FLY-EVAL++ - Single-Step Prediction - Field Validity: leaderboard
FLY-EVAL++ (arXiv 2609.04021v1, Table 1) single-step flight-state prediction (S1), D2 field score: numeric validity, certified range limits and per-second jump limits. Deterministic evidence-based verification of 45-field JSON predictions; percent, higher is better. Only the 21 of 66 screened models above 90% protocol compliance are listed.
Metric: S1 D2 field score (%). Source: arxiv.org. 21 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Grok 4 | 68 | #189 |
| 2 | Qwen 3 235B A22B | 63.8 | #323 |
| 3 | Qwen 3 32B | 63.7 | #449 |
| 4 | Qwen 2.5 32B | 63.6 | #544 |
| 5 | Qwen3-Next-80B | 63.4 | #321 |
| 6 | Claude Sonnet 4.5 | 61.9 | #148 |
| 7 | Llama 3.1 405B | 61.9 | #476 |
| 8 | Kimi K2 0905 | 61.8 | #292 |
| 9 | Gemini 2.5 Pro | 61.7 | #159 |
| 10 | DeepSeek V3 | 61.7 | #329 |
| 11 | DeepSeek R1 0528 | 61.7 | #232 |
| 12 | GPT-4o | 61.6 | #357 |
| 13 | GPT-5 | 61.6 | #105 |
| 14 | DeepSeek V3.1 | 61.5 | #275 |
| 15 | Kimi K2 (Thinking) | 61.5 | #252 (Kimi K2) |
Interactive version: theaggregate.ai/benchmark?slug=fly-eval-plus-plus-single-step-prediction-field-validity · How It Works · Data refreshed daily, snapshot 2026-11-05.