Organizations: Institute of Cyberspace Security, College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023, China, also with the Binjiang Institute of Artificial Intelligence, Zhejiang University of Technology, Hangzhou 310056, China. · Institute of Cyberspace Security, Zhejiang University of Technology, Hangzhou 310023, China, and also with the Research Center of Electromagnetic Space Security, Binjiang Institute of Artificial Intelligence, ZJUT, Hangzhou, 310056, China. · College of Information and Communication Engineering, Harbin Engineering University, Harbin 150001, China. · College of Telecommunications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China.
Deep neural networks for electromagnetic signal processing are typically developed as single-task experts with fixed inputs and objectives, limiting cross-task transferability. Electromagnetic foundation models offer broader task coverage but incur substantial pretraining and inference costs. To balance accuracy and efficiency, we adopt model routing to dynamically dispatch unknown signals to matched experts within a heterogeneous pool. We construct HeteroSignal-7, a multi-task routing benchmark with 140,000 samples from seven datasets covering six task types, paired with seven classification and waveform-recovery experts and a unified evaluation protocol. We then propose GLCR, a two-stage routing method combining global dispatch with local discrimination of similar tasks. First, a lightweight one-dimensional convolutional neural network (CNN) with an additive-margin head predicts one of seven routing classes. Second, a gradient-boosted-tree specialist uses four physical statistics to re-evaluate samples whose top two candidates match a confusion pair automatically identified from validation errors. The pipeline is trained across the full signal-to-noise ratio (SNR) range and requires neither SNR labels nor noise estimation at inference time. GLCR outperforms all 19 baselines, achieving 96.16% routing accuracy and a 76.30% system score, exceeding the respective second-best baselines by 1.16% and 1.14%. It achieves the lowest routing latency among the compared methods at 1.43 ms per sample and reduces expected overall system FLOPs by 5.9 times compared with executing all experts. This work provides a public evaluation platform and an efficient, reproducible baseline for heterogeneous expert routing in electromagnetic signal processing.
Figures & tables
Fig. 1: Number of publications on machine-learning methods for communications from 2016 to 2025, retrieved from OpenAlex [ 1 ] .
Fig. 2: Three paradigms for multi-task electromagnetic signal processing: (a) transfer learning with a frozen source-task backbone and task-specific heads, (b) an electromagnetic foundation model pretrained on multi-task data and adapted through task-specific heads, and (c) model routing that selects task-specific experts from a heterogeneous pool.
Fig. 3: Cross-task transfer with frozen backbones and retrained probe heads. (a) Cross-dataset transfer of three AMC models trained on RML2016.10a. (b) Head-swapped cross-evaluation of five classification experts. Red marks indicate dataset-matched entries, which differ from the full-expert results in Table II . Waveform-output tasks are excluded from (b).
Fig. 4: Example samples of the seven datasets in the unified 1,024-point router input format (amplitude-normalized I/Q). Zero-padded datasets (RML2016.10a, RadChar) are shown over their non-zero segments; native lengths are annotated in the panel titles, and FS-SEI is center-cropped from its native 4,800-point sequence.
Dataset
Task
Signal Type
Signal Length
Sampling Rate
Data Source
Number of Samples
RML2016.10a [ 38 ]
AMC
11 modulation classes, SNR: −20 – 18 dB
128
1 MHz
Simulation
20,000
RML2018.01a [ 2 ]
AMC
24 modulation classes, SNR: −20 – 30 dB
1,024
1 MHz
Simulation
20,000
SD-BENCH3 [ 2 ]
SD
24 modulation classes with AWGN, SNR: −3 – 20 dB
1,024
1 MHz
Simulation
20,000
EM-Infer-Comm [ 9 ]
WII
9 interference types, SNR: −20 – 20 dB
1,024
20 MHz
Simulation
20,000
FS-SEI [ 39 ]
RFFI
90 ADS-B emitters, SNR: not annotated
4,800
50 MHz
Real-world
20,000
RadChar [ 40 ]
RWC
5 radar waveform classes, SNR: −20 – 20 dB
512
3.2 MHz
Simulation
20,000
TABLE I: The HeteroSignal-7 Benchmark. Bold entries are constructed in this work using source data or generation protocols from the cited references.
Task
Dataset
Expert Model
Venue
Params (M)
FLOPs (M)
Test Metric
Test Score
AMC
RML2016.10a
IQFormer [ 22 ]
IEEE TCCN 2025
0.36
42
OA
65.15%
RML2018.01a
AWN [ 23 ]
IEEE TCCN 2023
0.38
70
OA
61.96%
SD
SD-BENCH3
Conv-TasNet [ 27 ]
IEEE/ACM TASLP 2019
2.18
293
SNR gain
+14.53 dB
WII
EM-Infer-Comm
FOSR [ 4 ]
IEEE TCOM 2026
0.82
89
OA
79.35%
RFFI
FS-SEI
CV-TSLANet [ 5 ]
IEEE IoT-J 2025
6.46
889
OA
86.75%
RWC
RadChar
JDMRNet-MTL [ 3 ]
IEEE TAES 2023
1.92
54
OA
89.45%
TABLE II: Expert Model Pool of HeteroSignal-7. All experts are trained from scratch.
Fig. 5: Overview of the proposed GLCR framework: (a) unified preprocessing of HeteroSignal-7 samples; (b) Stage-1 global decision using a lightweight CNN with an additive-margin head; (c) top-2 confusion-pair triggering and Stage-2 local re-evaluation using four physical statistics (DSP4) and a binary HGBT specialist; and (d) dispatching to the selected frozen expert.
Fig. 6: Motivation for local re-evaluation. (a) Spectrograms of an AMC input from RML2018.01a and a denoising input from SD-BENCH3 exhibit similar time-frequency structures. (b) Distribution of validation routing errors over unordered task pairs.
Fig. 7: Means of the four DSP4 statistics on the training splits, stratified by class and SNR tier over the 0–10 dB confusion region. The tiers follow the native SNR grids of the two datasets (RML2018.01a: even-numbered tiers; SD-BENCH3: 0/3/6/9 dB). The class means differ at every shown tier; the dashed curve in (a) is the theoretical attenuation law γ/(1+γ) normalized to the SD mean at 0 dB.
Category
Method
Routing Acc. (%)
System Score (%)
Signal Models
AMC-Net
95.01
74.89
ResNet
94.83
75.16
PET-CGDNN
92.14
73.83
Feature-based Classifiers
All-layer concatenation + MLP
94.19
74.86
se_out + MLP
94.03
74.73
Routing Methods
AWN + CSCR
93.83
74.67
TABLE III: Routing Accuracy and System Score on the Cleaned HeteroSignal-7 Test Split (13,975 Samples; Seed 0). Oracle is the upper bound and is excluded from comparison. Best and second-best results are in bold and underlined , respectively.
Fig. 8: System score on the cleaned test split versus routing-pipeline CPU latency per sample for the compared routers; a representative subset of the nineteen baselines is shown for clarity.
Fig. 9: I/Q waveforms and routing decisions of the compared methods for two case-study samples.
Stage-1 classifier
Stage-2 specialist
Results
Configuration
CNN
Margin
Input
Classifier
Acc. (%)
Sys. (%)
Δ (%)
Panel A: construction path
MLP on frozen AWN features ( se_out )
94.03
74.73
–
End-to-end CNN with a plain linear head
✓
94.74
75.22
+0.71
+ feature & weight normalization ( m=0 )
✓
94.54
75.17
−0.20
+ additive margin m=0.3 (stage-1 classifier)
✓
✓
94.91
75.34
+0.37
TABLE IV: Ablation study of GLCR on HeteroSignal-7 (✓: enabled). Δ denotes routing accuracy difference relative to the preceding row in Panel A, the full configuration in Panel B, and HGBT in Panel C.
Adaptive model routing selects models or specialized experts per input to balance prediction quality, computational cost, latency, and resource constraints.
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
Sparse expert models can distribute traffic evenly while still grouping incompatible training signals within the same experts. We study routing as a gradient-partitioning problem and introduce gradient-aligned routing (GAR), whose load-normalized router objective rewards grouping observations with aligned gradients. On five multi-task text-classification mixtures, we compare GAR with task-loss-only routing, gradient-combination and gradient-conflict methods, and load-balancing losses. With a fully trainable RoBERTa backbone and classification-head experts, GAR has the highest aggregate validation accuracy, 1.07 percentage points above task-loss-only routing. With frozen DeBERTa and Qwen3-1.7B backbones and low-rank adapter experts, it again ranks first, 1.10 points above task-loss-only routing, with better-balanced expert load and higher gradient-mass purity, the share of each expert's gradient-norm mass from its dominant task; the load-balancing losses flatten load further but leave this purity near its task-loss-only level. Top-1 routing, trainable full-parameter feed-forward network (FFN) experts, and a larger backbone also show positive aggregate gains. The results distinguish expert-load balance from gradient-based routing organization and indicate the predictive value of gradient-informed routing in multi-task text classification.
Yuchen Li, Mingyu Du, Zongqi Fan +2
University of Sydney · University of New South Wales
The paradigm of Multimodal Large Language Models (MLLMs) offers a promising blueprint for advancing the electromagnetic (EM) domain. However, prevailing approaches often deviate from the native MLLM paradigm, instead using task-specific or pipelined architectures that lead to fundamental limitations in model performance and generalization. Fully realizing the MLLM potential in EM domain requires overcoming three main challenges: (1) Data. The scarcity of high-quality datasets with paired EM signals and descriptive text annotations used for MLLMs pre-training; (2) Benchmark. The absence of comprehensive benchmarks to systematically evaluate and compare the performance of models on EM signal-to-text tasks; (3) Model. A critical fragility in low Signal-to-Noise Ratio (SNR) environments, where critical signal features can be obscured, leading to significant performance degradation. To address these challenges, we introduce a tripartite contribution to establish a foundation for MLLMs in the EM domain. First, to overcome data scarcity, we construct and release EM-100k, a large-scale dataset comprising over 100,000 EM signal-text pairs. Second, to enable rigorous and standardized evaluation, we propose EM-Bench, the most comprehensive benchmark featuring diverse downstream tasks spanning from perception to reasoning. Finally, to tackle the core modeling challenge, we present MERLIN, a novel training framework designed not only to align low-level signal representations with high-level semantic text, but also to explicitly enhance model robustness and performance in challenging low-SNR environments. Comprehensive experiments validate our method, showing that MERLIN is state-of-the-art in the EM-Bench and exhibits remarkable robustness in low-SNR settings.
Junyu Shen, Zhendong She, Chenghanyu Zhang +13
Tsinghua University · Tianjin University · Beijing University of Posts and Telecommunications +6