Autonomous Driving on realistic-highway n10 20-step protocol
0No Collision RateLlama-3.2-3B
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Llama-3.2-3BMode=Deterministic forecast replay, Model=Llama-3.2-3B2026.05 | 0 | 0 | 0 | 0 | 0 | -0.8 | 1.22 | 0.01 | — | |
| Qwen3-235B-A22BMode=Deterministic forecast replay, Model=Qwen3-235B-A22B2026.05 | 90 | 94.5 | 81.9 | 24 | 3.3 | 2.85 | 0.18 | 0.01 | 30.2 | |
| DeepSeek-v4-flashMode=Reactive, Model=DeepSeek-v4-flash2026.05 | 100 | 100 | 81 | 21 | 0.5 | 20.31 | 1.34 | 1.37 | — | |
| Claude Haiku 4.5Mode=Reactive, Model=Claude Haiku 4.52026.05 | 100 | 100 | 72 | 38 | 0 | 2.05 | 1.44 | 0.11 | — | |
| GPT-5.4 miniMode=Reactive, Model=GPT-5.4 mini2026.05 | 100 | 100 | 75.3 | 21.5 | 1 | 0.97 | 1.26 | 0.03 | — | |
| Qwen3-235B-A22BMode=Reactive, Model=Qwen3-235B-A22B2026.05 | 100 | 100 | 67.6 | 17 | 0 | 1.86 | 1.37 | 0.05 | — | |
| Llama-3.2-3BMode=Reactive, Model=Llama-3.2-3B2026.05 | 100 | 100 | 9.5 | 0 | 0 | 0.28 | 1.19 | 0.01 | — | |
| DeepSeek-v4-flashMode=Deterministic forecast replay, Model=DeepSeek-v4-flash2026.05 | 100 | 100 | 80.9 | 16.5 | 0 | 17.46 | 0.16 | 0.13 | 8.7 | |
| Claude Haiku 4.5Mode=Deterministic forecast replay, Model=Claude Haiku 4.52026.05 | 100 | 100 | 80 | 16 | 0 | 0.26 | 0.18 | 0.01 | 3.9 | |
| GPT-5.4 miniMode=Deterministic forecast replay, Model=GPT-5.4 mini2026.05 | 100 | 100 | 75.8 | 12 | 0 | -1.22 | 0.16 | 0 | 19.1 | |
| Steins;Gate DriveMode=Dual, Model=DeepSeek-v4-flash2026.05 | 100 | 100 | 80.2 | 19.5 | 0 | 51.16 | 1.37 | 1.18 | 18.7 | |
| Steins;Gate DriveMode=Dual, Model=GPT-5.4 mini2026.05 | 100 | 100 | 74.1 | 9 | 0.5 | 0.4 | 1.29 | 0.03 | 17 | |
| Steins;Gate DriveMode=Dual, Model=Qwen3-235B-A22B2026.05 | 100 | 100 | 72.8 | 39 | 0 | 5.64 | 1.42 | 0.04 | 37.8 | |
| Steins;Gate DriveMode=Dual, Model=Llama-3.2-3B2026.05 | 100 | 100 | 22.6 | 68 | 0 | -1.03 | 1.13 | 0.01 | 0 |