 00009-of-00010.tar
├── Temporal Association/
│   └── ticbench-temporal-00000-of-00004.tar ... 00003-of-00004.tar
├── manifest.jsonl
├── dataset_index.json
└── SHA256SUMS
assets/
ticbench_loader.py

The paths stored inside the shards preserve the original domain layout:

spatial_qa_shuffled/question_*/dataset_caption.jsonl
visual_construct_final/question_*/qa_pairs_*.json
time_qa/question_*/qa_pairs*.json

manifest.jsonl provides one row per QA record with its QID, domain, paper-facing task type, question folder, QA filename, record index, and shard. dataset_index.json summarizes record counts, duplicate-QID diagnostics, shard sizes, and SHA-256 hashes. SHA256SUMS supports independent integrity checks.

Quick start

Install the Hub client and download only the domain you need:

pip install -U huggingface_hub

from huggingface_hub import snapshot_download

local_repo = snapshot_download(
    repo_id="pino10010/TIC-Bench",
    repo_type="dataset",
    allow_patterns=[
        "data/manifest.jsonl",
        "data/dataset_index.json",
        "data/Spatial Association/ticbench-spatial-*.tar",
        "ticbench_loader.py",
    ],
)
print(local_repo)

Use the included helper to inspect the manifest and safely extract a domain:

import sys
from pathlib import Path

sys.path.insert(0, local_repo)
from ticbench_loader import iter_manifest, extract_domain

root = Path(local_repo)
spatial_rows = list(iter_manifest(root, domain="spatial"))
print(len(spatial_rows), spatial_rows[0])

extract_domain(root, domain="spatial", output_dir="./ticbench-spatial")

You can also locate one QID before downloading or extracting unrelated shards:

from ticbench_loader import find_qid

for row in find_qid(root, "qid_20260606_235445_0001"):
    print(row["shard"], row["question_folder"], row["qa_file"])

QA formats

• Spatial: dataset_caption.jsonl; each line is a QA record with an interleaved sequence, reference answer, choices, and metadata. Use this final caption format rather than the earlier dataset.jsonl intermediate format.
• Logical: qa_pairs_*.json; the qa_pairs array contains question, answer, reasoning depth, graph type, QID, and review/filtering fields.
• Temporal: qa_pairs*.json; the qa array contains question, reference answer, reasoning steps, task type, QID, and review fields, alongside the story description.

Stored logical labels linear/star, loop, and branch_merge/cross correspond to Linear, Cyclic, and Convergent. Stored temporal labels Linear, Loop, and Branch correspond to Sequential, Retrospective, and Parallel.

Data integrity note

The release contains 2,280 unique QIDs. Four QIDs in the current Temporal source are each attached to two distinct QA records, so the physical Temporal row count is 730 rather than 726. The manifest preserves all 730 records and marks their qid_occurrences; it does not silently discard either record. The affected QIDs are listed in dataset_index.json.

Evaluation

The benchmark evaluation covers ten multimodal LLMs across fifteen inference settings. Five open-source models are tested in both thinking and non-thinking modes, while five closed-source models use their default inference settings. Model answers are judged by three independent evaluators, with human resolution for inconsistent judgments. See the paper for the complete methodology and the official TIC-Bench repository for inference and three-judge evaluation code.

Source data and responsible use

The Spatial domain uses COCO photographs and CVOGL aerial imagery. The Temporal domain uses selected frames from publicly available movie, television, and animation clips. The Logical domain contains generated scene imagery.

No single permissive license is asserted for every underlying image. Users must comply with the terms and copyright restrictions of the respective source material, use the data for research and evaluation, avoid redistributing source media outside the permissions that apply to it, and perform their own legal review before commercial or public downstream use.

The benchmark may contain people, fictional characters, place imagery, and model-generated content. It should not be used for identity inference, surveillance, or decisions about individuals.

Citation

If you find TIC-Bench useful, please cite our paper:

@misc{wang2026deeplyinterleavedtextimagecontexts,
  title         = {Deeply Interleaved Text-Image Contexts for Multimodal LLMs Assessment},
  author        = {Zihao Wang and Xi Xiang and Yuwen Sun and Yingyu Li and Yabo Zhang and Yihan Zeng and Fan Li and Wangmeng Zuo},
  year          = {2026},
  eprint        = {2609.02573},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.02573}
}