{"ID":2830062,"CreatedAt":"2026-06-01T04:54:23.091178241Z","UpdatedAt":"2026-06-01T04:54:23.091178241Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2512.10226","arxiv_id":"2512.10226","title":"Latent Chain-of-Thought World Modeling for End-to-End Driving","abstract":"Recent Vision-Language-Action (VLA) models for autonomous driving explore inference-time reasoning as a way to improve driving performance and safety in challenging scenarios. Most prior work uses natural language to express chain-of-thought (CoT) reasoning before producing driving actions. However, text may not be the most efficient representation for reasoning. In this work, we present Latent-CoT-Drive (LCDrive): a model that expresses CoT in a latent language that captures possible outcomes of the driving actions being considered. Our approach unifies CoT reasoning and decision making by representing both in an action-aligned latent space. Instead of natural language, the model reasons by interleaving (1) action-proposal tokens, which use the same vocabulary as the model's output actions; and (2) world model tokens, which are grounded in a learned latent world model and express future outcomes of these actions. We cold start latent CoT by supervising the model's action proposals and world model tokens based on ground-truth future rollouts of the scene. We then post-train with closed-loop reinforcement learning to strengthen reasoning capabilities. On a large-scale end-to-end driving benchmark, LCDrive achieves faster inference, better trajectory quality, and larger improvements from interactive reinforcement learning compared to both non-reasoning and text-reasoning baselines.","short_abstract":"Recent Vision-Language-Action (VLA) models for autonomous driving explore inference-time reasoning as a way to improve driving performance and safety in challenging scenarios. Most prior work uses natural language to express chain-of-thought (CoT) reasoning before producing driving actions. However, text may not be the...","url_abs":"https://arxiv.org/abs/2512.10226","url_pdf":"https://arxiv.org/pdf/2512.10226v2","authors":"[\"Shuhan Tan\",\"Kashyap Chitta\",\"Yuxiao Chen\",\"Ran Tian\",\"Yurong You\",\"Yan Wang\",\"Wenjie Luo\",\"Yulong Cao\",\"Philipp Krahenbuhl\",\"Marco Pavone\",\"Boris Ivanovic\"]","published":"2025-12-11T02:22:07Z","proceeding":"cs.CV","tasks":"[\"cs.CV\",\"cs.RO\"]","methods":"[\"Reinforcement Learning\"]","has_code":false}