VisNec: Measuring and Leveraging Visual Necessity for Multimodal Instruction Tuning

ECCV 2026

Mingkang Dong* Hongyi Cai* Jie Li Sifan Zhou Bin Ren Kunyu Peng Yuqian Fu

*Equal contribution    Corresponding author

VisNec pipeline overview

VisNec computes the gap between multimodal loss and text-only loss, then performs semantic clustering to keep a visually necessary and task-diverse subset for VLM fine-tuning.

Abstract

The effectiveness of multimodal instruction tuning depends not only on dataset scale, but critically on whether training samples genuinely require visual reasoning. Existing instruction datasets often contain visually redundant samples that are solvable from text alone, as well as multimodally misaligned supervision that can degrade learning.

We propose VisNec, a principled data selection framework that measures the marginal contribution of visual input during instruction tuning. By comparing predictive loss with and without visual context, VisNec identifies whether a training instance is vision-critical, redundant, or misaligned. Combined with semantic clustering, VisNec selects compact, diverse, and visually necessary subsets for data-efficient multimodal instruction tuning.

Case Study

VisNec scores form an interpretable spectrum rather than a simple ranking. Low VisNec reveals mismatched supervision where the annotated answer conflicts with the image; medium VisNec often corresponds to redundant samples whose answers are visually independent; high VisNec captures visually dependent samples that require dense image evidence.

VisNec case study spectrum
Low VisNec Misalignment

The image contradicts the annotation, so the sample should be filtered rather than trusted.

Medium VisNec Redundancy

The answer can be inferred from general knowledge or language priors with limited visual need.

High VisNec Visually dependent

The instruction requires localized visual evidence, making the sample valuable for tuning.

Sensitivity Study

VisNec is robust to the semantic clustering granularity and remains consistently strong across different sampling ratios, with the top-15% subset providing the best relative performance.

Sensitivity study for number of clusters and sampling ratio

Baseline Comparisons

Under the same 15% data budget, VisNec consistently outperforms competitive data selection baselines on both LLaVA-665K and Vision-Flan-186K. Rel. reports average relative performance normalized by the corresponding full-data baseline.

LLaVA-v1.5-7B on LLaVA-665K

15% sampling ratio, 98K selected samples.

Method VQAv2 GQA LLaVA-Wild SQA-I TextVQA MME-P MMBench (en) MMBench (cn) POPE MM-Vet Rel.
Full-Data (665K)79.163.067.968.457.91476.964.358.386.430.0100.0
Random75.355.158.867.854.31397.561.053.584.930.294.2
Self-Filter74.056.360.662.351.41356.548.145.486.329.089.3
EL2N76.154.759.066.550.21405.258.248.583.330.091.9
TypiClust76.059.865.268.253.31396.264.357.185.629.796.9
IFD74.057.862.366.551.81307.257.250.686.628.192.1
PreSel76.557.965.670.155.21457.764.856.585.429.697.7
XMAS75.161.262.467.055.21485.764.354.185.831.097.3
COINCIDE76.160.264.967.754.81414.960.553.986.428.595.8
CLIP-Score71.756.961.364.553.41380.351.048.084.029.992.0
CoIDO75.861.066.768.256.01419.563.355.585.631.497.7
ICONS76.360.765.365.355.21435.663.155.885.730.497.1
VisNec (ours)78.060.869.867.956.21457.264.959.186.032.1100.2

LLaVA-v1.5-7B on Vision-Flan-186K

15% sampling ratio, 28K selected samples.

Method VQAv2 GQA LLaVA-Wild SQA-I TextVQA MME-P MMBench (en) MMBench (cn) POPE MM-Vet Rel.
Full-Data (186K)69.646.035.755.638.31238.153.448.285.727.7100.0
Random66.543.833.562.138.71238.643.643.183.028.196.7
Self-Filter64.942.532.159.342.61262.242.143.880.925.195.0
EL2N63.642.232.862.742.41253.844.737.779.527.195.2
TypiClust65.843.130.459.237.71194.132.445.181.628.292.0
IFD65.042.429.857.842.01210.930.440.882.626.991.6
PreSel64.141.939.466.239.71218.850.445.484.129.1100.6
XMAS65.549.434.258.142.41151.251.144.076.224.396.9
COINCIDE66.044.534.663.933.01184.449.648.284.326.197.1
CLIP-Score63.041.631.262.339.71058.037.544.282.028.092.2
CoIDO66.746.837.666.243.21298.851.447.385.628.3103.6
ICONS67.248.835.460.249.91252.551.345.483.028.6103.2
VisNec (ours)65.057.139.574.551.61505.562.653.182.032.1115.8

Method Overview

01

Estimate Visual Necessity

Compute the loss gap between multimodal and text-only inference for each instruction sample.

02

Preserve Task Diversity

Cluster semantic instructions and select high-necessity examples within each cluster.

03

Tune Efficiently

Fine-tune VLMs with a top-15% VisNec subset instead of the full instruction dataset.

Released Artifacts

Citation

If you find this work useful, please cite VisNec:

@misc{dong2026visnec,
      title={VisNec: Measuring and Leveraging Visual Necessity for Multimodal Instruction Tuning},
      author={Mingkang Dong and Hongyi Cai and Jie Li and Sifan Zhou and Bin Ren and Kunyu Peng and Yuqian Fu},
      year={2026},
      eprint={2603.01195},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.01195},
}