MARVIS renders embeddings as visualizations and uses a 3B VLM to reason over them, beating Gemini 2.0 by 16% on average across four modalities while closing much of the gap to domain-specific models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
MARVIS: Modality Adaptive Reasoning over VISualizations
MARVIS renders embeddings as visualizations and uses a 3B VLM to reason over them, beating Gemini 2.0 by 16% on average across four modalities while closing much of the gap to domain-specific models.