Skip to content

Cola

Aggregate vision-language outputs with a language model

Visit official site

What Cola is

The method is described as a way to aggregate the outputs of two or more vision-language models through a coordinating language model.

What you can do with it

  • Combine several VLM outputs with an LM
  • Use the fine-tuned Cola-FT variant
  • Use the zero-shot Cola-Zero variant

Model notes

  • InstructBLIP is named among applicable VLMs
  • Seven datasets are listed, including VQA v2 and GQA
  • Gains and robustness to VLM errors are reported

Pros & cons

✓ What we like

  • Works with models such as InstructBLIP
  • Seven datasets are named in the material

! What to watch out for

  • The method needs a coordinating language model
  • Results depend on the assembled VLM outputs

FAQ

What does the name expand to?

The material expands Cola as COordinative LAnguage model or visual reasoning.

Which datasets are named?

The material lists VQA v2, OK-VQA, A-OKVQA, e-SNLI-VE, VSR, CLEVR and GQA.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review