A new research paper explores the capabilities of multimodal large language models (MLLMs) in generating aesthetic critiques, comparing their output to human evaluations. The study utilized the Reddit Photo Critique Dataset to assess five open-weight MLLMs across various prompt conditions. Findings indicate that while MLLMs can produce comprehensive critiques, their style differs significantly from human evaluations, often being more verbose and less selective, which challenges current evaluation metrics. AI
IMPACT Highlights limitations in current MLLM evaluation methods for open-ended generation, suggesting a need for new training and assessment strategies.
RANK_REASON Research paper published on arXiv evaluating MLLM capabilities.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
