SurgVLA-Bench - Level 2 Gauze Pick: leaderboard
Picking the instructed piece of gauze among visually similar distractors in the SurRoL laparoscopic simulator, a conditional task of SurgVLA-Bench (Sun et al., 2026) that tests target discrimination from the instruction. Models are LoRA fine-tuned on the Level 2 demonstrations; binary success rate in percent over 50 trials with a 100-step limit; higher is better.
Metric: Success rate (%; T2-1 gauze pick among visually similar distractors, SurgVLA-Bench Level 2, 50 trials in the SurRoL laparoscopic simulator with a 100-step limit; LoRA fine-tuned on the level's demonstrations). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | π0.5 | 36 | |
| 2 | OpenVLA | 8 | |
| 3 | π0 | 0 | |
| 4 | SmolVLA | 0 |
Interactive version: theaggregate.ai/benchmark?slug=surgvla-bench-level-2-gauze-pick · How It Works · Data refreshed daily, snapshot 2026-11-12.