A new paper on arXiv investigates why test-time adaptation (TTA) may not carry over from contrastive vision-language models to generative multimodal large language models. The authors note that vision-language models suffer significant accuracy drops under common corruptions, a known weakness in real-world deployment.

The paper points out that TTA has been effective for CLIP-like contrastive models, helping them recover from distribution shifts at inference time. However, the abstract suggests that applying the same representation-based approach to generative multimodal LLMs is not straightforward, and the work sets out to explore the limits of that transfer.

Because the abstract is truncated, the full findings are not yet visible from the announcement. Still, the framing highlights an important gap: methods that work for one family of models may not automatically benefit another, especially when the architectures and objectives differ.