Saved in:
Bibliographic Details
Main Authors: Jain, Animesh, Stergiou, Alexandros
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2508.07833
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915918292975616
author Jain, Animesh
Stergiou, Alexandros
author_facet Jain, Animesh
Stergiou, Alexandros
contents Vision Language Models (VLMs) encode multimodal inputs over large, complex, and difficult-to-interpret architectures, which limit transparency and trust. We propose a Multimodal Inversion for Model Interpretation and Conceptualization (MIMIC) framework that inverts the internal encodings of VLMs. MIMIC uses a joint VLM-based inversion and a feature alignment objective to account for VLM's autoregressive processing. It additionally includes a triplet of regularizers for spatial alignment, natural image smoothness, and semantic realism. We evaluate MIMIC both quantitatively and qualitatively by inverting visual concepts across a range of free-form VLM outputs of varying length. Reported results include both standard visual quality metrics and semantic text-based metrics. To the best of our knowledge, this is the first model inversion approach addressing visual interpretations of VLM concepts.
format Preprint
id arxiv_https___arxiv_org_abs_2508_07833
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MIMIC: Multimodal Inversion for Model Interpretation and Conceptualization
Jain, Animesh
Stergiou, Alexandros
Computer Vision and Pattern Recognition
Vision Language Models (VLMs) encode multimodal inputs over large, complex, and difficult-to-interpret architectures, which limit transparency and trust. We propose a Multimodal Inversion for Model Interpretation and Conceptualization (MIMIC) framework that inverts the internal encodings of VLMs. MIMIC uses a joint VLM-based inversion and a feature alignment objective to account for VLM's autoregressive processing. It additionally includes a triplet of regularizers for spatial alignment, natural image smoothness, and semantic realism. We evaluate MIMIC both quantitatively and qualitatively by inverting visual concepts across a range of free-form VLM outputs of varying length. Reported results include both standard visual quality metrics and semantic text-based metrics. To the best of our knowledge, this is the first model inversion approach addressing visual interpretations of VLM concepts.
title MIMIC: Multimodal Inversion for Model Interpretation and Conceptualization
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2508.07833