Researchers led by Boris Radovič explored a middle ground between cloud‑only AI agents and tiny on‑device models. The cloud writes a program, but when it needs to interpret user data it calls a small local model. Tests on the AppWorld benchmark showed the hybrid design lifted the code‑only agent’s success rate in all six model‑partition combos, adding as much as 14.9 percentage points and closing much of the gap to a fully cloud‑hosted ReAct loop while preserving privacy.
Why it matters
Edge device users can get more accurate AI assistants without sending their private data to the cloud.
