Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow

Chengxin Liu · Wonseok Choi · Chenshuang Zhang · Tae-Hyun Oh

CVPR 2026

Official implementation of the CVPR 2026 paper, "Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow".

Highlights

• Token Dynamics: LLM token dynamics reveal which visual tokens are relevant to the question, at least to some extent.
• Training-Free: Adaptive Information Flow is a training-free, test-time method that lets text tokens selectively attend to relevant visual tokens and improves VLM perception.
• Effective: The proposed method shows consistent improvements over LLaVA-1.5 and Qwen2.5-VL across diverse vision-centric tasks.

Code

ETA: June/July.

Citation

If you find our code or paper helpful for your research, please consider citing:

@inproceedings{liu2026aif,
    title = {Aligning What Vision-Language Models See and Perceive with Adaptive Information Flow},
    author = {Chengxin, Liu and Wonseok, Choi and Chenshuang, Zhang and Tae-Hyun, Oh},
    booktitle = {CVPR},
    year = {2026}
}

Contact

Chengxin Liu (cxliu@kaist.ac.kr)