Electromagnetic signal modeling is increasingly moving toward self-supervised pretraining on heterogeneous I/Q observations. However, an observed waveform is jointly shaped by signal generation, propagation environments, and acquisition processes, which mix reusable structural regularities with observation-specific variations. Existing methods usually pursue unified pretraining with reconstruction objectives on mixed observations. This heterogeneity creates two key mismatches for reconstruction pretraining: (1) appearance mismatch, induced by propagation-dependent waveform variations; and (2) temporal mismatch, induced by acquisition-dependent sampling coordinates. These mismatches make observation-specific appearances and sampling coordinates easier to reconstruct than stable signal structures, inducing observational shortcuts. We identify this failure mode as observation-biased pretraining. To address this, we propose Structural Pretraining with Time Alignment for heterogeneous EM signals. It mitigates temporal mismatch with RoPE-fs, which maps token positions to sampling-rate-normalized physical-time positions, and reduces reliance on appearance shortcuts by replacing waveform reconstruction with masked prediction of discrete structural units. These units are organized into atomic structural units and emergent structural units that capture variable-length compositional patterns beyond fixed patch boundaries. Across diverse EM benchmarks, our method improves both full-supervision and few-shot adaptation over existing EM pretraining baselines, showing that the learned representations better preserve structural consistency when observations drift in waveform appearance and temporal scale.
Figures & tables
Figure 1: Motivation and diagnosis of observation-biased EM pretraining. Heterogeneous EM observations arise from source generation, propagation, and acquisition, causing appearance and temporal mismatches. Reconstruction objectives can easily exploit these observation-specific cues, leading representations to organize by observation conditions rather than reusable structures. The right panels summarize the resulting sampling-rate extrapolation degradation and the improved downstream performance achieved by our pretraining method.
Figure 2: Representation-space visualization under observational heterogeneity. Left: cross-dataset heterogeneity. Right: sampling-rate heterogeneity. Statistics report cosine similarities for A: same structure, different observations; B: different structures, same observation. Modulation-colored visualizations are provided in Figure 5 .
Figure 3: Overview of structural pretraining with time alignment. I/Q patches are encoded with RoPE-fs, while masked prediction of atomic and emergent structural-unit labels provides multi-level structural supervision instead of raw waveform reconstruction.
Task
AMC
RWC
WII
RFFI
Method
RML16A
RML16B
RML16C
RML18A
Panoradio
RadChar
EI-Comm
EI-Radar
EM-AIS
ResNet
50.04
54.88
55.97
43.06
71.18
87.61
76.79
71.64
42.02
Transformer
59.27
63.10
65.41
59.45
84.93
88.77
80.44
77.05
39.12
SpectrumFM
63.72
65.35
73.37
–
–
–
–
–
–
EMind
62.51
65.45
74.34
63.83
81.97
88.37
81.70
79.19
57.07
Ours
64.07 (+1.56)
66.23 (+0.78)
79.21 (+4.87)
63.91 (+0.08)
95.13 (+13.16)
89.85 (+1.48)
82.97 (+1.27)
79.87 (+0.68)
68.90 (+11.83)
Table 1: Full-supervision fine-tuning results on downstream EM benchmarks.
Dataset
RML16A
RML16B
RML16C
RML18A
Method
50-shot
100-shot
50-shot
100-shot
50-shot
100-shot
50-shot
100-shot
ResNet
39.26
41.55
42.42
43.56
44.08
46.50
31.23
32.24
Transformer
44.08
45.70
46.19
49.22
48.41
50.43
35.04
38.59
EMind
48.38
50.10
49.08
51.20
55.01
57.09
41.97
49.13
Ours
50.76 (+2.38)
52.40 (+2.30)
50.26 (+1.18)
53.31 (+2.11)
57.27 (+2.26)
60.42 (+3.33)
45.97 (+4.00)
52.31 (+3.18)
Table 2: Few-shot fine-tuning results on AMC benchmarks.
Table 3: Ablation and sampling-rate extrapolation analyses.
Figure 4: Waveform patterns of learned structural units. (a) Representative atomic structural units (ASUs). (b) Representative emergent structural units (ESUs) and their constituent atomic sequences.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Representation-space visualizations colored by modulation class, complementing the observation-conditioned visualizations in Figure 2 .
Model
Condition
A
B
Δ=A−B
Cohen’s d
p
EMind
Cross-dataset
13.32±0.05
32.29±0.37
−18.97±0.42
−0.889±0.015
<0.001
EMind
Cross-sampling-rate
19.78±0.13
37.22±0.07
−17.44±0.06
−0.681±0.003
<0.001
SpectrumFM
Cross-dataset
84.53±0.01
85.82±0.02
−1.29±0.02
−0.219±0.002
<0.001
SpectrumFM
Cross-sampling-rate
79.58±0.02
85.61±0.02
−6.03±0.01
−0.727±0.001
<0.001
Appendix
Table 4: Cross-model cosine similarity analysis under observational heterogeneity. A denotes same-structure pairs across different observations, and B denotes different-structure pairs within the same observation. Cosine similarities are multiplied by 100, with Δ=A−B . Results are reported as mean ± standard deviation over three runs.
Model
Dataset
Modulation
Chance
EMind
99.22
62.14
33.33
SpectrumFM
79.06
57.89
33.33
Appendix
Table 5: Linear-probing accuracy (%) for dataset-source and modulation classification using frozen pretrained representations. Both tasks contain three classes.
Dataset Name
Signal Length
Sampling Rate
Used Samples
EM-Comm
1024
20 MHz
300,000
HisarMod2019.1
1024
1 MHz
300,000
RadarCommDataset
128
10 MHz
300,000
EM-Radar
1024
5 MHz
300,000
WiSig
256
25 MHz
300,000
EM-Infer-Radar-v2
2048
20 MHz
120,000
Appendix
Table 6: Pretraining subset used in this work.
Dataset
Length
Classes
SNR (dB)
Rate
RML2016.10A
128
11
−20∼18
1 MHz
RML2016.10B
128
10
−20∼18
1 MHz
RML2016.04C
128
11
−20∼18
1 MHz
RML2018.01A
1024
24
−20∼20
3.2 MHz
Panoradio HF
2048
18
−10∼25
172.8 kHz
RadChar
512
5
−20∼20
3.2 MHz
Appendix
Table 7: Fine-tuning datasets.
Ablation
Setting
RML16A
RadChar
Targets
Atomic-only
62.78
89.11
Emergent-only
62.23
89.15
Multi-level
64.07
89.85
Iterations
Iter0
62.04
89.01
Iter0 continued
62.22
89.29
Iter1 refinement
64.07
89.85
Appendix
Table 8: Additional component ablations on RML2016.10A and RadChar (accuracy, %). Iter0 continued uses an additional 100 training epochs with fixed dictionaries.
Hyperparameter
Settings
RML2016.10A
RadChar
Reference rate f0
0.5 / 1 / 2 MHz
62.01 / 64.07 / 63.51
88.82 / 89.85 / 89.76
Patch size P
4 / 8 / 16
61.91 / 64.07 / 56.56
89.57 / 89.85 / 88.60
ASU vocabulary size Ka
256 / 512 / 1024
61.98 / 64.07 / 62.20
88.95 / 89.85 / 89.11
ESU merge count Ke
256 / 512 / 1024
62.62 / 64.07 / 62.58
89.17 / 89.85 / 88.92
Maximum ESU length
3 / 5 / 8
62.93 / 64.07 / 62.81
89.69 / 89.85 / 89.68
Mask ratio
0.03 / 0.06 / 0.10
63.23 / 64.07 / 62.09
89.68 / 89.85 / 89.18
Appendix
Table 9: Hyperparameter sensitivity on RML2016.10A and RadChar (accuracy, %). Results correspond to the settings in the listed order, with the adopted configurations highlighted in bold.
Figure 6: Factor-level sampling-rate extrapolation results on RML18A.
Dataset
Standard
−5%
+5%
−10%
+10%
Missing
RML2016.10A
64.07
63.13
63.04
63.18
62.91
64.07
RadChar
89.85
89.59
89.36
89.00
88.72
89.24
Appendix
Table 10: Robustness to inaccurate and missing sampling-rate metadata (accuracy, %). Sampling-rate errors are introduced at inference time, while missing metadata is handled using f0 during both fine-tuning and inference.
Table 11: Dictionary quality analysis.
Perturbation
TMR ASU ↑
TMR ESU ↑
Cosine ASU ↑
Cosine ESU ↑
AWGN
62.9 → 64.2
67.8 → 69.9
91.8 → 91.7
93.0 → 89.6
Phase rotation
5.7 → 19.3
4.4 → 12.4
45.5 → 74.1
45.4 → 62.3
CFO
9.6 → 15.3
8.6 → 9.7
2.0 → 67.1
1.0 → 54.9
Cyclic shift
2.5 → 8.8
4.3 → 10.2
-0.8 → 63.0
-0.4 → 53.7
Overall
20.2 → 26.9
21.3 → 25.6
34.6 → 74.0
34.8 → 65.1
Appendix
Table 12: Stability of atomic and emergent structural units under signal perturbations. Each entry compares raw I/Q-based clustering with pretrained encoder-based clustering (Raw → Pretrained). TMR is reported in %, and cosine similarities are multiplied by 100. Higher values indicate greater consistency between paired views.
Large language models have expanded beyond text to vision, audio, and other modalities, but understanding raw electromagnetic (EM) signals remains challenging. Wideband recordings can contain millions of samples, whereas existing waveform-based EM models are primarily evaluated on short in-phase and quadrature (IQ) segments, leaving long-recording understanding and detailed description insufficiently explored. We introduce LEMMO, a large EM model for instruction-driven recognition and description of long raw IQ recordings. LEMMO combines a linear-attention EM encoder with anti-aliased temporal resampling to process up to eight million complex samples. Its EM-text interaction module conditions learnable queries on instructions before retrieving signal evidence. Compact queries support recognition, while structured queries with auxiliary factual supervision organize scene-level properties, signal attributes, and cross-signal relationships for detailed description. We also construct and release a unified 1.38-TB corpus of real-world and synthetic EM signals, including LEMMO-ShortQA and LEMMO-LongDesc. Compared with MERLIN-LoRA, LEMMO improves mean EM-Bench perception accuracy by 15.77 %, mean strategy-generation ROUGE-L F1 by 4.38 points, and mean LEMMO-ShortQA accuracy by 14.31 %. It also improves categorical and numerical fact F1 on both LEMMO-LongDesc tasks while preserving its language backbone's MMLU performance. Our code and datasets are available at https://github.com/N1rv4n4/LEMMO and https://huggingface.co/datasets/DaftJun/RFLAS, respectively.
Mingkai Li, Jiajun Wu, Yihan Mao +6
College of Information Science and Electronic Engineering Zhejiang University
Multimodal Large Language Models have demonstrated powerful cross-modal understanding and reasoning capabilities in general domains. However, in the electromagnetic (EM) domain, they still face challenges such as data scarcity and insufficient integration of domain knowledge. This paper proposes PReD, the first foundation model for the EM domain that covers the intelligent closed-loop of "perception, recognition, decision-making." We constructed a high-quality multitask EM dataset, PReD-1.3M, and an evaluation benchmark, PReD-Bench. The dataset encompasses multi-perspective representations such as raw time-domain waveform, frequency-domain spectrograms, and constellation diagrams, covering typical features of communication and radar signals. It supports a range of core tasks, including signal detection, modulation recognition, parameter estimation, protocol recognition, radio frequency fingerprint recognition, and anti-jamming decision-making. PReD adopts a multi-stage training strategy that unifies multiple tasks for EM signals. It achieves closed-loop optimization from end-to-end signal understanding to language-driven reasoning and decision-making, significantly enhancing EM domain expertise while maintaining general multimodal capabilities. Experimental results show that PReD achieves state-of-the-art performance on PReD-Bench constructed from both open-source and self-collected signal datasets. These results collectively validate the feasibility and potential of vision-aligned foundation models in advancing the understanding and reasoning of EM signals.
Zehua Han, Jing Xiao, Yiqi Duan +13
Beihang University · Civil Aviation University of China · Beijing Institute of Technology +10
Biosignals acquired from different locations on the body often provide temporally ordered views of the same underlying physiological process. However, most existing self supervised learning methods treat these signals as interchangeable views, overlooking the directional temporal dynamics that link them. A canonical example is the relationship between electrocardiography (ECG), which captures the electrical activation initiating each heartbeat, and photoplethysmography (PPG), which records the resulting peripheral pulse delayed by vascular dynamics. To capture this structured relationship, we introduce xMAE, a biosignal pretraining framework that leverages masked cross modal reconstruction across temporally ordered biosignals as a training time constraint to encourage physiologically meaningful timing structure in the learned representations. We show that pretraining with xMAE yields representations that outperform both unimodal and multimodal baselines on 15 of 19 downstream tasks, including cardiovascular outcome prediction, abnormal laboratory test detection, sleep staging, and demographic inference, while generalizing across devices, body locations, and acquisition settings. Further analysis suggests that the ECG PPG timing structure is reflected in the learned PPG representations. More broadly, xMAE demonstrates the effectiveness of incorporating temporal structure into multimodal pretraining when signals observe different stages of a shared underlying process. Code is available at https://github.com/hzhou3/xMAE.
Hao Zhou, Simon A. Lee, Cyrus Tanade +12
Samsung Research America · Algorithm Team, Digital Health Lab, Samsung Research America · The Pennsylvania State University +1