In‑vehicle assistants must turn natural‑language requests into vehicle function calls while operating under tight memory and latency limits, making small language models (SLMs) attractive for on‑device deployment. A core design choice is how the function surface is presented: either as dedicated Functional Tokens (FT) or by embedding function schemas directly in the prompt (Schema‑in‑Prompt, SIP).
FT assigns a unique token to each function, keeping inference compact but restricting calls to functions seen during training. SIP includes the full schema in the prompt, enabling generalization to unseen functions at the cost of longer prompts and higher inference overhead.
We built a benchmark covering 79 Android Automotive functions and 9,822 single‑turn examples, including seen functions, held‑out functions, and out‑of‑scope requests that require refusal.
Four SLMs ranging from 270 M to 1.7 B parameters were fine‑tuned under matched conditions. For functions seen during training, scaling yields limited gains: the 270 M model matches the 1.7 B model, and the best overall performance appears at 0.6 B. For held‑out functions, FT achieves zero accuracy by construction, whereas SIP generalizes and improves markedly with scale. For out‑of‑scope requests, FT may erroneously emit an unavailable function it was trained to produce, while SIP more reliably refuses based on the offered functions.
This flexibility comes with higher memory consumption and latency. Our theoretical analysis shows that SIP enables generalization by exposing explicit schemas, and longer schema contexts increase inference cost.
Overall, the way the function surface is represented, rather than model size alone, determines the capabilities and failure modes of SLM‑based vehicle function calling. Review