Latent Chain-of-Thought World Modeling for End-to-End Driving

arXiv 25.12 2025 43.5 method, application

TLDR

Latent-CoT-Drive uses latent chain-of-thought with world model tokens for end-to-end driving, achieving faster inference and better trajectory quality.

Reasoning

The paper introduces a novel latent reasoning space that unifies action proposals and world model tokens, supported by cold-start supervision and closed-loop RL post-training. Strengths include improved inference speed and trajectory quality over text-based reasoning baselines, but the approach is domain-specific and lacks discussion of generalization or safety limitations.

Read-first score

Read-first score 43.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 41.

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,baseline,benchmark,evaluation

Recency 6%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 29%

58.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 420.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

generative world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Introduces latent chain-of-thought (CoT) reasoning using action-proposal tokens and world model tokens instead of natural language
Unifies CoT reasoning and decision making in an action-aligned latent space
Cold starts latent CoT by supervising action proposals and world model tokens with ground-truth future rollouts
Post-trains with closed-loop reinforcement learning to strengthen reasoning capabilities

Methodology

LCDrive models reasoning as interleaved sequences of action-proposal tokens (from the same vocabulary as output actions) and world model tokens (grounded in a learned latent world model that predicts future outcomes). The model is first cold-started via supervised learning on ground-truth future rollouts, then fine-tuned with closed-loop reinforcement learning. Evaluation is performed on a large-scale end-to-end driving benchmark against non-reasoning and text-reasoning baselines.

Key Results

LCDrive achieves faster inference, better trajectory quality, and larger improvements from interactive reinforcement learning compared to both non-reasoning and text-reasoning baselines.

Limitations

Requires ground-truth future rollouts for cold-start supervision, which may be expensive or unavailable in real-world settings
Latent representations may be less interpretable than natural language, making debugging or safety analysis more challenging

Resources

Paper