UAV-DualCog - Flight Behavior Recognition (Composite): leaderboard

Metric: Composite-level behavior accuracy (%; 1,024 first-person UAV flight videos whose executed composite flight behavior must be named with its temporal interval; structured JSON answers on renders of simulated AerialVLN scenes, the discrete answer scored whatever the predicted box or interval; instant mode with explicit thinking disabled where a model allows it; higher is better). Source: arxiv.org. Saturation forecast: Around 2039. 29 models tracked.

Top models

#ModelScore
1InternVL3.5-8B26.2
2GLM-4.6V (Non-reasoning)23.2
3Gemini 3.1 Flash Lite (Non-reasoning)22.8
4MiMo-V2-Omni22
5Qwen 3.5 122B A10B (Non-reasoning)21.9
6Qwen 3.5 35B A3B (Non-reasoning)18.4
7Kimi K2.5 (Non-reasoning)12.8
8Qwen 3.5 27B (Non-reasoning)12.2
9Kimi K2.6 (Non-reasoning)12.1
10Qwen 3.6 Plus (Non-reasoning)7.6
11Qwen 3.5 9B (Non-reasoning)6.6
12Qwen 3.7 Plus (Non-reasoning)6.6
13Qwen 3.5 397B A17B (Non-reasoning)5.7
14Qwen 3.5 4B (Non-reasoning)2.3
15Qwen 3.5 Plus (Non-reasoning)2.2

Interactive version: theaggregate.ai/benchmark?slug=uav-dualcog-flight-behavior-recognition-composite · How It Works · Data refreshed daily, snapshot 2026-09-29.