ARTFEED — Contemporary Art Intelligence

Unified Backbone-Expert Framework with Relation-Token and Residual-Classifier for Automatic Modulation Recognition

other · 2026-08-18

A new unified backbone-expert framework for automatic modulation recognition (AMR) has been developed by researchers to overcome performance limitations associated with different observation lengths. This framework combines a standard convolutional state-space backbone with two tailored interfaces: relation tokens designed for short sequences and a gated multi-scale residual refinement module for longer ones. The relation tokens incorporate explicit lag-aware complex-plane descriptors prior to encoding, while the residual module enhances feature maps and utilizes fixed-averaging classifier collaboration. On RML2016.10b, the framework achieves average accuracies of 67.28% ± 0.14%, and 87.19% ± 0.77% on HisarMod2019, confirmed through three-seed ablations and native-length cross-configuration tests. This research is detailed in arXiv paper 2608.15160, which was announced as a cross submission.

Key facts

  • Framework uses a common convolutional state-space backbone.
  • Relation tokens are used for short sequences to compensate for information loss.
  • Gated multi-scale residual refinement module corrects feature maps for long sequences.
  • Fixed-averaging classifier collaboration harnesses complementary evidence.
  • Achieves 67.28% ± 0.14% accuracy on RML2016.10b.
  • Achieves 87.19% ± 0.77% accuracy on HisarMod2019.
  • Validated through three-seed ablations and native-length cross-configuration tests.
  • Paper available on arXiv with ID 2608.15160.

Entities

Institutions

  • arXiv

Sources