Topic: model selection

The choice of large language model (LLM) backbone significantly impacts vision-language model (VLM) performance but lacks clear guidelines, as existing compute-based scaling laws do not reliably predict VLM success across different model families.

arxiv.org↗