[CVPR 2026] LaST-VLA: Beyond Textual CoT—Thinking in Latent Spatio-Temporal Space for Autonomous Driving

Executive Summary

TL;DR: LaST-VLA is a groundbreaking Vision-Language-Action (VLA) model that moves the "reasoning" process of autonomous vehicles from error-prone text sequences into a physically-grounded Latent Spatio-Temporal Space. By distilling knowledge from world models and 3D foundation models, it achieves a record-breaking 91.3 PDMS on NAVSIM, proving that models can "think" more effectively in continuous vectors than in discrete words.

In the landscape of autonomous driving, this work represents a major shift from modular pipelines and explicit textual reasoning toward internally grounded latent intelligence, solving the "hallucination" problem that plagues current LLM-based drivers.

Problem & Motivation: The "Semantic Gap" and Redundant Over-thinking

Recent VLA models like DriveVLM use Chain-of-Thought (CoT) to explain driving decisions. However, the authors identify two fatal flaws:

1. Semantic-Perceptual Decoupling: Converting dense 3D visual data into sparse text (e.g., "there is a car") loses geometric precision, leading to "hallucinations" where the planner ignores visual evidence in favor of flawed linguistic logic.
2. Efficiency Bottleneck: Generating long strings of text before outputting a trajectory (tokens like "The road is clear...") adds hundreds of milliseconds of latency—unacceptable in high-speed driving.

While some researchers proposed "Implicit CoT" (latent reasoning), these often become "physics-agnostic" black boxes that are unstable during training because they lack intermediate constraints.

Methodology: The Core of Latent Spatio-Temporal CoT

LaST-VLA solves this by forcing the model's internal hidden states to align with the laws of physics.

1. Dual-Feature Alignment

The model extracts "Latent Thoughts" ($H$) and splits them into two streams:

• Geometry Stream ($H_{geo}$): Aligned with the VGGT (Visual Geometry Grounded Transformer) to capture 3D occupancy and depth.
• Dynamics Stream ($H_{dyn}$): Aligned with Cosmos (World Model) to predict how the scene evolves over time (forsight).

2. Progressive Two-Stage SFT

To ensure the model doesn't "cheat" by looking directly at the image for the final answer, the authors use Structured Causal Masking. In Stage 1, the model is forced to compress all visual information into these latent tokens. In Stage 2, it learns to plan trajectories based on these "grounded thoughts."

Figure 1: The LaST-VLA Architecture showing the alignment between internal hidden states and external foundation models.

Experiments & Results: Setting the New SOTA

The results are decisive. LaST-VLA doesn't just outperform other VLMs; it beats specialized end-to-end models that use LiDAR.

Key Performance Metrics:

• NAVSIM v1: Reached 91.3 PDMS, surpassing the previous best (Recogdrive) by a significant margin.
• Safety Gains: Significant improvements in No at-Fault Collision (NC) and Drivable Area Compliance (DAC), directly attributed to the geometric latent supervision.
• Spatial Reasoning: On the SURDS benchmark, it improved Yaw accuracy by 25% and Pixel location estimation by 10% over the InternVL3 base.

Table 1: Performance on NAVSIM v1. Note the superior PDMS and Safety metrics compared to UniAD and DriveVLA.

The Stability of Grounded Reasoning

One of the most impressive findings is shown in the RL training curves. Unsupervised latent reasoning (orange) oscillates wildly, while LaST-VLA's physically-grounded approach (red) converges stably to a higher performance peak.

Figure 2: RL training stability—physical supervision prevents model collapse and ensures steady performance gains.

Critical Analysis & Conclusion

Takeaways

LaST-VLA proves that the future of VLA models isn't necessarily in talking about the world, but in representing its physical properties internally. By shifting the reasoning to a continuous latent space, we get the best of both worlds: the high-level reasoning of LLMs and the metric precision of classical robotics.

Limitations

A key limitation noted by the authors is the reliance on a single front-view camera. Failure cases often involve sharp turns where the target trajectory leaves the camera's FOV. Future iterations will likely need surround-view (360-degree) inputs to maintain spatial grounding in complex intersections.

Future Outlook

This paradigm of "Latent Spatio-Temporal CoT" is highly extensible. We can expect future VLAs to incorporate even more specialized "thought" tokens, perhaps for legal rules or interactive social behaviors, moving us closer to a "Supervised Physical Intelligence."

Title Reference: LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (Luo et al., 2026)