Vision language models can nail the same answer twice while their internal reasoning quietly falls apart.
A paper posted to arXiv (arXiv:2602.06652) tested modern VLMs on the SEEDBench, MMMU, and POPE benchmarks, tracking not just final answers but the internal embeddings behind them. The paper's authors found that models often kept their output the same even as internal representations drifted substantially. For perturbations like text overlays, that drift approached the same magnitude you'd see between two completely different images. Bigger models did not fix this: they scored higher on accuracy but showed equal or greater sensitivity to perturbations, meaning sharper decision boundaries that are also more fragile. The effect cut both ways on downstream tasks. Perturbations hurt reasoning when they scrambled how a model combines coarse and fine visual detail, but on hallucination tests they sometimes made models answer more conservatively and cut false positives.
This matters because benchmark leaderboards mostly grade VLMs on whether the final answer matches a label, not on whether the model reached it the same way twice. A model that quietly reroutes its internal processing while giving consistent outputs is a model whose behavior could shift unpredictably the moment real world inputs drift outside a benchmark's test distribution.
Scaling has fixed plenty of VLM problems. Internal stability, according to this paper, apparently isn't one of them.