Small language models (SLMs) are attractive as local controllers for agents because they cut remote inference, latency, and deployment footprint, yet structured tool errors can cause a step to fail. Existing routers pick a model once per query, while an agent’s decision points form a sequence whose difficulty changes with intermediate observations. We therefore propose STEPGATE, an uncertainty‑aware handoff framework that scores each local SLM action and escalates challenging steps to a stronger model. On a 52‑task single‑step BFCL test split, the Qwen2.5‑1.5B/7B pair achieves 82.7% task success with a 30.8% escalation rate, far above local‑only (67.3%) and random escalation (33.8% escalation) at 75.4%. In a multi‑turn evaluation, STEPGATE attains 69.0% trajectory success and 84.0% action success while using only 30.0% cloud actions, outperforming local‑only (48.0%/70.5%), random escalation (60.0%/78.2%) and query‑level routing (57.0%/77.1%). The strong‑only backend reaches 82.0% trajectory success at 100% cloud usage. These results suggest that step‑level escalation recovers most of the performance gap to the stronger Qwen2.5‑7B backend at a matched cloud‑action rate while transmitting fewer tokens. Limitations include evaluation on a single model family and a single stronger backend, scripted tasks, small test sets, multi‑turn comparisons that rely on paired intervals and statistical tests, and risk tiers that serve as research annotations rather than formal safety guarantees.
Review