Async-SFT
Cross-source consensus on Async-SFT from 1 sources and 5 claims.
1 sources · 5 claims
How it works
Preparation
Comparisons
Evidence quality
Highlighted claims
- The paper uses a clock-based training setup to mimic asynchronous inference for small edge-scale models. — Building Interactive Real-Time Agents with Asynchronous I/O and Speculative Tool Calling
- Small open-source models require Async-SFT to reliably follow the asynchronous protocol. — Building Interactive Real-Time Agents with Asynchronous I/O and Speculative Tool Calling
- Normal supervised fine-tuning improves synchronous baselines but transfers poorly to AsyncIO behavior. — Building Interactive Real-Time Agents with Asynchronous I/O and Speculative Tool Calling
- Correction behavior is taught by adding erroneous early tool calls that must later be modified or removed. — Building Interactive Real-Time Agents with Asynchronous I/O and Speculative Tool Calling
- Synthetic training data are generated by segmenting user queries into streaming chunks and aligning speech timestamps with tool-call timing. — Building Interactive Real-Time Agents with Asynchronous I/O and Speculative Tool Calling