🔢 mathematics

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

This paper proposes FM-Receiver, a foundation model-enabled unified neural receiver that integrates inner and outer wireless processing into a single AI-native framework using a grouped error correction code Transformer and a three-stage pre-training strategy to achieve superior performance and zero-shot generalization across diverse system configurations.

Original authors: Tianyue Zheng, Chao Jiang, Linglong Dai

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC0 1.0

Original authors: Tianyue Zheng, Chao Jiang, Linglong Dai

Original paper dedicated to the public domain under CC0 1.0 (http://creativecommons.org/publicdomain/zero/1.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to listen to a friend shouting a secret message across a noisy, crowded stadium. In the old days, your brain (the receiver) would have to do this in two separate steps: first, you'd squint and filter out the crowd noise to hear the words clearly (the "outer" receiver), and then, once you had the words, you'd use a separate rulebook to fix any typos or missing letters (the "inner" receiver). The problem? These two steps didn't talk to each other. If the first step made a tiny mistake, the second step had no way to say, "Hey, I think that word was actually supposed to be different!"

This paper introduces a new kind of "super-listener" called FM-Receiver. Think of it as a single, super-smart AI brain that does both jobs at the same time. Instead of splitting the work, it learns to hear the noise and fix the message in one giant, unified leap. The authors call this a "Foundation Model," which is like a student who has read every book in the library before ever taking a test, so they can handle almost any question thrown at them.

The Big Problem: The "Bit" vs. "Symbol" Mismatch

The paper points out a specific glitch in how previous AI receivers worked. Imagine the outer receiver speaks in "words" (symbols, which are groups of bits), but the old AI decoders only understood "letters" (individual bits). It's like trying to translate a whole sentence by looking at one letter at a time; it's slow and confusing.

The authors explicitly rule out the idea of just sticking a standard AI on the outside and a traditional rulebook on the inside. They argue that keeping them separate prevents the system from learning how to work together efficiently. They also argue against the idea that AI decoders should only work on perfect, clean data; in the real world, the data coming from the first step is messy, and the decoder needs to handle that mess.

The Solution: The "Grouped" Super-Decoder

To fix the mismatch, the team invented a new tool called G-ECCT (Grouped Error Correction Code Transformer).

The Analogy: Instead of reading one letter at a time, this new decoder grabs a whole "word" (a symbol made of multiple bits) and treats it as a single unit. It's like reading a whole sentence at once rather than stumbling over every single letter.
The Result: This makes the process much faster and allows the "hearing" part and the "fixing" part to blend together seamlessly.

How They Taught It: The Three-Stage Bootcamp

You can't just throw a complex AI into the stadium and expect it to work. The authors designed a three-stage training strategy to make sure the AI is ready for anything:

Stage 1 (The Ear): They taught the AI how to filter out noise and hear the signal clearly, using a mix of different crowd sizes and stadium layouts.
Stage 2 (The Brain): They taught the new "Grouped" decoder how to fix errors, but they started with a clean, quiet room (simulated perfect data) so it could learn the rules of grammar without getting confused by noise yet.
Stage 3 (The Team): Finally, they put the "Ear" and the "Brain" back together and trained them as a team, letting them learn from each other's mistakes in messy, noisy conditions.

What the Simulations Showed

The authors ran thousands of simulations (computer tests) to see how this new receiver performed. They didn't just test it in one scenario; they threw everything at it.

Crowd Size: Whether there was 1 user or 4 users shouting at once, the FM-Receiver was the best at keeping the message clear. In the toughest 4-user scenario, it beat the next-best method by more than 2 dB (a measure of signal quality).
Different Languages: They tested it with different "languages" (modulation schemes like QPSK, 16-QAM, and 64-QAM). It won in all of them, though the gap was biggest with the simpler languages.
The "Zero-Shot" Magic: This is the coolest part. They trained the AI on data from 3.5 GHz, 4.9 GHz, and 7.0 GHz. Then, they tested it on a brand new frequency, 28 GHz, which it had never seen before. Without any extra training, it still worked better than the old methods. It also handled a completely new type of "stadium" (channel model) it had never seen, while other AI methods failed and fell back to the performance of basic, old-school receivers.

Speed and Size

The paper notes that while this new receiver is a "big" model (with about 38.98 million parameters, compared to smaller models with fewer), it is surprisingly fast. In their tests, it processed data in just 5.511 ms per batch. This is more than twice as fast as the previous best AI method and nearly 9 times faster than the traditional method that uses a slow, step-by-step error-checking loop.

The Bottom Line

The authors suggest that this unified approach is a major step toward "AI-native" wireless systems, where the whole receiver is built from the ground up to be smart. They don't claim it's a solved problem for the real world yet, but their simulations show it has the potential to be much more reliable and efficient than what we have today, especially when things get messy or change unexpectedly. They also hint that future work might need to shrink the model down so it can fit on smaller, cheaper hardware, but for now, the big, smart version looks very promising.

Technical Summary: FM-Receiver

Problem Statement

Current neural receiver architectures in wireless communications typically apply deep learning only to the "outer" receiver (channel estimation, equalization, and demapping) while retaining conventional algorithms, such as Belief Propagation (BP), for the "inner" receiver (channel decoding). This separation creates two critical limitations:

Lack of Joint Optimization: The outer and inner receivers are optimized independently, preventing the system from learning the complex, non-linear relationships between signal processing and decoding, which leads to performance degradation, particularly in challenging multi-user or low signal-to-noise ratio (SNR) scenarios.
Architectural Mismatch: Existing AI-based channel decoders often operate at the bit level, whereas the outer receiver processes modulated symbols (groups of bits). This mismatch hinders the seamless integration of the two stages into a single, unified neural framework and increases computational complexity.

Furthermore, existing neural receivers often struggle to generalize across diverse system configurations (e.g., varying numbers of users, frequency bands, and modulation schemes) without extensive retraining, limiting their applicability to AI-native air interfaces.

Methodology

The authors propose FM-Receiver, a Foundation Model (FM)-enabled unified neural receiver that integrates the outer and inner receivers into a single end-to-end framework. The core methodology consists of three main components:

1. Grouped Error Correction Code Transformer (G-ECCT)

To resolve the mismatch between symbol-level outer processing and bit-level decoding, the authors introduce G-ECCT.

Symbol-Level Decoding: Instead of treating individual bits as tokens, G-ECCT groups the $m$ bits of a modulated symbol into a single token. This aligns the decoding granularity with the outer receiver's output.
Masked Attention: A specialized group-based mask generation scheme is designed to incorporate the parity-check matrix structure into the self-attention mechanism. This allows the Transformer to learn dependencies among symbol groups while respecting the code constraints.
Efficiency: By processing symbols rather than individual bits, the computational complexity is reduced by a factor of $m^2$ compared to bit-level approaches.

2. Unified FM-Receiver Architecture

The overall architecture takes the received time-frequency resource grid (post-FFT) as input and directly outputs recovered transmitted bits.

Backbone: The network utilizes a stacked Transformer backbone. The first $N_1$ blocks employ full self-attention to perform outer receiver tasks (channel estimation, equalization, and demapping) to generate Log-Likelihood Ratios (LLRs).
Decoding: The subsequent $N_2$ blocks utilize masked self-attention (G-ECCT) to perform symbol-level channel decoding on the generated LLRs.
Auxiliary Supervision: A read-out module provides refined channel estimates during training to improve convergence and stability.

3. Three-Stage Configuration-Adaptive Pre-Training

To ensure robust generalization across diverse scenarios, a three-stage pre-training strategy is employed:

Stage 1 (Outer Receiver): Pre-trains the outer receiver blocks on diverse datasets to learn robust frequency-domain processing and interference suppression.
Stage 2 (G-ECCT): Pre-trains the G-ECCT blocks using clean LLRs generated from an Additive White Gaussian Noise (AWGN) channel. This allows the decoder to converge rapidly without overfitting to the imperfect LLR distributions of the outer receiver.
Stage 3 (Joint Optimization): Fine-tunes the entire unified network end-to-end using diverse datasets, allowing the outer and inner components to adapt to each other.

Key Contributions

Unified Framework: The paper presents the first architecture to fully unify the outer and inner receivers using Foundation Models, eliminating the need for conventional channel decoding algorithms.
G-ECCT Design: The introduction of symbol-level decoding via G-ECCT bridges the gap between modulation and coding, enabling seamless integration and reducing computational complexity.
Generalization Strategy: The proposed three-stage pre-training strategy enables the model to adapt to unseen system configurations, including varying numbers of users, frequency bands, and Modulation and Coding Schemes (MCSs).
Zero-Shot Capability: The model demonstrates the ability to generalize to unseen frequency bands (e.g., 28 GHz) and channel models without fine-tuning.

Simulation Results

Extensive simulations were conducted using 3GPP-compliant link-level simulators (Sionna) across 15 datasets covering various channel models (CDL-A to E), frequency bands (3.5 GHz to 28 GHz), and user configurations.

Performance Gains: FM-Receiver consistently outperformed baselines, including conventional receivers (LS/LMMSE + BP) and learning-based methods (CNN+BP, Transformer+BP). In a challenging 4-user scenario, the proposed method achieved a performance gain of over 2 dB compared to baselines.
Modulation Robustness: The method showed superior performance across QPSK, 16-QAM, and 64-QAM. For QPSK, gains exceeded 3 dB over traditional baselines; for 64-QAM, gains remained above 1.5 dB.
Zero-Shot Generalization:
- Frequency Bands: When tested on an unseen 28 GHz band (trained on 3.5–7.0 GHz), FM-Receiver maintained strong performance, outperforming the "Transformer+BP" baseline by approximately 1 dB more than in-band results.
- Channel Models: When tested on an unseen channel model (CDL-B), FM-Receiver maintained a ~1 dB gain over the best conventional baseline, whereas other neural receivers degraded significantly.
Scalability: Performance improved with larger model sizes (up to 86.66M parameters) and more diverse pre-training datasets.
Computational Efficiency: Despite having more parameters (38.98M) than some baselines, FM-Receiver achieved the lowest inference latency (5.511 ms per batch) compared to iterative BP-based methods (e.g., 49.980 ms for LMMSE+LMMSE+BP), due to its single-pass feed-forward architecture.

Significance and Claims

The paper claims that FM-Receiver represents a significant step toward realizing AI-native wireless communications. By unifying the entire receiver chain into a single foundation model, the approach overcomes the limitations of modular, separately optimized designs. The authors assert that this architecture not only delivers superior bit error rate (BER) performance across diverse and challenging scenarios but also offers strong zero-shot generalization capabilities essential for future 6G networks where system configurations may vary dynamically. The reduction in inference latency further suggests the practical viability of deploying such complex neural receivers in real-time systems, moving beyond the theoretical benefits of AI to practical implementation.