Async-SFT

Normal supervised fine-tuning improves synchronous baselines but transfers poorly to AsyncIO behavior. Small open-source models require Async-SFT to reliably follow the asynchronous protocol. The paper uses a clock-based training setup to mimic asynchronous inference for small edge-scale models. Correction behavior is taught by adding erroneous early tool calls that must later be modified or removed. Synthetic training data are generated by segmenting user queries into streaming chunks and aligning speech timestamps with tool-call timing.