Inference-Driven Uplink for 6G: Architecture, Principles, and Challenges

Next-generation wireless networks (6G) face a critical uplink challenge arising from stringent device-side resource constraints and the growing demand for intelligence services. This article introduces InferCom, an inference-driven communication architecture designed to enable robust 6G uplink transmission under low signal-to-noise (SNR) conditions. InferCom adopts a compute-asymmetric architecture, featuring a lightweight transmitter and an inference-capable receiver empowered by generative artificial intelligence (GenAI) models, together with a quality-of-experience (QoE)-aware retransmission mechanism. Grounded in the information bottleneck (IB) theory, InferCom redefines uplink communications through task-agnostic compression, inference-driven reconstruction, error-distribution channel coding, and QoE-aware feedback. The case study demonstrates that InferCom outperforms conventional 5G NR and Deep- JSCC in terms of transmitter-side computational complexity, required SNRs and retransmission efficiency. Finally, we outline key challenges and research directions toward making InferCom a practical enabler of human-centric, intelligent and sustainable wireless networks.

💡 Research Summary

This paper proposes “InferCom,” a novel inference-driven communication architecture designed to overcome the critical uplink bottleneck anticipated in 6G networks. The bottleneck arises from the conflict between the stringent computational, energy, and bandwidth constraints of edge devices and the growing demand for high-data-rate, intelligent services like immersive XR and digital twins.

InferCom’s core innovation is its compute-asymmetric architecture, which fundamentally shifts the processing burden from the resource-constrained transmitter to the resource-abundant receiver. The transmitter is designed to be lightweight and inference-oriented. Instead of running complex neural networks for semantic extraction, it applies simple, task-agnostic compression (e.g., mean filtering, downsampling) to create a coarse representation that preserves structural and statistical cues essential for inference. The receiver, empowered by a large-scale Generative AI (GenAI) model, acts as a powerful semantic inference engine. Leveraging its strong generative prior learned from vast datasets, it can reconstruct task-relevant, high-quality outputs from the degraded and noisy received signals, even under low Signal-to-Noise Ratio conditions where conventional systems fail.

The architecture integrates three key components: the lightweight transmitter, the GenAI-powered receiver, and a QoE-aware retransmission mechanism. This feedback mechanism is a significant departure from conventional CRC-based protocols. Retransmissions are triggered only when the reconstructed output fails to meet a semantic adequacy or Quality of Experience metric, avoiding unnecessary delays and overhead when bit-level errors are tolerable for the end-task.

The design principles of InferCom are grounded in an extended interpretation of the Information Bottleneck theory. It introduces four key shifts: 1) Simple, task-agnostic compression at the transmitter, 2) Inference-driven reconstruction at the receiver, 3) Error-distributing channel coding that shapes errors into a form more manageable by the GenAI model, rather than eliminating them entirely, and 4) QoE-aware retransmissions that align feedback with semantic relevance.

A case study using image transmission demonstrates InferCom’s advantages. Under block Rayleigh fading channels, it significantly outperforms conventional 5G NR (using JPEG and LDPC) and Deep-JSCC baselines. The metrics show superior performance in transmitter-side computational complexity, the required SNR to achieve a target QoE (based on perceptual quality and semantic similarity), and retransmission efficiency. Finally, the paper outlines practical challenges for deployment, including the real-time latency of large GenAI models, defining adaptive QoE metrics, and supporting diverse data modalities, pointing toward future research directions for making InferCom a practical enabler of human-centric and intelligent 6G networks.