From Semantic to Token Communication: The Next Paradigm for Large-Model-Driven 6G Intelligent Connectivity

Yu Ma; Zhen Gao; Li Qiao; Xiaoyuan Zhang; Mahdi Boloursaz Mashhadi; Yin Xu; Wenjun Xu; Xiaodong Xu

eess.SPSep 9, 2026

From Semantic to Token Communication: The Next Paradigm for Large-Model-Driven 6G Intelligent Connectivity

Authors: Yu Ma, Zhen Gao, Li Qiao, Xiaoyuan Zhang, Mahdi Boloursaz Mashhadi, Yin Xu, Wenjun Xu, Xiaodong Xu,

Organizations: Beijing Institute of Technology, Beijing 100081, China · Zhongguancun Academy, Beijing 100094, China · Department of Electrical and Computer Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong · 5G/6G Innovation Centre (5G/6GIC), Institute for Communication Systems, University of Surrey, Guildford GU2 7XH, UK · Cooperative Medianet Innovation Center, Shanghai Jiao Tong University, Shanghai 200240, China · State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China · School of Information Science and Engineering, Southeast University, Nanjing 210096, China · School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, UK · Faculty of Information Science and Technology, Ocean University of China, Qingdao 266100, China · Information Systems Technology and Design Pillar, Singapore University of Technology and Design, Singapore 487372, Singapore

Abstract

The ambitious requirements of sixth-generation (6G) networks are driving communication systems from reliable bit delivery toward meaning-aware and task-oriented connectivity. Large models (LMs), with strong multimodal understanding and generation capabilities, have accelerated this shift and made semantic communication (SemCom) increasingly practical. Yet current LM-driven SemCom remains fragmented: semantic representations are typically tied to specific modalities, models, or tasks. While the bit provides a universal unit for digital transport, there is still no analogous unit for representing and processing semantics, which limits interoperability, theoretical unification, and scalable system design. We argue that tokens provide a natural candidate for this missing abstraction. Two trends support this: unified multimodal LMs now encode text, images, audio, video, and robot actions in one token space, while distributed LM inference already generates substantial token-level traffic through expert routing, cache transfer, and speculative decoding. Token communication (TokenCom) emerges by unifying these trends, using the LM's native processing unit as a communication abstraction above the bit level and enabling importance assignment, error handling, and resource allocation directly at token granularity. This survey traces the evolution from LM-driven SemCom to TokenCom. We review three major directions of LM-driven SemCom: source-centric semantic coding, channel semantics for physical-layer tasks, and collaborative edge-device intelligence. We then examine the token abstraction, the transmission techniques it requires, and two emerging paradigms, namely TokenCom for LM services and for embodied and agentic intelligence. Finally, we identify open challenges toward unified, scalable, and AI-native 6G communication systems.

Figures & tables

Figure 1: Evolution from bit-level communication to semantic-level communication and further to token-level communication enabled by LMs. The three flows label the unit at which each paradigm represents and optimizes information. All three still require physical transmission, in either digital or analog form.

		Foundations		Research directions & paradigms
Reference	Year	LM	Token	Sem. coding	Ch. semantics	Edge collab.	LM serv.	Embodied
Iyer et al. [ 20 ]	2023	✗	✗	✓	✗	▲	✗	✗
Getu et al. [ 21 ]	2024	✗	✗	✓	▲	▲	✗	✗
Guo et al. [ 22 ]	2025	✗	✗	▲	✗	▲	✗	▲
Liang et al. [ 23 ]	2025	✓	✗	✓	✗	▲	▲	✗
Xia et al. [ 24 ]	2025	▲	✗	✓	✗	✓	▲	✗

Table 1: Comparison of this survey with existing SemCom surveys, aligned with the three research directions and two emerging paradigms identified in this paper. Foundations — LM: large-model fundamentals; Token: token as the communication unit. Research directions and paradigms — Sem. coding: source-centric semantic coding; Ch. semantics: channel semantics for physical-layer tasks; Edge collab.: collaborative edge-device intelligence; LM serv.: token communication for LM services; Embodied: embodied and agentic intelligence ( ✓ : substantive coverage; ▲ : partial; ✗ : not covered)

Panel of Figure 2 — Figure 2: Organization of this survey paper

Panel of Figure 3 — Figure 3: (a) Multimodal large language model (MSA denotes multi-head self-attention and LN denotes layer normalization), and other generative AI models, including (b) generative adversarial network, (c) variational autoencoder, and (d) diffusion probabilistic model.

Architecture	Model	Year	Modality	Token type	Key characteristic
Encoder–decoder	Transformer [ 34 ]	2017	Text	D	Replaces recurrence with parallelizable attention
Encoder-only	BERT [ 33 ]	2019	Text	D	Bidirectional masked-LM pretraining
	ViT [ 31 ]	2021	Image	C (patch embeddings)	Images processed as patch-token sequences
Decoder-only (AR)	GPT-3 [ 48 ]	2020	Text	D	Few-shot in-context learning at 175B scale
	LLaMA [ 36 ]	2023	Text	D	Open-weight LLM family optimized for inference cost
	GPT-4o [ 10 ]	2024	Text , Image , Audio , Video	n.d.	End-to-end real-time multimodal interaction

Table 2: Representative transformer-based and other generative AI models, grouped by architecture family. Modalities are color-coded as Text , Image , Audio , and Video . Token type — D: discrete; C: continuous; D+C: hybrid; n.d.: not publicly disclosed; parenthetical notes indicate the specific representation where noteworthy

Figure 4: Characteristics of LMs and advantages of tokens

Figure 5: Quality-controllable image tokenization with Ada-TokenCom [ 58 ] : reconstructions from a decreasing number of transmitted tokens.

Tokenizer	Year	Structure	Codebook	Tokens	Input size	Generation paradigm
VQ-VAE [ 49 ]	2017	2D grid	512	$32{\times}32$	$128^{2}$	Raster-scan AR (PixelCNN prior)
VQGAN [ 50 ]	2021	2D grid	1,024	$16{\times}16$	$256^{2}$	Raster-scan AR transformer
MaskGIT [ 51 ]	2022	2D grid	1,024	$16{\times}16$	$256^{2}$	Masked parallel decoding
RQ-VAE [ 56 ]	2022	2D grid, depth 4	16,384	$8{\times}8{\times}4$	$256^{2}$	Raster-scan AR, depth-stacked
TiTok [ 52 ]	2024	1D sequence	4,096	32	$256^{2}$	Masked parallel decoding
LlamaGen [ 54 ]	2024	2D grid	16,384	$16{\times}16$	$256^{2}$	Raster-scan AR next-token

Table 3: Comparison of representative discrete visual tokenizers

Aspect	Discrete tokens	Continuous tokens
Representation	Indices in a finite codebook or vocabulary	Real-valued embedding vectors
Interpretability	Arbitrary labels; index proximity carries no semantic relation	Embedding distance correlates with semantic similarity in spaces trained or aligned for it
Compression	Index only; highly compact	Full vector; quantization needed for a finite rate
LM integration	Input by embedding lookup; output by the native softmax head	Input by a trained projector; output needs a non-categorical head
Transmission and robustness	Digital channel coding; residual erasures regenerated from context, avoiding retransmission	Analog JSCC-style symbol mapping; small perturbations give graceful degradation

Table 4: Discrete versus continuous tokens from a communication perspective

Figure 6: Data and computation distribution in a SemCom system

Scenario	Token rate	Token size (bits/token)	Success rate	Delay budget
Human $\leftrightarrow$ Agent	30K $\sim$ 100K/s	10 $\sim$ 20 (small), $\sim$ 400 (large)	99.9% (text), 80 $\sim$ 99% (others)	0.1 $\sim$ 1 s
Robot $\leftrightarrow$ Agent	30K $\sim$ 60K/s	10 $\sim$ 20 (small), $\sim$ 400 (large)	80 $\sim$ 99%	10 $\sim$ 15 ms
Agent $\leftrightarrow$ Agent	up to 30K/s	$\sim$ 20 bits	99.9% (text), 90 $\sim$ 99% (others)	1 $\sim$ 15 ms

Table 5: Example traffic characteristics and service requirements for token communications. Adapted from [ 60 ]

		Source	Coding		Model type			Target
Reference	Year	modality	SSCC	JSCC	DL	LM	GAI	Recon.	Task
DeepSC [ 63 ]	2021	Text	✗	✓	✓	✗	✗	✓	✗
DeepJSCC [ 64 ]	2019	Image	✗	✓	✓	✗	✗	✓	✗
Qiao et al. [ 13 ]	2024	Image	✓	✗	✗	✓	✓	✓	✗
LAM-SC [ 15 ]	2024	Image	✓	✗	✓	✓	✗	✓	✗
D 2 -JSCC [ 65 ]	2025	Image	✗	✓	✓	✗	✗	✓	✗

Table 6: Overview of representative SemCom studies reviewed in Section 4 , by source modality, coding paradigm, model type, and design target. DL: conventional task-specific deep learning networks; LM: pre-trained large/foundation model; GAI: generative model (GAN/VAE/diffusion); Recon.: reconstruction fidelity; Task: downstream task performance

Figure 7: Basic framework of a SemCom system

Aspect	SSCC: separate semantic & channel coding	JSCC: jointly trained end-to-end coding
Architecture	Two-stage and modular: modality-specific semantic extraction produces compact semantics, which are compressed into bit streams and protected by conventional channel coding	A single neural network maps the source directly to channel symbols [ 64 ] , or distinct semantic and channel codecs are trained jointly [ 63 ]
Semantic extraction	Off-the-shelf pre-trained models plugged in per modality, e.g., CLIP [ 32 ] (image), YOLO [ 78 ] (objects), Wav2vec and speech recognition models [ 79 , 80 ] (speech), and PointNet [ 81 ] (3D point cloud)	Learned implicitly by convolutional or transformer-based encoders trained together with the channel codec, without a stand-alone extraction stage
Optimization	Each stage optimized independently; receiver-side generative models recover content without end-to-end training [ 13 ]	Single end-to-end objective over source distortion, semantic fidelity, or downstream-task accuracy, see Eq. ( 17 )
Channel behavior	Cliff effect under extreme channel conditions [ 64 ] ; mitigated by masking-based semantic completion [ 72 ]	Graceful degradation as channel quality varies; further enhanced by adaptive coding [ 82 , 83 ] and diffusion-based denoising [ 44 , 47 ]
Digital compatibility	Native: the bit-stream interface reuses existing coding and modulation stacks	Analog symbols by default; digital variants restore constellation compatibility [ 84 , 85 ] or a bit-stream interface [ 65 ]
Interpretability and flexibility	High: explicit human-readable semantics (texts, sketches, semantic maps), plus freely composable auxiliary modules (memory, masking)	Limited: learned latent features lack explicit semantic modeling; interpretability partly regained by transmitting shared-codebook indices [ 86 ]

Table 7: Contrast between the SSCC and JSCC paradigms for source semantic communication (upper part) and classification of the reviewed systems by source modality (lower part), where each system is classified by its own coding architecture rather than by the subsection in which it is cited

Name	Backbone	Params	Pre-training		Fine-tuning		Downstream tasks
			Datasets	Size	Datasets	Method
Direction 1: fine-tuning an LLM pre-trained on text
ChannelGPT [ 101 ]	GPT-2	124M	WebText	40 GB	Multimodal environment	Few-parameter tuning	Channel generation and prediction
Sheng et al. [ 102 ]	GPT-2	n.d.	WebText	40 GB	DeepMIMO, 73.7K	Input/output modules only	mmWave beam prediction
Zheng et al. [ 103 ]	LLaMA2	7B	Public online data	2T tokens	3GPP UMa, 50K/task	LoRA	Precoding, signal detection, channel prediction
Direction 2: pretraining a foundation model on wireless data

Table 8: Comparison of LM-based approaches to channel semantics for physical-layer tasks. For the upper block, the pretraining columns describe the corpus of the backbone LLM rather than data reported by the cited work. n.d.: not disclosed in the cited work; “—”: stage not used. LoRA: low-rank adaptation; UMa: urban macro; ISAC: integrated sensing and communication; LoS/NLoS: line-of-sight/non-line-of-sight

Figure 8: Token communication framework [ 12 ]

Figure 9: Simulation results for generative image TokenCom as functions of SNR and PER. TokenCom with cross-modal information (CMI) sustains high semantic similarity (CLIP), perceptual quality (learned perceptual image patch similarity, LPIPS), and pixel-level fidelity (PSNR) as the PER increases, attaining TCE gains of up to 67.3% over conventional codebook-based baselines. The baselines are the learned image codec Cheng2020 [ 121 ] and VQGAN [ 50 ] , and “+ R” denotes a scheme in which the receiver requests retransmission of every corrupted packet. The right panel shows representative image reconstructions under decreasing PER. Reproduced from [ 12 ] .

How $w_{i}$ is obtained	Representative work	Decision driven in that work	Limitation
Attention a token receives from the other modalities, averaged and ranked to retain the top $k\%$	LMECM-SC [ 122 ]	Token selection	Measures cross-modal salience rather than task value. Demonstrated as a way of reducing training tokens
Contextual predictability, read inversely from the prior entropy of each position under an MLM shared by transmitter and receiver, so that low-entropy tokens are masked	Shin et al. [ 127 ]	Whether to transmit or mask	Task-independent by construction, and therefore carrying no information about task value
Semantic loss caused by losing each frame, or the number of important words it contains, scored by a pre-trained language model	SIAC [ 138 ]	Power allocation	Defined for text only
Downstream task contribution, taken from the gradient of the correct-class probability with respect to each feature channel	IRCSC [ 139 ]	How many semantic features to retain	Defined per feature channel rather than per token, and only locally valid as a linear approximation
Tokenizer structure, in which coarse scales are decoded first and finer scales encode residuals conditioned on them	VAR [ 57 ]	Coarse-to-fine generation order	Free of cost, but that work performs no transmission, and the scale order need not follow task importance
Feature-wise masking, applied to a representation whose importance order has been reshaped by decreasing retention probabilities and increasing distortion levels across the feature index	ISFR [ 140 ]	Channel matching, feature selection, modulation and power	Requires retraining the semantic encoder

Table 9: Sources of the token priority weight

w_{i}

, and the decision each has been used to drive. Some weights are defined on feature channels or on tokenizer scales rather than on individual tokens

Panel of Figure 10 — Figure 10: Token communication inside LM service infrastructure. (a) KV-cache transfer in disaggregated LLM serving. The prefill cluster processes the prompt and transfers the resulting KV cache to the decode cluster, which generates tokens autoregressively while extending the cache at each step. The link depicts a cross-node deployment [ 151 , 152 ] . (b) Token-level speculative decoding over a device–edge wireless link, drawn for $\gamma=4$ with the third draft rejected. The uplink carries the $\gamma$ draft identifiers and their per-step top- $K$ logits. The target LLM at the edge verifies all drafts in one parallel pass, keeps the accepted prefix up to the first rejection at index $j$ , discards the drafts after it, and returns only $j$ and one token $x_{j}$ ( $j=\gamma+1$ when all are accepted) [ 18 ] . (c) Communication-media selection over a wireless link. Because tokens are generated during inference, the sending agent holds both media at transmission time, and neither dominates: the token identifiers cost little bandwidth but oblige the receiver to re-encode the context with a full prefill pass, whereas the KV cache skips the prefill at a much higher bandwidth cost [ 153 ] .

Aspect	Token identifier transmission	KV-cache transmission
Per-token payload	$\log_{2}\lvert\mathcal{V}\rvert$ bits for the identifier; 16 bytes for the text-based exchange measured in [ 158 ]	$2N_{\mathrm{layer}}N_{\mathrm{kv}}d_{\mathrm{head}}b$ bits per model; 88 KB for four sharers combined in the same experiment
Receiver-side cost	One prefill pass over the received context	No prefill for same-model transfer. Under Cache-to-Cache the receiver still prefills its own input and adds a projection and fusion pass [ 159 ]
Cross-model use	A shared tokenizer or tokenizer negotiation as in [ 137 ]	Token-level and layer-level alignment plus a fuser trained for the model pair [ 159 ]
Effect of residual errors	Index errors can be concealed using the receiver’s generative prior over token sequences (Section 5.5 )	Perturbed tensor entries enter the attention computation. Task-level concealment remains an open problem
Retransmission	Held at one transmission per packet in [ 12 ] by concealing erasures at the receiver	Relies on automatic repeat request (ARQ) or a lower code rate to preserve tensor fidelity
Privacy	The exchanged text is directly readable	Not directly readable, though the cache still encodes the source content

Table 10: Token identifier transmission versus KV-cache transmission as the communication medium between two LM endpoints. Payload values are those reported by Chen et al. [ 158 ] for a four-sharer cache collaboration setting.

\lvert\mathcal{V}\rvert

: vocabulary size;

N_{\mathrm{layer}}

: number of layers;

N_{\mathrm{kv}}

: number of key-value heads;

d_{\mathrm{head}}

: head dimension;

b

: bits per element

Semantic Communication

Semantic communication sends meaning- or task-relevant information so receivers can act without reconstructing every transmitted bit or signal.

See all 103 papers

Token Communications: A Large Model-Driven Framework for Cross-modal Context-aware Semantic Communications
Feb 17, 2025Li Qiao, Mahdi Boloursaz Mashhadi, Zhen Gao +3Semantic Communication
School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China · 5GIC & 6GIC, Institute for Communication Systems (ICS), University of Surrey, Guildford, United Kingdom · Centre for Wireless Communications, University of Oulu, 90014 Oulu, Finland +1
ToDMA: Large Model-Driven Massive Token Communications for Semantic Multiple Access
May 16, 2025Li Qiao, Mahdi Boloursaz Mashhadi, Zhen Gao +2Semantic Communication
The University of Hong Kong, Hong Kong · 5GIC & 6GIC, Institute for Communication Systems (ICS), University of Surrey, GU2 7XH Guildford, U.K. · Beijing Institute of Technology, Beijing 100081, China +2
Ada-TokenCom: Rate-Adaptive Token Communications via Large-Model-Driven Token Compression and Generation
Aug 28, 2026Zijun Zhang, Li Qiao, Mahdi Boloursaz Mashhadi +3Token Compression Semantic Communication
School of Information and Electronics, Beijing Institute of Technology (Zhuhai), Zhuhai 519088, China · Department of Electrical and Computer Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong SAR · University of Surrey, Guildford, U.K. +1

Feb 17, 2025J/K move · Enter open · S save

Token Communications: A Large Model-Driven Framework for Cross-modal Context-aware Semantic Communications

Li Qiao, Mahdi Boloursaz Mashhadi, Zhen Gao, Rahim Tafazolli+2

School of Information and Electronics, Beijing Institute of Technology, Beijing 100081, China · 5GIC & 6GIC, Institute for Communication Systems (ICS), University of Surrey, Guildford, United Kingdom · Centre for Wireless Communications, University of Oulu, 90014 Oulu, Finland+1

Semantic Communication

Open paper AskPDF arXiv

In this paper, we introduce token communications (TokCom), a large model-driven framework to leverage cross-modal context information in generative semantic communications (GenSC). TokCom is a new paradigm, motivated by the recent success of generative foundation models and multimodal large language models (GFM/MLLMs), where the communication units are tokens, enabling efficient transformer-based token processing at the transmitter and receiver. In this paper, we introduce the potential opportunities and challenges of leveraging context in GenSC, explore how to integrate GFM/MLLMs-based token processing into semantic communication systems to leverage cross-modal context effectively at affordable complexity, present the key principles for efficient TokCom at various layers in future wireless networks. In a typical image semantic communication setup, we demonstrate a significant improvement of the bandwidth efficiency, achieved by TokCom by leveraging the context information among tokens. Finally, the potential research directions are identified to facilitate adoption of TokCom in future wireless networks.

Abstract

Semantic Communication

Explore similar work

Token Communications: A Large Model-Driven Framework for Cross-modal Context-aware Semantic Communications

ToDMA: Large Model-Driven Massive Token Communications for Semantic Multiple Access

Ada-TokenCom: Rate-Adaptive Token Communications via Large-Model-Driven Token Compression and Generation