SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model

Abstract: Vision-Language-Action (VLA) models offer a promising autonomous driving paradigm for leveraging world knowledge and reasoning capabilities, especially in long-tail scenarios. However, existing VLA models often struggle with the high latency in action generation using an autoregressive generation framework and exhibit limited robustness. In this paper, we propose SpanVLA, a novel end-to-end autonomous driving framework, integrating an autoregressive reasoning and a flow-matching action expert. First, SpanVLA introduces an efficient bridge to leverage the vision and reasoning guidance of VLM to efficiently plan future trajectories using a flow-matching policy conditioned on historical trajectory initialization, which significantly reduces inference time. Second, to further improve the performance and robustness of the SpanVLA model, we propose a GRPO-based post-training method to enable the VLA model not only to learn from positive driving samples but also to learn how to avoid the typical negative behaviors and learn recovery behaviors. We further introduce mReasoning, a new real-world driving reasoning dataset, focusing on complex, reasoning-demanding scenarios and negative-recovery samples. Extensive experiments on the NAVSIM (v1 and v2) demonstrate the competitive performance of the SpanVLA model. Additionally, the qualitative results across diverse scenarios highlight the planning performance and robustness of our model.

• The paper introduces a sparse-layer flow-matching action bridge with historical trajectory initialization, reducing 50-waypoint inference time from 2.57 to 0.67 seconds while raising NAVSIM v1 PDMS to 90.3.
• The paper uses GRPO reinforcement fine-tuning with positive, negative, and expert-recovery trajectories, improving performance from 82.1 to 90.3 PDMS on NAVSIM v1 and from 79.4 to 86.4 EPDMS on v2.
• The paper presents the 30K-scenario mReasoning dataset and shows that bounded reward shaping can reduce overly cautious driving and support corrective maneuvers, although deployment throughput and reward design remain open challenges.

Overview

SpanVLA is an end-to-end autonomous driving framework that couples a vision-LLM (VLM) backbone with a lightweight flow-matching action expert, and introduces a reinforcement fine-tuning (RFT) scheme that exploits real-world negative-recovery samples. The paper addresses two shortcomings of existing VLA driving models: the high latency of autoregressive action decoding, which scales linearly with trajectory length, and the reliance on imitation learning from positive demonstrations only, which limits robustness in long-tail scenarios. The system is built on Qwen2.5-VL-3B and evaluated on NAVSIM v1 and v2, where it reports state-of-the-art results among camera-only VLA methods alongside substantial inference-time reductions.

Architecture: VLM Backbone with Flow-Matching Action Bridging

The VLM backbone consumes multi-frame, multi-view camera streams (front, front-left, front-right by default), ego-state text tokens, and high-level navigation instructions. It performs autoregressive decoding to produce structured chain-of-thought reasoning, following AutoVLA's adaptive "fast/slow thinking" mechanism: for simple scenarios the model emits only action tokens, while complex scenarios trigger explicit CoT reasoning before planning. During supervised fine-tuning (SFT), discrete action tokens from a 2048-entry codebook are appended after reasoning to unify reasoning and planning within a single sequence-level language modeling loss.

The key architectural contribution is the efficient action bridging module. Rather than decoding trajectories token-by-token through the full VLM, SpanVLA extracts KV-cache features from sparse VLM layers (every two layers) and feeds them to a flow-matching action expert. Critically, unlike prior diffusion or flow-matching planners such as ReCogDrive and Alpamayo that denoise from pure Gaussian noise, SpanVLA initializes the flow from historical trajectory embeddings and learns the optimal-transport displacement map from past actions to future actions, with Gaussian noise injected into the historical embeddings during training for robustness. Gradients from the action bridge are stopped with respect to both the VLM backbone and the bridge during its separate fine-tuning stage, preserving learned representations.

The latency results are the strongest quantitative claims in the paper. On NAVSIM v1 navtest, autoregressive decoding of a 50-waypoint trajectory takes 2.57 s total, whereas SpanVLA's flow-matching expert generates the same-length trajectory in 0.67 s — a 74% reduction — while also improving PDMS from 89.1 (AutoVLA) to 90.3. Because generation is parallel rather than sequential, runtime is nearly invariant to waypoint count (0.08 s for both 10- and 50-waypoint outputs). Ablations show this design matters: last-layer-only feature extraction drops PDMS to 79.3, dense full-layer caching reaches 88.1 at higher cost, and historical initialization contributes +3.9 PDMS over the same sparse configuration without it.

GRPO-Based RFT with Negative-Recovery Samples

The second contribution targets the imitation-only training regime. The authors curate negative-recovery data — suboptimal real-world ego trajectories paired with expert corrections collected during early-stage exploratory testing — and incorporate them into Group Relative Policy Optimization (GRPO). Each sample is labeled positive, negative, or recovery, and the reward combines:

• Driving reward: PDMS (v1) or EPDMS (v2), used directly for positive samples.
• Negative-behavior penalty: a bounded L2-based shaping term penalizing proximity to the undesirable reference trajectory; bounding prevents the policy from escaping toward arbitrary deviated actions.
• Recovery-behavior reward: the same functional form applied with positive sign toward the expert correction, preserving multi-modality of recovery solutions.
• Reasoning penalties: a sigmoid length penalty on CoT tokens plus a rule-based action–reasoning alignment check that overrides the CoT reward with a fixed penalty on inconsistency.

Training follows a two-stage schedule: a 2K positive-only warm-up followed by 4K mixed steps under a fixed 6K RFT budget. Data-recipe ablations identify 2K warm-up + 3K positive + 0.5K negative + 0.5K recovery as the best mixture; increasing negatives beyond ~25% yields diminishing returns, and removing the warm-up degrades performance. Ablating the shaping terms shows PDMS alone is insufficient under P+N+R training because it provides no explicit directional signal away from negative modes, and performance is sensitive to both the shaping weight and the proximity threshold $\delta$.

The empirical effect is large: post-RFT SpanVLA improves from 82.1 to 90.3 PDMS on NAVSIM v1 and from 79.4 to 86.4 EPDMS on NAVSIM v2 navtest, exceeding DriveVLA-W0 (90.2 / 86.1) and all conventional end-to-end baselines including DiffusionDrive. On the harder navhard benchmark with pseudo closed-loop simulation, SpanVLA achieves 40.1 EPDMS versus 36.9 for RAP, with notably better No-Collision scores in stage two (86.9 vs. 83.2). Qualitative results indicate the negative samples reduce overly conservative behaviors (hesitant lane changes, yield-induced stopping) while recovery samples enable corrective maneuvers such as temporary lane borrowing around construction zones.

The mReasoning Dataset

mReasoning comprises 30K reasoning-demanding scenarios curated from in-house expert driving logs across Las Vegas, Boston, Pittsburgh, and Singapore, covering cut-ins, VRU interactions, construction zones, and stop signs. CoT annotations are generated via an automated pipeline built on Gemini-3-Pro that retains only scenario elements causally relevant to the ego decision, producing compact traces designed to avoid the verbosity of prior datasets. Human quality checking of 250 random samples yielded 80.2% accuracy under a strict binary protocol, with erroneous samples corrected or removed. The negative-recovery subset adds 3K + 3K scenarios; the authors state this is the first dataset with real-world negative-recovery samples. They also extend the NAVSIM evaluation pipeline to support PDMS-series scoring on mReasoning, enabling its use as an RFT environment. A limitation worth noting: the dataset derives from a single operator's fleet and has not yet been released as a public benchmark with a comprehensive test set.

Limitations and Open Questions

The paper is candid about deployment readiness. Current inference runs at roughly 1.5 Hz (33 ms per token for VLM encoding versus Alpamayo's optimized 1.75 ms), which the authors deem insufficient for direct real-world deployment, though they argue hardware-level optimization is orthogonal to their methodological contributions. Second, the effectiveness of negative-recovery learning depends on hand-designed bounded L2 shaping terms whose weight and activation threshold require careful tuning; designing more principled rewards for such data remains open. Third, the sparse-layer feature extraction trades accuracy for speed — interval-2 sampling outperforms interval-4 but lags full caching when historical initialization is absent — leaving the optimal granularity question unresolved. Finally, the human verification covered only 250 of 30K annotations, so residual annotation noise at scale is not quantified.

Conclusion

SpanVLA demonstrates that decoupling trajectory generation from autoregressive VLM decoding via sparse-layer flow-matching bridging with historical initialization simultaneously reduces inference time by up to 74% and improves planning scores, and that GRPO-based RFT over real-world negative-recovery samples provides gains beyond positive-only post-training. The combination of the mReasoning dataset and the bounded shaping rewards offers a concrete recipe for learning from failure data, though deployment-grade throughput and more principled reward design remain open problems the paper explicitly defers.