LLM-Designed Autoencoders Fail to Beat a Baseline on RF Fingerprinting — and Cost More Training Time

arXiv 2607.21564·low signal

A systematic real-world evaluation of open-set RF fingerprint identification across five probe points on a BPSK receiver chain found probe placement dominates everything: timing recovery and, secondarily, carrier recovery enable low false-acceptance operation, while other stages need a false-acceptance ratio above 0.1 to hit a 0.9 true-acceptance ratio. To test whether the finding survives model selection, the authors benchmarked several LLM-designed autoencoders under a controlled pipeline holding preprocessing and MSE scoring fixed — those architectures confirmed the probe dependence but did not outperform the baseline at the chosen operating point and typically increased training time. A useful published negative result on LLM-as-architecture-designer.