Streaming Intent Model SI Achieves Breakthrough in Autonomous Driving Control

A new paper formalizes action emergence as a target capability for autonomous driving: generating safe, semantically appropriate actions in arbitrary long-tail traffic scenes through reasoning rather than retrieval or interpolation. The authors argue previous paradigms fail—autoregressive decoders collapse multimodality, while diffusion/flow generators lack steerability by intent. Their solution, Streaming Intent, makes driving intent semantically streamed via continuous chain-of-thought and temporally streamed across clips for coherent commitments over the driving horizon. The resulting VLA model, SI, autoregressively decodes a four-step chain-of-thought to emit an intent token, then uses classifier-free guidance on a flow-matching action head with only two denoising steps.

On the Waymo End-to-End benchmark, SI achieves competitive performance with an RFS score of 7.96 on the validation set and 7.74 on the test set. More importantly, the model demonstrates—for the first time in a fully end-to-end VLA—intent-faithful controllability: for a fixed scene, varying the intent class at inference yields qualitatively distinct yet consistently high-quality plans. This arises purely from data-driven learning without any pre-built trajectory bank or hand-coded post-hoc selector. The work opens new directions for interpretable, intent-driven autonomous driving systems.