ActiveFly-Bench - Observation Behavior Planning: leaderboard
ActiveFly-Bench (Tsinghua University and collaborators, 2026) studies aerial embodied perception for drones. In this task a vision-language model sees the first-person start view and an embodied question and picks, from four options, the flight and camera behavior that would reveal the answer; accuracy in percent over about 1,300 questions, higher is better.
Metric: Accuracy (%; 1.3k four-option questions: which flight and gimbal behavior from the first-person start view would reveal the answer to an aerial embodied question; the vision-language model answers alone). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 72.5 | #123 |
| 2 | Gemini 2.5 Pro | 70.4 | #266 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=activefly-bench-observation-behavior-planning · How It Works · Data refreshed daily, snapshot 2026-11-13.