SIS-Bench - Action Prediction: leaderboard
SIS-Bench (2026) tests multimodal LLMs on UAV self-awareness and spatial cognition with four-option questions about drone videos. The action prediction task asks where the drone ends up after a given action sequence in a long flight; accuracy in percent, higher is better.
Metric: Accuracy (%; 263 action prediction questions; four-option multiple choice on UAV videos (at most 32 frames), zero-shot, open models through vLLM with at most 128 generated tokens and proprietary models through their APIs; higher is better). Source: arxiv.org. Saturation forecast: Around July 2028. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 Plus | 69.2 | #232 |
| 2 | GPT-5.4 | 65.8 | #123 |
| 3 | Kimi K2.5 | 65.4 | #256 |
| 4 | Qwen 3 VL 8B (Thinking) | 64.6 | #887 |
| 5 | Seed 1.8 | 63.1 | #242 |
| 6 | Gemini 3 Flash (Preview) | 61.6 | #144 |
| 7 | Qwen 2.5 VL 7B Instruct | 60.8 | #1833 |
| 8 | Qwen 3 VL 30B A3B Instruct | 59.7 | #871 |
| 9 | Qwen 2 VL 7B Instruct | 58.9 | #2553 |
| 10 | Qwen 3 VL 8B Instruct | 57 | #981 |
| 11 | Qwen 3 VL 4B Instruct | 57 | #1279 |
| 12 | InternVL3-14B | 55.5 | #1182 |
| 13 | GLM-4.1V-9B (Thinking) | 48.7 | #1060 (GLM-4.1V-9B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sis-bench-action-prediction · How It Works · Data refreshed daily, snapshot 2026-11-13.