Vision & Language Benchmark

BrowseComp-VL

BrowseComp-VL - multimodal browsing-agent benchmark on hard visual web-search questions, proposed alongside WebWatcher (Alibaba-NLP): 'We propose a new benchmark, BrowseComp-VL, to evaluate the capabilities of multimodal agents' (https://huggingface.co/Alibaba-NLP/WebWatcher-32B); lineage of OpenAI BrowseComp (https://openai.com/index/browsecomp/). No standalone dataset/repo page surfaced; source documented as the introducing artifact.

Category

Vision & Language

Score type

percent

Scale

100%

higher is better

Models tested

3

What is BrowseComp-VL?

BrowseComp (Browsing Competence) is an OpenAI benchmark for deep web research. Each task is a question whose answer is deliberately hard to find — the model has to navigate many web pages, follow leads and cross-check sources. A correct final answer is the only thing that counts.

What does it test?

BrowseComp-VL measures understanding images combined with text. In detail, it checks:

  • searching the web step by step instead of guessing
  • reading and understanding many pages per task
  • combining information from several sources
  • persistence on questions that are nearly impossible without browsing

How to read the score

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Frequently asked questions

What is the BrowseComp-VL benchmark?

BrowseComp (Browsing Competence) is an OpenAI benchmark for deep web research. Each task is a question whose answer is deliberately hard to find — the model has to navigate many web pages, follow leads and cross-check sources. A correct final answer is the only thing that counts.

What does BrowseComp-VL test?

BrowseComp-VL mainly measures understanding images combined with text. Concretely, it checks: searching the web step by step instead of guessing; reading and understanding many pages per task; combining information from several sources.

How is the BrowseComp-VL score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get BrowseComp-VL data and scores?

The official source is https://huggingface.co/Alibaba-NLP/WebWatcher-32B. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/browsecomp-vl/scores) or pull it directly from the Open LLM Radar API with any HTTP client.

Model scores 3 models tested

Reset
# Model Score Date Source Verified
1 GLM-5.3-Flash
Zhipu AI
61%
— Report Unverified
2 DeepSeek-V4.1-Flash-1017
DeepSeek
60.6%
— Report Unverified
3 Sage-S2.1-Flash
Shanghai AI Laboratory
60.5%
— Report Verified

What now?

You've seen how models score on BrowseComp-VL — here's what to do next.