LLM-designed autoencoders failed to beat a fingerprinting baseline

LLM-designed autoencoders lost to the baseline in a test of RF fingerprint identification, per the paper posted to arXiv. That matters for anyone weighing whether an LLM can shortcut architecture design. Here, the AI models added training time without adding accuracy.

The autoencoder comparison wasn't the paper's main result. It was a check on whether the real finding, that probe placement decides fingerprinting accuracy, would survive a change in model architecture.

The team tested five probe points on a BPSK receiver chain, then benchmarked LLM-designed autoencoders with preprocessing and MSE scoring held fixed.

The LLM architectures confirmed the same probe-placement effect. None beat the baseline at the chosen operating point, and training typically took longer.

The take

Handing architecture design to an LLM doesn't guarantee a better model. Here it produced a slower one that proved the baseline was already right. The open question is whether this holds outside RF fingerprinting, where LLM-designed models get judged against a baseline instead of praised for novelty.

That kind of result rarely makes it into a paper. Most published work reports what worked, not what a fancier approach failed to beat.