Skip to content

RefCOCO average

A referring-expression grounding benchmark averaged across RefCOCO variants to test whether a model can localize described objects correctly.

8 models

Higher is better.

  1. 1Qwen3.6-27BAlibaba92.5
  2. 2Qwen3.5 397BAlibaba92.3
  3. 3Qwen3.6-35B-A3BAlibaba92.0
  4. 4Nemotron 3 Nano Omni 30B A3BNVIDIA90.5
  5. 5LFM2.5-VL-3BLiquid AI87.9
  6. 6Kimi K2.5Moonshot AI87.8
  7. 7Gemini 3 ProGoogle84.1
  8. 8Interfaze BetaInterfaze82.1
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Grounded visual localization
Tasks
Referring-expression grounding
Authors
RefCOCO dataset authors (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.