Semantic Recognition on AIDE 5->5 rollout (test)
73.35DBR (F1)Non-causal (bidir)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Non-causal (bidir)Rollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Non-causal2026.05 | 73.35 | 74.46 | 88.65 | 68.82 | |
| Driver-WM (full pretrained)Rollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal, Pre-training status=Full pretrained2026.05 | 73.22 | 73.97 | 90.54 | 62.09 | |
| Driver-WM (main)Rollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal2026.05 | 68.07 | 72.61 | 90.15 | 68.34 | |
| Cross-Attn onlyRollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal2026.05 | 62.52 | 68.75 | 87.94 | 69.09 | |
| MotionBERTRollout Protocol=5→5, VLM Interface=None, Causality=Causal2026.05 | 56.7 | 52.71 | — | — | |
| Static pooling (no rollout)Rollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal (no rollout)2026.05 | 54.04 | 60.07 | 26.39 | 14 | |
| Offline Enc-DecRollout Protocol=5→5, VLM Interface=None, Causality=Offline2026.05 | 53.82 | 50.68 | — | — | |
| Late fusionRollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal2026.05 | 45.87 | 49.1 | 87.85 | 65.97 | |
| SiMLPeRollout Protocol=5→5, VLM Interface=None, Causality=Causal2026.05 | 37.26 | 41.69 | — | — | |
| Single-streamRollout Protocol=5→5, VLM Interface=Qwen3-VL, Causality=Causal2026.05 | 24.56 | 24.35 | 26.39 | 14 | |
| ST-GCNRollout Protocol=5→5, VLM Interface=None, Causality=Causal2026.05 | 18.96 | 22.66 | — | — | |
| Zero-VelocityRollout Protocol=5→5, VLM Interface=None, Causality=Causal2026.05 | 8.27 | 14.94 | — | — |