DynVLA: Learning World Dynamics for Action Reasoning in Autonomous Driving

arXiv 26.3 2026 50.5 method

TLDR

DynVLA introduces Dynamics CoT, forecasting compact world dynamics via a tokenizer for informed action reasoning in autonomous driving.

Reasoning

The paper presents a novel CoT paradigm that compresses future world dynamics into tokens, decoupling ego and environment dynamics, and validates it on multiple benchmarks including a real-world dataset. Strengths include compact representation and empirical gains; weaknesses are domain specificity and lack of comparison to broader world model frameworks.

Read-first score

Read-first score 50.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 18.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset,github

Topical relevance 42%

25.7

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 582.

Keyword Scores

world dynamics prediction

world model

model-based reinforcement learning world model

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Introduces Dynamics CoT, a new Chain-of-Thought paradigm that forecasts compact world dynamics before action generation.
Proposes a Dynamics Tokenizer that compresses future evolution into a small set of dynamics tokens.
Decouples ego-centric and environment-centric dynamics for more accurate world dynamics modeling in interaction-intensive scenarios.
Trains the model with supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) to generate dynamics tokens before actions.

Methodology

DynVLA is a driving Vision-Language-Action (VLA) model that employs Dynamics CoT: it first uses a Dynamics Tokenizer to compress future world evolution into a compact set of dynamics tokens, then decouples these tokens into ego-centric and environment-centric components. The model is trained via SFT and RFT to generate dynamics tokens prior to action prediction, enabling physically grounded decision-making while maintaining latency-efficient inference. Evaluation is conducted on NAVSIM, Bench2Drive, and a large-scale in-house dataset, comparing against Textual CoT and Visual CoT baselines.

Key Results

DynVLA consistently outperforms both Textual CoT and Visual CoT methods across all three datasets (NAVSIM, Bench2Drive, in-house), demonstrating the effectiveness and practical value of Dynamics CoT for autonomous driving.

Resources

Paper