A new paper on arXiv, titled "On the Limits of Scale: Inverse Scaling in Vision-Language Models," reports evidence that the usual scaling assumptions do not always hold. While scaling laws suggest that increasing model size, training data, and compute generally yields monotonic performance improvements, the authors say this often holds only at the aggregate level.
The study presents systematic evidence that vision-language models can exhibit inverse scaling — meaning larger models may perform worse on certain tasks. The abstract does not specify which tasks or benchmarks are affected, but the finding points to limits in current scaling approaches.
If confirmed, this complicates the common practice of simply scaling up models to gain capability. The paper is available on arXiv under identifier 2610.27360.