Generative Communications: A Controllable Generation Paradigm for 6G

Definition and Core Properties

Generative Communications (GenCom) redefines the communication objective: the transmitter sends only minimal semantic cues, control signals, or latent representations, while the receiver leverages shared large‑language, diffusion, or multimodal models together with a knowledge base to generate task‑aligned outputs. Two essential properties characterize GenCom:

Native intelligence : semantic understanding, reasoning, and generation are integral to the communication link.

Generation‑driven quality : success is measured by the receiver’s ability to produce correct, controllable content from the sparse cues.

Core Mechanisms

Transmission Information – sparse cues such as text prompts, segmentation maps, latent codes, low‑resolution observations, or control vectors that trigger and constrain generation.

Generative Model Priors – pre‑trained large models that embed structured knowledge and latent spaces, enabling the receiver to infer and synthesize content without transmitting the full data.

Knowledge Base – shared factual repositories that provide grounding, consistency, and version control; synchronization mechanisms prevent drift between sender and receiver.

Advantages over Traditional Communication

GenCom shifts the goal from accurate bit‑level delivery to task‑level semantic quality. By transmitting far fewer bits, it maintains semantic relevance under high noise, low bandwidth, or constrained devices. The receiver’s generative capability supports cross‑modal reconstruction, intent‑driven transmission, adaptive resource scheduling, and task‑level quality control. This advantage requires alignment of models, knowledge, and evaluation metrics across sender, receiver, and control layers.

6G GenCom Architecture

The architecture consists of two layers:

Transmission Layer – handles end‑to‑end signal expression, wireless channel transmission, on‑device inference, and generation. The sender extracts task‑relevant representations, encodes them into channel symbols, and transmits them; the receiver decodes the symbols, aligns them with conditional embeddings, and invokes large‑model inference to synthesize the final output.

Control Layer – sits above the transmission layer and coordinates global intelligence tasks such as self‑learning, network resource orchestration, knowledge synchronization, and generative resource management. Control agents collect data and rewards, fine‑tune models, manage knowledge versions, and adapt inference strategies based on network state, user intent, and model load.

Key Enabling Technologies

Joint source‑channel‑generative coding that integrates the generative model’s latent space into the coding objective.

Controlled generation via conditional prompts, latent control codes, and constraint decoding to ensure outputs respect sender intent and task constraints.

Communication‑aware large models that understand channel state, latency, and resource budgets; these models can be compressed, pruned, quantized, or distilled for edge deployment.

Knowledge‑driven generation and synchronization (incremental model broadcast, knowledge‑graph updates, version tracking, provenance verification) to keep distributed nodes semantically consistent without massive data transfer.

New evaluation metrics beyond traditional QoS: semantic similarity, perceptual quality, factual consistency, control success rate, task completion, and generation stability.

Comparison with Related Paradigms

GenCom overlaps with semantic communication, knowledge‑assisted communication, and model‑enhanced communication but differs in focus. Semantic communication extracts task‑relevant semantics and may use a generative module as post‑processing; GenCom jointly optimizes encoding, transmission, and generation, making the receiver’s generative ability a core component. Knowledge‑assisted approaches reduce payload via shared facts; GenCom further incorporates model priors and controlled generation into the system design.

Typical Applications

Extended reality content delivery : transmit sparse viewpoints, scene descriptions, or mesh structures; the receiver synthesizes high‑fidelity immersive experiences, adapting quality to user view, device capability, and network state.

Multi‑UAV cooperative communication : UAVs share semantic summaries or target cues; other nodes reconstruct full scenes using shared maps, task priors, and generative models, enabling efficient collaborative perception and planning.

Base‑station resource allocation : intelligent agents infer user intent, channel conditions, and resource constraints to dynamically allocate spectrum, beamforming, power, model deployment, and compute scheduling.

Semantic‑fidelity scalable transmission : transmission granularity adapts to bandwidth and device capability. Text + low‑resolution images achieve high semantic similarity with reduced symbol overhead; pure text further compresses transmission at the cost of visual fidelity.

Future Research Directions

Unified theoretical framework : quantify the minimal transmission needed given model priors, knowledge, and task constraints; define task‑level capacity and generation error bounds.

Real‑time processing : design low‑latency generative models, edge‑side lightweight inference, caching, progressive generation, and joint transmission‑computation scheduling to meet stringent latency requirements of XR, vehicular, and UAV control.

Edge‑collaborative inference : investigate model partitioning, intermediate representation sharing, and knowledge synchronization across distributed edge nodes under privacy, bandwidth, and latency constraints.

Security and resilience : develop robust detection of manipulated prompts, poisoned knowledge bases, and malicious latent codes; implement semantic consistency verification, trustworthy provenance, and safe‑fallback mechanisms.

Conclusion

GenCom represents a paradigm shift where communication systems aim to control generation rather than merely copy data. By transmitting minimal semantic cues and leveraging shared generative priors and knowledge, 6G networks can achieve ultra‑efficient transmission, semantic‑level robustness, and new intelligent networking functions, provided that open challenges in theory, latency, edge collaboration, knowledge synchronization, metrics, and security are addressed.

Code example

来源：专知
本文
约4500字
，建议阅读
8
分钟
本文提出的生成式通信代表了一种重要方向：通信系统的目标从“复制数据”转向“控制生成”。