VLM Agreement Doesn't Always Mean Correctness, Study Finds

The assumption that a model's agreement across varied inputs signifies correctness is widely used for reliability signals and self-training in AI. However, this coupling is rarely directly measured, especially in natural image contexts where semantic preservation is only assumed. This research addresses this gap by using scientific figures, which are programmatically generated from data, allowing for the creation of semantically equivalent images with an exact answer key. Researchers developed RENDEQ, a generator that produces "render-equivalence sets" for scientific figures. Using RENDEQ, they measured the agreement-accuracy coupling on three open-weight Vision-Language Models (VLMs). The study found that re-rendering perturbations outperformed simple resampling in both accuracy and reliability. While agreement generally beat a baseline of mean token log-probability, a critical finding emerged: fine-tuning models on their own cross-render consensus paradoxically led to a decrease in accuracy across all tested replication runs. This suggests that agreement only certifies correctness above a certain threshold, which is influenced by how diffuse a model's errors are. The objective of rewarding agreement can inadvertently destroy this necessary diffuseness, leading to worse performance. The dispersion in VLM errors was heavily concentrated in specific style factors, such as the plotting library used, highlighting the sensitivity of models to presentation variations.