Researchers have developed RENDEQ, a tool for generating semantically equivalent re-renderings of scientific figures to better measure the relationship between a model's agreement on perturbed inputs and its actual correctness. Testing on three open-weight vision-language models (VLMs), they found that re-rendering improved both accuracy and reliability compared to resampling. However, fine-tuning models on their own cross-render consensus led to a decrease in accuracy, contradicting previous findings on natural images and suggesting that objectives rewarding agreement can harm performance. AI
IMPACT Introduces a novel method for evaluating VLM reliability, potentially improving how model agreement is used to predict correctness.
RANK_REASON The cluster contains a research paper detailing a new method and experimental findings. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
