LEMMO
A large electromagnetic model for instruction-driven recognition and description that combines pretrained language models with electromagnetic signal processing, supporting understanding of long raw IQ recordings with up to eight million complex sample points.
Timeline (last 90 days)
LEMMO is a large electromagnetic model for instruction-driven recognition and description, specifically handling long raw IQ recordings.
LEMMO's architecture includes a linear-attention electromagnetic encoder and anti-aliased temporal resampling, capable of handling up to eight million complex sampling points.
Before retrieving signal evidence, LEMMO's electromagnetic-text interaction module conditions learnable queries on the instruction; compact queries are used for recognition tasks, and structured queries are used to organize scene-level attributes, signal attributes, and cross-signal relationships to support detailed descriptions.
Compared with MERLIN-LoRA, LEMMO improves average perception accuracy on EM-Bench by 15.77%.
Compared with MERLIN-LoRA, LEMMO improves ROUGE-L F1 on strategy generation tasks by an average of 4.38 points.
Compared with MERLIN-LoRA, LEMMO improves average accuracy on LEMMO-ShortQA by 14.31%.
On the two tasks of LEMMO-LongDesc, LEMMO improved both classification and numerical fact F1.
While improving electromagnetic signal understanding, LEMMO preserves the language backbone's MMLU performance.
The paper abstract does not disclose the model parameter count, training cost, or robustness details under different signal-to-noise ratio conditions.
The actual deployment efficiency of long-recording understanding and generalization to unknown signal types remain directions requiring further verification.
All Facts (10)
LEMMO is a large electromagnetic model for instruction-driven recognition and description, specifically handling long raw IQ recordings.
50%UnverifiedLEMMO's architecture includes a linear-attention electromagnetic encoder and anti-aliased temporal resampling, capable of handling up to eight million complex sampling points.
50%UnverifiedBefore retrieving signal evidence, LEMMO's electromagnetic-text interaction module conditions learnable queries on the instruction; compact queries are used for recognition tasks, and structured queries are used to organize scene-level attributes, signal attributes, and cross-signal relationships to support detailed descriptions.
50%UnverifiedCompared with MERLIN-LoRA, LEMMO improves average perception accuracy on EM-Bench by 15.77%.
50%UnverifiedCompared with MERLIN-LoRA, LEMMO improves ROUGE-L F1 on strategy generation tasks by an average of 4.38 points.
50%UnverifiedCompared with MERLIN-LoRA, LEMMO improves average accuracy on LEMMO-ShortQA by 14.31%.
50%UnverifiedOn the two tasks of LEMMO-LongDesc, LEMMO improved both classification and numerical fact F1.
50%UnverifiedWhile improving electromagnetic signal understanding, LEMMO preserves the language backbone's MMLU performance.
50%UnverifiedThe paper abstract does not disclose the model parameter count, training cost, or robustness details under different signal-to-noise ratio conditions.
50%UnverifiedThe actual deployment efficiency of long-recording understanding and generalization to unknown signal types remain directions requiring further verification.
50%