STEER: Route-Aware Adaptive Reasoning for Autonomous Driving
Authors: Yangang Zou, Pei Liu, Nan Song, Bozhou Zhang, Mingyu Guo, Jun Ma, Jiankang Deng, Xiatian Zhu, Li Zhang
Organizations: School of Data Science, Fudan University · Shanghai Innovation Institute · Logos Intelligence Lab · HKUST(GZ) · HKUST · Li Auto Inc. · Imperial College London · University of Surrey
Vision-language action (VLA) models excel for end-to-end autonomous driving, yet their tendency to over-invoke chain-of-thought (CoT) reasoning mirrors a human cognitive pitfall: overthinking. Just as deliberate reasoning can slow or mislead human judgment in routine tasks, excessive CoT in VLA incurs unnecessary compute overhead and can paradoxically degrade planning performance. Recent research tackles this via adaptive reasoning, which dynamically modulates inference complexity by selectively triggering CoT for complex scenes while responding directly to routine ones. While prior adaptive methods rely on implicit adaptation, recent work shows that explicitly optimizing the route between CoT and direct responses yields superior performance, assuming scene-difficulty annotations which are labor-intensive, poorly scalable, and prone to error due to the subjectivity of judging driving complexity. We explore, for the first time, whether adaptive reasoning with explicit route optimization can be achieved without any scene-difficulty labeling. We introduce STEER, a generic over-reasoning mitigation strategy featuring two key components: (i) routing uncertainty-aware rollout, which calibrates sampling to align routing diversity with model uncertainty; and (ii) cross-route advantage credit assignment, which introduces a differential metric to reinforce optimal routing decisions based on environmental rewards rather than subjective labels. Extensive experiments on NAVSIM v1/v2 demonstrate that STEER is competitive with state-of-the-art methods in both planning quality and the Pareto front of inference efficiency.
Figures & tables
Figure 1: Overview of STEER. STEER is built on an autoregressive vision-language model that takes a front-view driving image together with structured driving context as input, and generates a textual future ego trajectory as output. It formulates adaptive reasoning in autonomous driving as an explicit routing problem, where an initial decision selects either a reasoning route or a direct route before trajectory generation. The framework is trained in three stages: (1) driving-centric pretraining for domain adaptation, (2) route-balanced supervised fine-tuning to preserve support for both valid routes, and (3) route-aware policy optimization to learn the routing preference directly from task rewards, without any scene-difficulty labels.
Figure 2: Route-aware policy optimization in STEER. For each prompt, STEER first inspects the initial routing distribution to estimate routing uncertainty. When the routing decision is uncertain, the model performs route-contrastive rollout by explicitly sampling both the reasoning and direct routes under the same prompt; otherwise, it falls back to standard on-policy rollout. During optimization, STEER applies a standard content advantage to all response tokens and adds a cross-route advantage at the initial routing position, jointly optimizing response quality and route preference.
Sensors
Trajectory
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
Human GT
Human [ 4 ]
-
-
100.0
100.0
87.5
100.0
99.9
94.8
End-to-End Methods
Ego-MLP [ 16 ]
6x C
Continuous
93.1
78.3
63.2
84.0
100.0
66.4
UniAD [ 7 ]
6x C
Continuous
97.8
91.9
78.8
92.9
100.0
83.4
PARA-Drive [ 29 ]
6x C
Continuous
97.9
92.4
79.3
93.0
99.8
84.0
Table 1: Performance on NAVSIM v1 navtest using closed-loop metrics. ∗ denotes fine-tuned on the NAVSIM trajectory dataset. † denotes Best-of- N evaluation with six generated candidates.
Method
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
HC ↑
EC ↑
EPDMS ↑
VADv2 [ 3 ]
97.3
91.7
98.2
99.9
77.6
92.7
66.0
100
97.4
76.6
TransFuser [ 21 ]
97.7
92.8
98.3
99.9
79.2
92.8
67.6
100
95.3
77.8
HydraMDP [ 15 ]
97.5
96.3
98.3
99.9
80.1
93.0
65.5
100
97.4
79.8
HydraMDP++ [ 15 ]
97.9
96.5
98.9
100
79.2
93.4
67.2
100
97.7
80.6
ARTEMIS [ 5 ]
98.3
95.1
98.6
99.8
81.5
97.4
96.5
100
98.3
83.1
ReCogDrive [ 14 ]
98.3
95.2
99.5
99.8
87.1
97.5
96.6
98.3
86.5
83.6
Table 2: Performance on NAVSIM v2 navtest with extended metrics.
Training Stage
Route Policy
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
Avg. Resp. Tokens ↓
SFT
Forced direct
97.7
94.5
80.2
93.4
99.9
85.3
155
Forced reasoning
97.7
94.4
80.1
93.7
99.9
86.4
417
Adaptive
97.9
94.4
80.3
94.2
100.0
86.5
274
RL
Forced direct
97.6
96.5
89.0
92.1
99.9
89.9
156
Forced reasoning
97.5
97.0
90.3
91.9
100.0
90.8
415
Adaptive
97.4
97.3
90.2
92.3
100.0
91.2
315
Table 3: Adaptive routing versus fixed routing on NAVSIM v1. Adaptive routing is compared against forced direct and forced reasoning policies after route-balanced SFT and after RL.
Figure 3: Prompt-level route comparison. For each prompt, we compare the trajectory produced by the route selected by STEER (blue) with the counterfactual alternative route (red). In the top case, explicit reasoning yields safer obstacle clearance. In the bottom case, STEER correctly avoids over-reasoning and selects the direct route, which better follows the road geometry compared to the perturbed reasoning alternative.
ID
Routing Uncert.-
Cross-Route
RouteAcc (%) ↑
AUC (%) ↑
AvgRegret ↓
Avg. Resp. Tokens ↓
PDMS ↑
Aware Rollout
Advantage
1
×
×
69.3
73.9
0.034
363
89.5
2
✓
×
70.7
74.4
0.023
362
90.9
3
×
✓
77.5
76.5
0.024
350
89.9
4
✓
✓
81.1
79.1
0.018
315
91.2
Table 4: Ablation of route-aware policy optimization . Our proposed mechanisms effectively overcome exploration collapse and confounded credit assignment, achieving the best routing quality and planning performance.
ID
Length Norm.
Route Calib.
Reasoning Prob. (%)
Direct Prob. (%)
PDMS ↑
1
×
×
26.6
73.4
86.4
2
✓
×
43.2
56.8
86.1
3
✓
✓
49.1
50.9
86.5
Table 5: Ablation of route-balanced SFT. Length normalization and routing calibration alleviate the direct-route bias.
ID
Pretraining
RL
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
1
×
×
97.8
94.1
79.9
94.0
100.0
85.9
2
✓
×
97.9
94.4
80.3
94.2
100.0
86.5
3
✓
✓
97.4
97.3
90.2
92.3
100.0
91.2
4
×
✓
97.5
97.3
90.2
92.5
100.0
91.2
Table 6: Ablation of the training pipeline. Planning performance with and without driving-centric pretraining and RL.
Point
Mode A
Mode B
Mode C
Reasoning ratio
Start
85%
13%
2%
50%
Middle
44%
35%
21%
57%
End
17%
22%
61%
64%
Table 7: Rollout-mode dynamics during route-aware policy optimization. Mode columns give percentages of prompts.
Route policy
PDMS ↑
Time (s/sample) ↓
Forced direct
89.9
2.4
Forced reasoning
90.8
6.7
STEER
91.2
5.1
Table 8: Efficiency analysis. End-to-end inference latency (s/sample, left) and RL training cost (GPU-hours, right).
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Prompt template for reasoning-data generation.
Figure 5: Training data template used for driving-centric pretraining and SFT.
SFT
Route-aware policy optimization
Parameter
Value
Parameter
Value
Base model
Qwen2.5-VL-8B
Rollout group size n
16
Max pixels
2073600
Clipping coefficient ϵc
0.2
Global batch size
256
Minority-route threshold ϵ
0.05
Epochs
5
Low-entropy threshold τlow
0.25
Learning rate
4.0×10−5
High-entropy threshold τhigh
0.50
Appendix
Table 9: Main hyper-parameter settings for route-balanced SFT and route-aware policy optimization.
Setting
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
STEER
97.41±0.36
97.25±0.55
90.22±0.81
92.25±1.10
100.0±0.0
91.18±0.25
STEER (Best-of-6)
98.02±0.19
97.73±0.31
91.61±0.32
94.36±0.51
100.0±0.0
93.44±0.12
Appendix
Table 10: NAVSIM v1 evaluation statistics. Mean ± standard deviation over three evaluation seeds.
Metric
Score
Metric
Score
NC ↑
97.40±1.21
TTC ↑
97.12±0.92
DAC ↑
97.63±0.58
LK ↑
95.61±1.31
DDC ↑
98.05±1.10
HC ↑
97.81±0.08
TLC ↑
99.15±0.15
EC ↑
84.47±1.12
EP ↑
95.62±0.22
EPDMS ↑
87.15±0.33
Appendix
Table 11: NAVSIM v2 evaluation statistics for STEER. Mean ± standard deviation over three evaluation seeds.
Setting
RouteAcc (%) ↑
AUC (%) ↑
AvgRegret ↓
Avg. tokens ↓
PDMS ↑
Vanilla GRPO
69.33±0.68
73.91±0.31
0.034±0.005
363±6
89.45±0.31
STEER
81.09±0.34
79.06±0.40
0.018±0.002
315±4
91.18±0.25
Appendix
Table 12: Matched comparison with vanilla GRPO. Mean ± standard deviation over three evaluation seeds for IDs 1 and 4 of Table 4 .
Setting
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
With pretraining
97.41±0.36
97.25±0.55
90.22±0.81
92.25±1.10
100.0±0.0
91.18±0.25
Without pretraining
97.45±0.40
97.31±0.34
90.17±0.63
92.54±0.90
100.0±0.0
91.20±0.28
Appendix
Table 13: Pretraining ablation after route-aware RL. Mean ± standard deviation over three evaluation seeds; both settings use route-balanced SFT and the same RL procedure.
λ
Reasoning Prob. (%)
Direct Prob. (%)
Routing Uncertainty ↑
PDMS ↑
0
43.2
56.8
0.681
86.1
0.05
48.5
51.5
0.689
86.0
0.10
49.1
50.9
0.693
86.5
0.15
49.4
50.6
0.692
86.0
0.20
49.6
50.4
0.692
85.9
Appendix
Table 14: Sensitivity analysis of the route-balanced SFT calibration weight λ . Moderate route calibration improves route balance and yields the best downstream planning performance.
λr
RouteAcc (%) ↑
AUC (%) ↑
AvgRegret ↓
Avg. Resp. Tokens ↓
PDMS ↑
0
70.7
74.4
0.023
362
90.9
0.02
77.8
77.4
0.020
333
91.1
0.05
81.1
79.1
0.018
315
91.2
0.10
80.1
76.8
0.029
322
90.4
0.20
77.2
74.7
0.039
324
89.4
Appendix
Table 15: Sensitivity analysis of the RL route loss weight λr . A moderate route loss weight provides the best routing quality, efficiency, and planning performance.
Figure 6: Qualitative comparison with a representative baseline. We compare STEER (blue) with ReCogDrive [ 14 ] (red) on representative prompts. STEER produces trajectories that better align with the road geometry and intended maneuver.
Figure 7: Representative failure cases on the NAVSIM benchmark.
Figure 8: Additional qualitative results on the NAVSIM benchmark.
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.
Yuechen Luo, Fang Li, Shaoqing Xu +11
Tsinghua University · Xiaomi EV · Nanyang Technological University +2
Existing adaptive reasoning methods for driving Vision-Language-Action (VLA) models primarily focus on whether to reason, overlooking how reasoning should differ across driving situations. Our key insight is that while scene complexity informs reasoning depth, dynamic risk is equally critical for deciding how to reason in time-critical situations. We therefore propose CAR-VLA, a unified driving VLA model that jointly considers scene complexity and dynamic risk to guide reasoning depth, urgency, and focus. CAR-VLA maps four complexity--risk categories to three reasoning modes: Fast Intuition for direct trajectory generation in simple low-risk scenes, Slow Thinking for deliberate reasoning in complex low-risk scenes, and Reflex Response for compact, hazard-focused reasoning in high-risk scenes regardless of complexity. Rather than merely shortening deliberation, Reflex Response centers reasoning on the most critical hazard and the immediate safe response. We train CAR-VLA through progressive supervised learning that links scene assessment, reasoning-mode selection, and trajectory generation, followed by reasoning-augmented reinforcement learning to improve driving quality and reasoning behavior. Experiments on NAVSIM v1(91.1 PDMS), NAVSIM v2(90.3 EPDMS), and Navhard(35.0 EPDMS) demonstrate competitive driving performance. Qualitative comparisons on navtest and in-house high-risk scenarios further illustrate risk-aware reasoning and hazard-responsive trajectory generation. The code for this paper will be released publicly at: https://github.com/chenxl124578/CAR-VLA.git
Xiaolei Chen, Zhuolin He, Yuxuan Liang +13
Fudan University · Yinwang Intelligent Technology Co., Ltd · Fuzhou University +2
Vision-language models (VLMs) have advanced scene understanding and enabled explicit reasoning in end-to-end autonomous driving. However, existing methods insufficiently integrate spatial-physical evidence into planning reasoning, while reasoning adaptation remains coarse-grained and falls short of scene-specific planning demands. Furthermore, reasoning-path optimization for higher planning quality remains largely unexplored in autonomous-driving post-training. To address these limitations, we propose FactorDrive, an end-to-end autonomous driving framework for adaptive multi-step reasoning driven by planning-critical factors (PCFs). We first perform large-scale driving-domain instruction tuning to establish foundational driving knowledge. Building on this foundation, we construct PCF-CoT, a chain-of-thought (CoT) dataset that grounds planning reasoning in trajectory-relevant spatial-physical evidence and organizes reasoning around scene-specific PCFs, enabling the composition and depth of reasoning paths to adapt to different planning demands. We further introduce Quality Search-Guided Group Relative Policy Optimization (QS-GRPO), which guides Monte Carlo Tree Search (MCTS) with trajectory-level planning rewards to discover reasoning paths with higher planning quality and uses the resulting responses to optimize the policy through GRPO, thereby improving trajectory planning performance. Extensive experiments on both open-loop (nuScenes) and closed-loop-oriented (NAVSIM) benchmarks demonstrate that FactorDrive achieves state-of-the-art planning performance.
Guolei Huang, Tengfei She, Yuxuan Lu +3
1Southeast University · 2Universiti Malay · 3Ant Group +2