FM-Receiver: A Foundation Model Enabled Unified Inner and Outer Neural Receiver Towards AI-Native Wireless Communications

2026-07-14

AI Summary

This work addresses the limitation of existing neural receivers, which typically confine deep learning to the outer receiver while relying on conventional decoding methods for the inner receiver, thereby hindering end-to-end joint optimization. To overcome this, the authors propose FM-Receiver, the first framework to integrate foundation models into wireless receiver design, establishing a unified AI-native architecture that directly recovers transmitted bits from received signals. Built upon the Transformer architecture, FM-Receiver incorporates structured error-correcting code mechanisms and employs a novel three-stage pretraining strategy with configuration-aware adaptation, enabling integrated inner–outer receiver operation and symbol-level end-to-end channel decoding. Experimental results demonstrate that FM-Receiver outperforms current baselines across diverse system configurations and exhibits strong zero-shot generalization to unseen frequency bands and deployment scenarios.

Abstract

With the development of artificial intelligence (AI) techniques, neural receivers, which apply AI to improve wireless receivers have been developed. However, most existing neural receivers apply deep learning only to the outer receiver while retaining conventional channel decoding for the inner receiver, which prevents joint optimization and makes it difficult to build efficient and unified AI-native receivers. To address this issue, we propose a foundation model (FM)-enabled unified neural receiver, FM-Receiver, that integrates the outer and inner receivers into a single AI-native framework, by leveraging the strong representation capability of FMs. Specifically, we introduce a grouped error correction code Transformer that performs symbol-level channel decoding, enabling seamless integration of the inner and outer receiver. Building on this, we illustrate the proposed FM-Receiver, that directly takes the received signals as input of FM and outputs the recovered transmitted bits. In addition, a three-stage configuration-adaptive pre-training strategy is designed to improve the generalization ability to diverse system configurations and scenarios. Extensive simulations show that the proposed FM-Receiver achieves better performance than baselines across different system configurations. It also demonstrates strong zero-shot generalization to unseen frequency bands and scenarios.

Problem

• neural receiver
• AI-native wireless communications
• inner and outer receiver
• channel decoding
• foundation model

Innovation

• Foundation Model
• Neural Receiver
• Unified Inner-Outer Receiver
• Error Correction Code Transformer
• Zero-shot Generalization