AI News
All stories
Official

Zhejiang University releases LEMMO, a foundation model for electromagnetic signals

LEMMO classifies raw electromagnetic signals, including jamming and radar types, and beats the earlier MERLIN model on most perception tasks, according to the authors.

Research · · 2 min read

Researchers at Zhejiang University have released LEMMO, a model that classifies raw electromagnetic signals and generates jamming and anti-jamming signals. The authors say it outperforms MERLIN, a model published earlier this year by defense-affiliated Chinese researchers, on most perception tasks. Their arXiv preprint reports an aggregate perception score of 59.3 against 43.53 for MERLIN, and a generation score of 47.64 against 43.26. The code is published on GitHub. The repository notes that the reference runtime has been validated on NVIDIA H100 hardware, and that other CUDA accelerators and dependency versions have not yet been validated.

The model works on raw in-phase and quadrature (IQ) recordings. According to the abstract, existing waveform-based EM models are mostly tested on short IQ segments, while wideband recordings can contain millions of samples. LEMMO pairs a linear-attention encoder with anti-aliased temporal resampling, so it can process up to eight million complex samples. It uses learnable queries conditioned on a text instruction before they pull evidence from the signal. Compact queries handle recognition, while structured queries with auxiliary factual supervision handle detailed descriptions. The GitHub README lists two inference variants: a recognition variant accepting lengths from 128 to 5,000,000 samples, and a description variant accepting 16 to 8,000,000.

The classification tasks include communication-jamming recognition, modulation recognition, radar bandwidth, duty cycle, number of pulses, pulse repetition frequency and pulse width, protocol identification, radar-jamming recognition and segment detection. The generation tasks are anti-communication-jamming, anti-radar-jamming, and communication and radar jamming strategies. The results are uneven across tasks. On perception, LEMMO beats MERLIN and other models on every task except segment detection, where Gemini-3.5 Flash Lite wins. On generation, LEMMO has the top scores on anti-radar-jamming and radar jamming, while MERLIN wins the others.

The abstract compares LEMMO with a variant called MERLIN-LoRA and reports gains in mean EM-Bench perception accuracy of 15.77 percent, mean strategy-generation ROUGE-L F1 of 4.38 points, and mean LEMMO-ShortQA accuracy of 14.31 percent. The abstract does not say how the aggregate scores in the summary are weighted across tasks. The paper also says it releases a unified 1.38-TB corpus of real-world and synthetic signals. The repository's README validates lengths and hardware for its reference runtime, and says other CUDA accelerators and dependency versions have not yet been validated.

Import AI, Jack Clark's newsletter, framed the work as a sign that general-purpose generative pretraining is spreading into military signal analysis. It described LEMMO as the kind of model a military will use to characterise the electromagnetic battlespace.

What changed

Zhejiang University released LEMMO, a foundation model that classifies raw electromagnetic signals and generates jamming and anti-jamming signals.

Type: New capability · Applies to: LEMMO

Why it matters

General-purpose pretraining is being applied to signal analysis, which widens the set of domains where foundation models can be used for sensing and electronic warfare. Companies in defence and telecoms should expect more such models.

Original source

Where we saw it

1 source

Import AI (Jack Clark)

Zhejiang University

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works