Rich-ARQ: From 1-bit Acknowledgment to Rich Neural Coded Feedback

This paper reimagines the foundational feedback mechanism in wireless communication, transforming the prevailing 1-bit binary ACK/NACK with a high-dimensional, information-rich vector to transform passive acknowledgment into an active collaboration. We present Rich-ARQ, a paradigm that introduces neural-coded feedback for collaborative physical-layer channel coding between transmitter and receiver. To realize this vision in practice, we develop a novel asynchronous feedback code that eliminates stalling from feedback delays, adapts dynamically to channel fluctuations, and features a lightweight encoder suitable for on-device deployment. We materialize this concept into the first full-stack, standard-compliant software-defined radio prototype, which decouples AI inference from strict radio timing. Comprehensive over-the-air experiments demonstrate that Rich-ARQ achieves significant SNR gains over conventional 1-bit hybrid ARQ and remarkable latency reduction over prior learning-based feedback codes, moving the promise of intelligent feedback from theory to a practical, high-performance reality for next-generation networks.

💡 Research Summary

The paper introduces Rich‑ARQ, a novel wireless communication paradigm that replaces the conventional 1‑bit ACK/NACK feedback with a high‑dimensional, neural‑coded feedback vector, turning feedback from a passive binary indicator into an active collaborator in physical‑layer coding. In the Rich‑ARQ framework, the access point (AP) processes the received forward signal to extract a compact feature vector, encodes it into a feedback packet, and sends it back to the user equipment (UE). The UE’s encoder then consumes this rich feedback together with previously transmitted symbols to generate the next forward transmission, establishing a closed‑loop, collaborative coding cycle.

To make this vision practical, the authors address three major obstacles that have limited prior feedback‑coding schemes: (1) feedback latency that stalls the forward encoder, (2) sensitivity to time‑varying signal‑to‑noise ratio (SNR), and (3) the high computational burden of deep‑learning (DL) encoders on resource‑constrained devices. Their solution is the Asynchronous Feedback Code (AFC). AFC decouples forward transmission from the arrival of the latest feedback by allowing the encoder to use historical feedback information, thereby eliminating stalls and creating an overlapping pipeline that dramatically reduces end‑to‑end latency.

A second contribution is an SNR‑conditioned curriculum learning strategy augmented with Langevin perturbations. Training starts at high SNR and gradually introduces lower‑SNR conditions while adding stochastic noise, enabling the model to learn robust encoding and decoding policies across a wide SNR range that mirrors real indoor measurements (2.1 dB variation over 100 ms up to 14.3 dB over 2.7 s).

The third contribution is a lightweight AFC encoder designed for UE deployment. Through model pruning and sparse matrix operations, the encoder’s parameter count and FLOPs are drastically reduced, while the decoder—running on the AP, which has abundant power and compute resources—remains a full‑featured attention‑based DNN. This asymmetric design aligns with the star topology of modern cellular and Wi‑Fi networks, where the base station is resource‑rich and the devices are battery‑limited.

The authors built the first full‑stack, standard‑compliant software‑defined radio (SDR) prototype that integrates AFC with a 4G/5G physical layer. A deadline‑aware execution architecture separates AI inference from strict radio timing, ensuring that variable‑latency DNN processing does not violate air‑interface deadlines. Over‑the‑air experiments demonstrate that Rich‑ARQ achieves 8.8–9.5 dB SNR reduction compared with Turbo‑HARQ and Polar‑HARQ to reach a packet error rate of 10⁻⁴, corresponding to a 1.38× and 1.70× increase in maximum communication distance, respectively. Moreover, Rich‑ARQ maintains stable performance under realistic SNR fluctuations where prior DL‑based feedback codes degrade sharply, and it reduces end‑to‑end latency by 43.4 % relative to the state‑of‑the‑art DL feedback scheme.

In summary, Rich‑ARQ demonstrates that enriching feedback with neural‑coded vectors can unlock substantial gains in reliability, spectral efficiency, and latency, moving feedback coding from theoretical constructs to a deployable technology for next‑generation ultra‑reliable wireless systems. Future work may explore multi‑user scheduling, MIMO feedback channels, and integration into standardization bodies.