Understanding How MLLMs Describe Artworks Using Token Activation Maps

2026-06-26

AI Summary

This study investigates the opacity of visual reasoning in multimodal large language models (MLLMs) when performing art description tasks, questioning whether their outputs genuinely reflect image content. To this end, it introduces the first application of Token Activation Maps (TAM) to the domain of art, integrating SAM³ open-vocabulary segmentation with tailored prompt engineering to generate heatmaps that quantify the degree of visual grounding for each generated token across a cross-period art dataset. The findings reveal that MLLMs’ reliance on visual input is strongly modulated by token semantic type: they demonstrate reliable grounding in artist identification but are prone to hallucination in title generation. Furthermore, the work systematically uncovers disparities in visual grounding across stylistic, iconographic, and affective dimensions. Code and results are publicly released to ensure reproducibility.

Abstract

Multimodal Large Language Models (MLLMs) describe artworks with remarkable fluency, yet the visual reasoning behind their outputs remains opaque. When an MLLM names a style, identifies a subject, or recognizes an iconographic symbol, does it ground each claim in the relevant region of the canvas, draw on an undifferentiated visual signal, or rely primarily on textual priors? We study this using the Token Activation Map (TAM), which produces, for each generated token, a heatmap isolating the visual evidence specific to that token from prior-context interference. Applying TAM to a curated set of paintings spanning multiple periods and genres, we analyze grounding patterns across five semantically distinct token categories: common visual objects, style descriptors, metadata, iconographic tokens, and affective expressions. We find that visual grounding varies substantially with token semantics. We further show that MLLMs attempt to identify artworks and artists, achieving higher accuracy in artist attribution than in title prediction, where hallucinations are more frequent. Finally, we compare TAM with SAM~3 open-vocabulary segmentation. To ensure reproducibility, we release our code, experimental configurations, prompts, and qualitative results on the project page at https://nicolafan.github.io/tamart/.

Problem

• Multimodal Large Language Models
• visual grounding
• artwork description
• token semantics
• visual reasoning

Innovation

• Token Activation Map
• Multimodal Large Language Models
• visual grounding
• artwork analysis
• interpretability