Research map / Architectures and efficiency / RNNs and normalization
ETA: Efficiency through Thinking Ahead, A Dual Approach to Self-Driving with Large Models
Shadi Hamdan, Chonghao Sima, Zetong Yang, Hongyang Li, Fatma Guney
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add forecasting supervision from large model features (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 54.92 -> 69.53 | Improved |
| add the small model to the forecasting pipeline (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 42.49 -> 69.53 | Improved |
| add action mask supervision (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 42.67 -> 69.53 | Improved |
| add action mask loss (ablation) | Base Model · Bench2Drive (Leaderboard-v2) | Driving Score: 70.02 -> 74.33 | Improved |
| add large-model forecasting to the small model (ablation) | Small Model Only · Bench2Drive (Leaderboard-v2) | Driving Score: 61.30 -> 69.53 | Improved |
| add small model to forecasting for emergency braking (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | E. Brake ability: 16.36 -> 60.13 | Improved |
| add small model for traffic signs (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | T. Sign ability: 30.56 -> 43.64 | Improved |
| replace forecast features by ground truth in train and test (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 69.53 -> 74.12 | Improved |
| provide ground truth features only at test time (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 69.53 -> 60.72 | Hurt |
| use ViT-Base instead of first 8 CLIP-ViT-L layers as small model (ablation) | Async Model · Bench2Drive (Leaderboard-v2) | Driving Score: 69.53 -> 57.07 | Hurt |
| replace CLIP-ViT-L large encoder with CLIP-ViT-g (ablation) | Base Model · Bench2Drive (Leaderboard-v2) | Driving Score: 74.33 -> 73.19 | No clear effect |
| replace CLIP-ViT-L large encoder with InternViT (ablation) | Base Model · Bench2Drive (Leaderboard-v2) | Driving Score: 74.33 -> 71.53 | Hurt |
Closest papers by meaning
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.86)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.86)
- OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation (NeurIPS 2026 · similarity 0.85)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.85)
- LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.85)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
