While reasoning technology like Chain of Thought (CoT) has been widely adopted in Vision Language Action (VLA) models, it demonstrates promising capabilities in end to end autonomous driving. However, recent efforts to integrate CoT reasoning often fall short in simple scenarios, introducing unnecessary computational overhead without improving decision quality. To address this, we propose AdaThinkDrive, a novel VLA framework with a dual mode reasoning mechanism inspired by fast and slow thinking. First, our framework is pretrained on large scale autonomous driving (AD) scenarios using both question answering (QA) and trajectory datasets to acquire world knowledge and driving commonsense. During supervised fine tuning (SFT), we introduce a two mode dataset, fast answering (w/o CoT) and slow thinking (with CoT), enabling the model to distinguish between scenarios that require reasoning. Furthermore, an Adaptive Think Reward strategy is proposed in conjunction with the Group Relative Policy Optimization (GRPO), which rewards the model for selectively applying CoT by comparing trajectory quality across different reasoning modes. Extensive experiments on the Navsim benchmark show that AdaThinkDrive achieves a PDMS of 90.3, surpassing the best vision only baseline by 1.7 points. Moreover, ablations show that AdaThinkDrive surpasses both the never Think and always Think baselines, improving PDMS by 2.0 and 1.4, respectively. It also reduces inference time by 14% compared to the always Think baseline, demonstrating its ability to balance accuracy and efficiency through adaptive reasoning.
Figures & tables
Fig. 1 : Impact and Design of Adaptive Reasoning in Trajectory Prediction.
Fig. 2 : We present AdaThinkDrive , an end-to-end autonomous driving framework that adaptively selects between "Thinking" and "Non-Thinking" modes depending on scene complexity. Given vision and text inputs, the VLM dynamically determines its output mode through an adaptive reasoning mechanism. During the reinforcement learning of the three-stage training process, multiple reward including PDMS, format, and endpoint are combined with the proposed Adaptive Think Reward.
Fig. 3 : Visualization of dynamic agents for Think-style CoT supervision. (a)-(b) depict CIPO-1 agents (occupying the ego lane) and CIPO-2 agents (likely to merge), while (c)-(d) show Motion Interaction cases where agents’ future trajectories intersect with the ego vehicle’s trajectory.
Fig. 4 : Adaptive Think Reward: A Dynamic Reasoning Control Strategy. This reward adjusts the model’s reasoning behavior by identifying misclassified scenes. When scene-specific conditions are satisfied, it assigns rewards to either Thinking or Non-thinking responses accordingly.
Method
Image
Lidar
NC ↑
DAC ↑
TTC ↑
CF ↑
EP ↑
PDMS ↑
Constant Velocity
68.0
57.8
50.0
100
19.4
20.6
Ego Status MLP
93.0
77.3
83.6
100
62.8
65.6
UniAD [ 3 ]
✓
97.8
91.9
92.9
100
78.8
83.4
LFT [ 37 ]
✓
97.4
92.8
92.4
100
79.0
83.8
TransFuser [ 37 ]
✓
✓
97.7
92.8
92.8
100
84.0
84.0
PARA-Drive [ 38 ]
✓
97.9
92.4
93.0
99.8
79.3
84.0
TABLE I: The Performance Comparison on NAVSIM using Closed-Loop Metrics.
Model
Mode
NC ↑
DAC ↑
TTC ↑
CF ↑
EP ↑
PDMS ↑
Non-Think SFT
w/o
98.1
92.1
94.0
100
77.1
83.3
Think SFT
w
98.5
94.4
94.9
100
79.9
86.2
Non-Think RL
w/o
98.2
96.1
94.3
100
83.5
88.3
Think RL
w
98.2
96.4
94.6
100
84.2
88.9
AdaThinkDrive
Ours
98.4
97.8
95.2
100
84.4
90.3
TABLE II : Comparison of Think and Non-Think Models in SFT and RL (InternVL3-8B). We compare the performance of SFT and RL models in Thinking (w) and Non-Thinking (w/o) modes. "Ours" refers to adaptive think mode.
Method
Infer Time ↓
PDMS ↑
Non-Think RL
0.68
88.3
Think RL
0.86
88.9
AdaThinkDrive
0.74
90.3
TABLE III : Comparison of inference time and PDMS among the Non-Think Model, Think Model, and AdaThinkDrive. Inference time denotes the average time to predict 4-second trajectories on the NAVSIM Test dataset.
Model
Level
NC ↑
DAC ↑
TTC ↑
CF ↑
EP ↑
PDMS ↑
Non-Think RL AdaThinkDrive
Level 1
98.5
96.3
95.0
100
83.0
88.5
98.8
98.1
96.1
100
84.5
90.7
Think RL AdaThinkDrive
Level 3
98.7
92.9
95.4
100
84.7
87.8
99.4
94.6
96.3
100
86.6
89.8
TABLE IV : Comparison of Think and Non-Think RL Models in Simple (Level 1) and Challenging (Level 3) Scenarios.
Fig. 5 : Qualitative comparisons of trajectory predictions on the NAVSIM dataset. We compare AdaThinkDrive with the Think RL model in the simple scenario ( top ), and with the Non-Think RL model in a challenging scenario (bottom). Each example shows the input image, driving command, reasoning content (if any), and the predicted trajectory. Green, blue, and red represent the ground-truth, AdaThinkDrive planning, and Think/Non-Think Model planning, respectively.
Model
NC ↑
DAC ↑
TTC ↑
CF ↑
EP ↑
PDMS ↑
SFT
98.5
94.4
94.9
100
79.9
86.2
Pre+SFT
98.9
95.3
96.0
100
80.6
87.5
Pre+SFT+RL
98.4
97.8
95.2
100
84.4
90.3
TABLE VI : Ablation Study on AdaThinkDrive Components. We evaluate the effect of pre-training, supervised fine-tuning, and reinforcement learning on driving performance using NAVSIM evaluation metrics.
ID
P. & F.
E.
A.
NC ↑
DAC ↑
TTC ↑
CF ↑
EP ↑
PDMS ↑
1
✓
96.7
95.2
90.7
100
87.8
88.1
2
✓
✓
98.3
96.6
94.5
100
84.4
89.1
3
✓
✓
✓
98.4
97.8
95.2
100
84.4
90.3
TABLE VII : Ablation Studies of Reward Designs in GRPO. We evaluate the effect of PDMS Reward (P.), Format Reward (F.), Endpoint Reward (E.), and Adaptive Think Reward (A.) on driving performance using NAVSIM evaluation metrics.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
Stage 1
Stage 2
Stage 3
Epochs
2
2
2
Batch size
1
2
4
Learning rate
1×10−5
4×10−5
2×10−6
Gradient accumulation steps
4
2
16
Weight decay
0.05
0.05
0.05
Number of GPUs
16
16
64
Appendix
TABLE VIII : Training hyperparameters. A dash indicates a setting not used in supervised training.
Fig. 6 : Paired SFT supervision: structured reasoning followed by the trajectory (top), and direct trajectory prediction with the same output structure (bottom).
Fig. 7 : Qualitative reasoning examples from the pretraining stage, covering scene understanding, road-boundary relevance, traffic lights, and critical-object identification.
Fig. 8 : Additional adaptive-planning examples after reinforcement learning. Each case includes the observation, navigation command, ego state and history, model response, and predicted trajectory.
Autonomous driving develops vehicles that perceive road scenes, predict interactions, and plan and control safe movement with limited or no human input.
State Key Laboratory of Networking and Switching Technology · Beijing University of Posts and Telecommunications, Beijing, China · School of Digital Media & Design Arts