🚗 Autonomous Driving

🧪 ICML2026 · 1 paper notes

📌 Same area in other venues: 📷 CVPR2026 (88) · 🔬 ICLR2026 (18) · 🤖 AAAI2026 (57) · 🧠 NeurIPS2025 (49) · 📹 ICCV2025 (93)

DeepSight: Long-Horizon World Modeling via Latent States Prediction for End-to-End Autonomous Driving

DeepSight shifts "future world prediction" from explicit pixel reconstruction (codebook single-frame) to multi-frame parallel implicit prediction of DINOv3 semantic features in BEV space, with an additional on-demand Adaptive Chain-of-Thought. This enables Qwen2.5-VL-3B to achieve a Driving Score of 86.23 (+7.39) and Success Rate of 71.36% (+13.63) on Bench2Drive closed-loop, with only ~4% extra inference latency.