Zero-shot vision-language models (VLMs) such as CLIP demonstrate strong open-vocabulary recognition capabilities, yet their prediction confidences are often poorly calibrated. Existing calibration methods often estimate prediction stability by applying multiple augmentations in input space, introducing substantial computational overhead and limiting practical deployment. To this end, this paper proposes an embedding-space posterior calibration approach for zero-shot VLMs, which shifts the sampling process from the input space to the embedding space. Our approach defines and computes a confidence stability score that measures how stable the sample remains under posterior embedding perturbations. This score is further converted into a sample-wise temperature, so that unstable samples receive stronger confidence smoothing while stable samples remain close to their original predictions. Compared with the original CLIP model, our method reduces the mean Expected Calibration Error from 5.83 to 2.05, achieving a 64.8% relative reduction. Meanwhile, our method achieves this improvement with only a minor increase in GFLOPs, from 8.83 to 8.84.

Advancing from usable to collaborative autonomy requires driving systems to execute passenger instructions safely and reliably. This work formulates instruction realization as scheduling across multiple motion planners and presents a dual-loop framework that provides a transparent decision chain from natural language to vehicle control. The outer loop uses a small language model (SLM) for high-level, low-frequency semantic reasoning and schedule generation, while the inner loop performs low-level, high-frequency schedule execution and vehicle control. To compensate for the SLM’s limited capacity, the framework integrates receding-horizon scheduling to segment long-horizon instruction tasks, a domain-specific language (DSL) that restricts SLM outputs to a scheduling-oriented subspace, and reinforcement learning in high-fidelity urban traffic to refine the SLM’s DSL proficiency and scheduling performance. Experiments show that the framework improves instruction-completion rates while maintaining high safety and compliance relative to multiple baselines.