DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection

EMNLP 2026

Yuyang Hong^1,2,3,, Jinhui Guo^3,, Jiaqi Gu³, Lubin Fan^3,†,
Ruixiang Wang^1,2, Kun Ding^1,2, Yue Wu³, Shiming Xiang^1,2,†, Jieping Ye³

¹ University of Chinese Academy of Sciences
² MAIS, Institute of Automation, Chinese Academy of Sciences
³ Alibaba Token Hub, Alibaba Group

^* Equal contribution    ^† Corresponding authors

---

🔥 Highlights

• Sample-level quality: DIC measures the intrinsic consistency among the image, instruction, and response instead of relying only on global diversity or heuristic filtering.
• Two complementary signals: Visual Information Consistency (VIC) measures visual dependency, while Response Information Consistency (RIC) measures instruction-response coherence.
• Budget-aware selection: DICS adaptively balances intrinsic consistency and distributional diversity under different selection ratios.
• Less data, stronger performance: Using only 25% of LLaVA-1.5-665K, DICS reaches 101.40% relative performance, outperforming full-data fine-tuning.
• Scales to millions: On our DICS-6M corpus, a 25% subset outperforms full 6M-data training on InternVL3-8B.

🧩 Motivation

Visual instruction tuning is essential for vision-language alignment and instruction following. As multimodal datasets rapidly grow, selecting the most effective subset under a fixed training budget becomes increasingly important.

Existing methods mainly consider global properties such as diversity, sample relationships, importance, or heuristic quality rules. However, they often overlook a fundamental question:

Are the image, instruction, and response internally consistent within each individual sample?

DICS addresses this gap by explicitly evaluating sample-level intrinsic consistency before performing diversity-aware subset selection.

💡 Method

DICS first evaluates each visual instruction sample from visual and response perspectives, then performs adaptive diversity-aware selection according to the target data budget.

Data Intrinsic Consistency

DIC contains two complementary components:

• Visual Information Consistency (VIC): evaluates whether the image provides essential information for generating the response by comparing response losses with and without visual input.
• Response Information Consistency (RIC): evaluates whether the response is focused on and coherent with the instruction through a reverse instruction prediction task.

A high-quality sample should be reliable from both perspectives:

DIC = min(VIC, RIC)

Data Intrinsic Consistency Selection

DICS jointly considers DIC scores, representation diversity, and the target sampling budget:

• At low selection ratios, DICS prioritizes high-DIC samples and applies farthest-point sampling to preserve broad data coverage.
• At high selection ratios, DICS retains a high-quality core set and removes low-DIC samples with redundant representations.

This adaptive strategy provides robust subsets across different datasets, model architectures, and resource constraints.

📊 Main Results

LLaVA-1.5-665K

Using only 166K samples (25% of the full dataset), DICS achieves the strongest overall performance among all evaluated selection methods.

Method	Data Scale	MMMU	DocVQA	InfoVQA	POPE	Relative Performance
Full Dataset	665K	35.89	23.82	22.30	84.49	100.00%
Random Selection	166K	35.56	21.62	21.63	81.58	96.69%
PRISM	166K	36.11	23.87	21.87	82.18	97.66%
DICS (Ours)	166K	36.11	24.64	23.11	84.91	101.40%

Scaling to DICS-6M

We further curate DICS-6M and conduct large-scale visual instruction selection experiments with InternVL3-8B.

Selection Ratio	Data Scale	Relative Performance
15%	0.9M	99.26%
25%	1.5M	101.47%
50%	3.0M	100.32%
75%	4.5M	101.86%
85%	5.1M	103.10%

With 5.1M samples, DICS reaches 94.52% of the official InternVL3-8B-Instruct performance while using less than 25% of its reported 21.7M training samples.

✅ Contributions

1. We introduce Data Intrinsic Consistency, a unified self-scoring paradigm for measuring sample-level visual and response consistency.
2. We propose DICS, an adaptive selection strategy that integrates intrinsic consistency, diversity, and sampling efficiency.
3. We demonstrate strong robustness and scalability across datasets, model architectures, selection ratios, and evaluation benchmarks.
4. We curate DICS-6M for large-scale visual instruction selection research.

📢 News

• 2026: The DICS project page is now available.
• Paper, selected datasets, and trained models will be released progressively.

📝 Citation

If you find this work useful, please consider citing us:

@misc{hong2026dics,
      title={DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection}, 
      author={Yuyang Hong and Jinhui Guo and Jiaqi Gu and Lubin Fan and Ruixiang Wang and Kun Ding and Yue Wu and Shiming Xiang and Jieping Ye},
      year={2026},
      eprint={2608.30209},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.30209}, 
}

The citation entry will be updated after the paper is publicly released.

📬 Contact

For questions or discussions, please contact Yuyang Hong or open an issue in this repository.

---

⭐ If you find DICS useful, please consider starring this repository.