Shanghai Jiao Tong's RadioMaster Turns Plain English Into Working Radio Signals
RadioMaster uses three specialized AI agents to turn plain-language prompts into real, verified radio signals on physical SDR hardware -- no RF expertise required
Takeaways
- RadioMaster is a multi-agent framework that translates plain-language prompts into verified, working radio signals on real SDR hardware.
- Three agents collaborate: RadioWiki (domain knowledge retrieval), RadioAgent (I/Q sample and code generation), and RadioEmulator (closed-loop physical verification).
- The system covers 10 wireless protocols (Wi-Fi, BLE, LTE, 5G NR, LoRaWAN, Zigbee, and more) and mainstream SDR platforms like USRP and ADALM-Pluto.
- RadioBench, the first benchmark for radio signal generation, tests models across 3 difficulty levels using QAA, CPR, HDR, and SIR metrics.
- RadioMaster significantly outperforms all tested frontier LLMs (GPT, Claude, Gemini, DeepSeek, Qwen) on configuration viability and signal fidelity.
- No public code or model checkpoint has been released; the paper is available on arXiv.
Writing code that makes a radio chip transmit a Bluetooth beacon or a Wi-Fi frame is genuinely hard. You need to understand the protocol spec, translate it into digital baseband math, generate the right I/Q samples (the raw in-phase and quadrature numbers that represent the waveform), and then configure the hardware without violating its physical constraints. A new paper from researchers at Shanghai Jiao Tong University introduces RadioMaster, a fully autonomous multi-agent framework that collapses that entire pipeline into a single natural-language prompt.

Why LLMs alone fall flat on radio
The intuition that a powerful LLM could just "write the SDR code" turns out to be wrong in practice. The traditional manual workflow is notoriously tedious and prone to errors -- human operators must manually parse thick protocol specifications, mathematically translate them into digital baseband algorithms, and subsequently write intricate hardware execution code to configure SDR platforms. When the researchers tested current frontier models on this task, they found two fundamental failure modes:
- Domain ignorance: Current LLMs rely on generalized training corpora, lacking deep domain expertise and missing recent wireless protocol advancements.
- Hardware blindness: Models generate code that looks syntactically correct but violates physical constraints of the target SDR -- wrong sample rates, illegal gain values, unsupported center frequencies.
This disjointed process creates a severe bottleneck, significantly raising the barrier to entry and stifling rapid prototyping in wireless systems. The gap between "write me a BLE beacon" and a signal that actually comes out of an antenna is where every existing model fails.
Three agents, one pipeline
RadioMaster is built on three tightly coupled components that each tackle one of the failure modes. Think of it as a small team: one agent knows the theory, one writes and debugs the code, and one physically checks that the signal is correct before declaring success.