LEMMO connects long radio recordings to a language model—but detailed accuracy remains limited

Why longer recordings matter

The study, by Mingkai Li, Jiajun Wu and colleagues at Zhejiang University, works directly with in-phase and quadrature, or IQ, recordings. These paired measurements preserve sampled amplitude and phase information, rather than representing radio activity only as an image.

High sampling rates make even brief recordings large: the paper notes that 10 milliseconds at 100 million samples per second produces one million complex samples. Short excerpts may miss later bursts or relationships between transmissions. LEMMO targets recordings in which the relevant signals’ times and frequency bands are not known beforehand.

How LEMMO processes the signal

LEMMO divides the waveform into patches of 16 complex samples. A 12-layer encoder uses linear attention, whose computational cost grows approximately in proportion to sequence length. Before language training, the encoder learns by reconstructing masked waveform patches.

Recordings exceeding one million samples undergo additional compression. Each million-sample region produces 62,500 patch features, which downsampling and pooling reduce to 800. Low-pass processing is intended to reduce aliasing during downsampling. Two bidirectional attention layers then exchange information across regions. At the maximum supported length, eight million samples become 6,400 signal features.

Learnable queries connect those features to the question. They first attend to the instruction, then retrieve signal evidence. Recognition uses 16 queries. Detailed description uses 128, grouped around the overall scene, individual signal objects and relationships between them.

Additional training objectives teach the description representations to capture categories, counts, numerical attributes and relationships. These auxiliary predictions are not supplied as facts to the language model during generation. Responses come from Qwen3.5-4B, adapted through LoRA, which trains a limited set of added parameters while freezing the base language model.

What the researchers tested

The authors report assembling and releasing a 1.38-terabyte corpus combining real and synthetic signals. Their LEMMO-ShortQA dataset covers modulation, jamming patterns, drone-related radio signals and wireless technologies. Its test split contains 846 recordings with four question phrasings each.

LEMMO-LongDesc tests scene overviews and individual-signal descriptions using 802 description questions from 652 recordings. Four of its five signal families—LTE, 5G NR, basic modulation and industrial, scientific and medical-band scenes—are MATLAB-generated. The drone scenes use recorded signals.

The team also evaluates EM-Bench, an existing benchmark with 4,200 questions across 14 tasks. Importantly, this experiment uses a separate direct encoder-to-language-model interface, not the 16-query or 128-query modules. The paper therefore reports several task-specific configurations, rather than demonstrating one unchanged configuration across every benchmark.

Results: stronger recognition, incomplete descriptions

On EM-Bench, LEMMO reaches 59.30% mean perception accuracy versus 43.53% for MERLIN-LoRA: a gain of 15.77 percentage points. Its average strategy-generation ROUGE-L score rises from 43.26 to 47.64. ROUGE-L measures overlap with reference wording; it does not establish that a proposed strategy works operationally.

On ShortQA, average recognition accuracy increases from 55.74% to 70.05%. The largest improvement is drone recognition, from 41.88% to 73.75%, using five-million-sample recordings.

Description results are more qualified. For scene overviews, LEMMO scores 44.83 on both categorical-fact and numerical-fact F1. For individual-signal descriptions, those scores are 41.42 and 24.06. Both improve on MERLIN-LoRA, but remain far from complete factual recovery. In one illustrated 5G example, LEMMO identifies many signal properties correctly while reporting zero visible terminals instead of three and zero resource occupancy instead of 29%.

Longer observations generally help in a separate input-length experiment. However, numerical-fact F1 does not improve at every step: it falls at one million samples before reaching its highest reported value at four million.

What the evidence does not establish

The long-description comparison has a major constraint: MERLIN-LoRA receives only the first 4,096 samples, while LEMMO receives the complete recording. Those results measure the systems under different observation limits, not architectural quality with equal access to signal evidence.

The language backbones also differ. LEMMO uses Qwen3.5-4B; MERLIN-LoRA uses Qwen3-4B-Instruct-2507. The reported comparisons cannot isolate the contribution of each design choice.

Much of the long-description dataset is simulated, limiting what these tests establish about unfamiliar real-world environments. The paper reports no confidence intervals or real-time inference measurements. Its near-unchanged MMLU score supports retention on that particular knowledge benchmark, not a comprehensive claim about general capabilities. Overall, the preprint supports improved benchmark performance for long-recording interpretation, while its factual errors leave reliable detailed description unresolved.