GroundBench: leaderboard
GroundBench (Bian et al., 2026) re-targets 1,500 RefCOCO, RefCOCO+ and RefCOCOg referring expressions to polygon grounding: the model must trace the referred object with exactly N vertices, at five budgets from 8 to 64. Filled-region IoU with the canonical target contour on a 0-100 scale, counting unparseable answers as zero and averaged over the five budgets; higher is better.
Metric: Filled-region IoU (0-100), IoU x 100 (mean over the five vertex budgets N=8, 16, 24, 32 and 64; 1,500 RefCOCO, RefCOCO+ and RefCOCOg image, expression and referent triples (500 each) re-targeted to exact-N polygons; the model returns exactly N integer vertices and the filled polygon is scored by IoU against the canonical target; fixed denominator (an unparseable answer scores 0); first parseable output over one call plus up to two stateless retries). Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.5 Flash (High) | 57.7 | #98 (Gemini 3.5 Flash) |
| 2 | Seed 2.0 Pro | 42.7 | #209 |
| 3 | Kimi K2.7 Code | 33.9 | #180 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=groundbench · How It Works · Data refreshed daily, snapshot 2026-11-12.