AstroMind - Intent Inference: leaderboard
AstroMind (Liu et al., 2026) intent inference: from simulated, noisy orbital observations and intelligence reports of varying credibility, the model names the purpose of a spacecraft maneuver, such as station-keeping, phasing or rendezvous. Accuracy in percent over 66 questions (31 four-option, 35 open-ended mapped to labels by a DeepSeek-R1 judge) with a 4,096-token context; empty responses are excluded from scoring. Higher is better.
Metric: Accuracy (%; 66 maneuver-intent questions, 31 four-option and 35 open-ended answers mapped to the label set by a judge; empty responses excluded). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 32B | 67.74 | #960 |
| 2 | GPT-OSS-20B | 64.62 | #1150 |
| 3 | Gemma 3 27B | 56.92 | #1638 |
| 4 | QwQ-32B | 56.25 | #909 |
| 5 | Qwen 3 8B | 38.1 | #1838 |
| 6 | Gemma 3 4B | 25.76 | #5104 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=astromind-intent-inference · How It Works · Data refreshed daily, snapshot 2026-11-12.