Visual Instruction Data Selection

DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection

Select less. Learn more.

Yuyang Hong1,2,3,* Jinhui Guo3,* Jiaqi Gu3 Lubin Fan3,† Ruixiang Wang1,2 Kun Ding1,2 Yue Wu3 Shiming Xiang1,2,† Jieping Ye3
1 University of Chinese Academy of Sciences 2 Institute of Automation, Chinese Academy of Sciences 3 Alibaba Token Hub, Alibaba Group

* Equal contribution   † Corresponding authors

▤ Paper ⌘ Code ◫ Dataset · Coming Soon
DICS relative performance across selected data ratios compared with full-data training and other methods

DICS consistently leads competing selection methods and surpasses full-data fine-tuning from only 25% of LLaVA-1.5-665K.

25%of LLaVA databeats full-data fine-tuning
101.40%relative performanceon 12 diverse benchmarks
6Minstruction sampleslarge-scale selection study
94.52%of official performancewith <25% reported data
01 / Motivation

Quality begins inside each sample.

Abstract

Visual instruction tuning is crucial for advancing the vision-language alignment and instruction-following capabilities of Vision-Language Models. Yet, identifying optimal subsets under a fixed data budget remains a significant bottleneck.

We introduce Data Intrinsic Consistency (DIC), a self-scoring metric that measures sample-level consistency among the image, instruction, and response. Building upon DIC, Data Intrinsic Consistency Selection (DICS) adaptively balances high intra-sample consistency with global distributional diversity across data budgets.

Across datasets and model architectures, DICS consistently outperforms state-of-the-art selection methods—surpassing full-dataset fine-tuning with only 25% of LLaVA-1.5-665K.

02 / Method

Two views of intrinsic consistency.

DIC uses the model itself to quantify how much each component contributes—without relying on an external evaluator.

01
V

Visual Information Consistency

Does the image provide essential information for producing the response? VIC contrasts response losses with and without visual context.

VIC = ℒ(response | instruction) − ℒ(response | image, instruction)
02
R

Response Information Consistency

Is the response focused, accurate, and coherent with the instruction? RIC measures how the response helps recover the original instruction.

RIC = ℒ(instruction | image) − ℒ(instruction | image, response)
03
D

Data Intrinsic Consistency

A high-quality sample must be strong from both perspectives. The stricter dimension determines its final intrinsic consistency.

DIC = min(VIC, RIC)
Adaptive selection

Consistency meets diversity.

At low selection ratios, DICS prioritizes high-DIC samples while preserving broad coverage through farthest-point sampling. At high ratios, it removes low-consistency samples with redundant features.

Visual instruction
dataset
→
VICRICDIC
→
Diversity-aware
sampling
→
Optimal
subset
03 / Results

More performance from less data.

Comparison of DICS performance and data efficiency against baseline selection methods
LLaVA-1.5-665K · 25%

DICS is the only 25% selection method to beat full-data training.

Using 166K samples, DICS reaches 101.40% relative performance and leads on MMBench-CN, DocVQA, InfoVQA, and POPE.

  • DocVQA24.64
  • InfoVQA23.11
  • POPE84.91
DICS-6M · InternVL3-8B

Consistency scales to millions.

At 1.5M samples, DICS already outperforms full 6M training. At 5.1M, it reaches 103.10% relative performance.

15% · 0.9M99.26
25% · 1.5M101.47
50% · 3.0M100.32
75% · 4.5M101.86
85% · 5.1M103.10
04 / Takeaways

Simple principle. Broad impact.

01

Sample-level metric

DIC jointly measures visual and response consistency with one self-scoring paradigm.

02

Budget-aware selection

DICS adapts the quality–diversity trade-off to the requested subset size.

03

Robust at scale

Consistent gains across datasets, architectures, data scales, and evaluation capabilities.

05 / Citation

Cite DICS.

@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}, 
}

Preliminary citation; venue and publication fields will be updated after the paper is released.