ProactiveBench: leaderboard
Metric: Trajectory accuracy (%) averaged over the seven scenarios (ROD, VSOD, MVP-N, ImageNet-C, QuickDraw, ChangeIt, COCO): multi-turn multiple choice in which the model may name a category, abstain, or pick a proactive suggestion that yields a new frame, and a trajectory counts only when it ends in the correct category; samples a first-turn guess solves for at least 25% of the evaluated models are filtered out; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini | 34 | #172 |
| 2 | GPT-4.1 | 30.9 | #240 |
| 3 | GPT-5.2 | 24.8 | #105 |
| 4 | InternVL3-38B | 23 | #395 |
| 5 | Phi-4 Multimodal Instruct | 19.4 | #896 |
| 6 | InternVL3-78B | 15.6 | #345 |
| 7 | InternVL3-8B | 12.7 | #606 |
| 8 | Qwen 2.5 VL 32B Instruct | 10.8 | #443 |
| 9 | Qwen 2.5 VL 7B Instruct | 9.9 | #643 |
| 10 | Qwen 2.5 VL 72B Instruct | 7.5 | #364 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=proactivebench · How It Works · Data refreshed daily, snapshot 2026-10-11.