Reasoning-aware Speculative Decoding for Efficient Vision-Language-Action Models in Autonomous Driving

2026-06-30

AI Summary

This work addresses the high latency in causal chain reasoning exhibited by vision-language-action (VLA) planners in autonomous driving. To mitigate this, the authors propose an asymmetric dual-reasoner architecture: a routine reasoner rapidly generates standard actions based on trajectory history, while a deliberative reasoner handles unexpected scenarios requiring fresh visual evidence. The two reasoners operate synergistically within an enhanced speculative decoding framework. Additionally, FlatRoPE positional encoding is introduced to direct attention toward historical trajectories, and action-aware reinforcement learning (AARL) is employed for post-training optimization. This approach maintains decision quality while reducing inference time per reasoning step by approximately fourfold compared to the original Alpamayo planner, substantially improving real-time performance.

Abstract

Modern Vision-Language-Action (VLA) planners for autonomous driving emit a chain-of-causation (CoC) reasoning step \emph{before} producing a trajectory. The reasoning is autoregressive and dominates inference latency, while the trajectory head is parallel and cheap. Latency is an operational constraint in autonomous driving, so accelerating the reasoning step is the central problem we address. We observe that CoC reasoning has two qualitatively different needs: most tokens continue routine setup that follows naturally from the ego-trajectory history, and a small fraction encode commitments that require fresh visual evidence about an unexpected situation. We split this reasoning into two specialized paths: a \emph{routine reasoner} that handles the predictable continuation by attending to trajectory history, and a \emph{deliberative reasoner} (the unmodified VLA target) that handles novel cases by attending to current visual evidence, using the speculative decoding framework as the architectural template for how the two paths cooperate. Unlike standard speculative decoding, our routine reasoner is not a smaller replica of the target; the two reasoners are deliberately specialized to read different parts of the prompt. We propose two techniques to realize this. First, we introduce \textbf{FlatRoPE}, a 1D rotary positional embedding in the draft that breaks the rotational symmetry of the target's 3D M-RoPE, redirecting attention away from visual tokens and onto trajectory-history tokens. Second, we introduce \textbf{Action-aware RL (AARL)}, a post-training stage that uses an action-quality reward together with a static-reference KL anchor. Together, our two-reasoner system reduces the reasoning-step running time by approximately $4\times$ relative to the original Alpamayo planner.

Problem

• speculative decoding
• vision-language-action models
• autonomous driving
• reasoning latency
• chain-of-causation

Innovation

• Speculative Decoding
• Vision-Language-Action Models
• FlatRoPE
• Action-aware RL
• Autonomous Driving